You have revenue. You have retention. You have a legacy services business.

The metrics that are supposed to catch this don’t. Revenue growth, retention, and churn all read identically whether AI is doing the work or headcount is.

“You haven’t built an AI-native services business. You’ve built a services business with the wrong kind of funding.” That’s Jake Saper of Emergence Capital, who calls this Mirage PMF.

By the time gross margin tells you otherwise, twelve months of compounding have already passed. A venture cap table expects software economics. The delivery model is producing 33 percent.

The trap closes one hire at a time

The early deals close before you finish pitching them. You are cheaper, or faster, or better than what the client had before, often all three, and word gets around before the sales deck does. To meet the demand, you hire forward deployed engineers, implementation specialists, anyone who can get the client’s operation running and keep it running.

Every person you bring on to deliver the service is a person not building the AI platform that would eventually make delivery cheaper at scale. Revenue keeps growing because headcount keeps growing. The AI leverage curve stays flat because there is always a more urgent delivery problem than a platform investment.

She has a badge at the client site, a desk in the building, and a name people recognize in the hall. She knows the operations director down the hall by name, knows he has been waiting since Tuesday for this particular deliverable, and knows she will have to look him in the eye and explain why if it slips. The platform primitive she could instead be abstracting this afternoon has no face, no deadline, and no one waiting.

That is the trap. Not a bad decision. Not a failure of discipline. Just the steady, rational pull of the concrete over the abstract, compounding one quarter at a time.

Investors celebrate growth rate and wait for gross margin to tell them something went wrong. The question they are not asking is simpler: is the platform doing more of the work each quarter, or the headcount? The two look identical from the outside for longer than most people expect.

What Palantir actually built

Palantir is the only company that has run the forward deployed engineer model for long enough, and transparently enough, to show what happens when you don’t leave the knowledge in the field.

Palantir was founded in 2003, in the aftermath of the September 11 attacks, to solve problems for U.S. intelligence agencies that no existing software could handle, the kind where clients could not articulate what they needed because the solution had no precedent. Shyam Sankar, now Palantir’s CTO, who joined as employee 13 and formalized the FDE role by 2006, understood the organizational implication early. If a problem could have been solved through a requirements document, it would have been solved already. The FDE’s job was to learn what the client could not articulate, by being present where the actual work happened.

That insight produced a specific organizational structure: Palantir split its field work between Delta engineers, who built the bespoke solution for each client engagement, and Dev engineers, who took what the Delta built and abstracted it into a reusable Ontology primitive. Forward-deployed engineers operate upstream of the roadmap. Consultants operate downstream of the contract. The field learning did not stay with the engineer who earned it. It traveled into the platform through the Dev. Every subsequent deployment started higher up the stack because the previous engagement had been encoded into the product, not just delivered to a client.

Twenty years of platformization ran through a specific product sequence. The same move repeats in each step, at a larger scope each time: take something that requires bespoke human effort and turn it into reusable machinery. Gotham encoded Ontology primitives for U.S. intelligence and defense clients. Metropolis proved the approach could work commercially, but only inside one vertical, rebuilt from scratch per client per industry.

Foundry fixed that and made the ontology logic industry-agnostic and self-service, dropping integration time from years to weeks. Apollo applied the same move to Palantir’s own delivery machinery, productizing the deployment process itself. AIP applied it to AI, building the grounding layer once so every customer deployment inherits it. AI FDE, generally available in March 2026, automated the configuration and iteration work FDEs had spent twenty years doing by hand. The object being platformized gets bigger each time: one client, one industry, all industries, the delivery process, AI itself, the human operating the platform, and finally the substrate other companies build on.

Between 2020 and 2025, Palantir generated four times the revenue from less than two times the people: revenue from $1.1 billion to $4.48 billion, headcount from 3,100 to 4,429, gross margin from 67 to 82 percent. The platform was doing the work the additional people would otherwise have done.

The first sign that the platform was absorbing that work appeared in deployment speed. By mid-2020, average installation time had fallen fivefold, to fourteen days. The margin line was still catching up.

The org chart spread. The mechanism stayed put.

The FDE org chart spread fast. The Palantir mechanism stayed at Palantir.

FDE job postings grew 1,165 percent year-over-year between January and October 2025. The role has moved from Palantir idiosyncrasy to industry standard in roughly eighteen months. The role spread faster than the mechanism it was built around.

The column that matters is the one where field learning routes back into the platform.

Each FDE deployment is a different economic bet. The FDE role is the correct organizational response to a specific structural condition in enterprise AI today. Deploying AI agents into workflows that have never been automated means no requirements document can be written, and the customer’s business process has to change to accommodate how agents work. Every industry player fielding this problem is responding to it differently.

Jesse Zhang, co-founder of Decagon, separates two kinds of work that most companies deploying FDEs collapse into one title: discovery, where neither the company nor the customer yet knows what the right solution looks like, and delivery, where the workflow is known and the work is execution against a spec. The companies in the high-margin, low-Mirage-PMF column are doing the former, routing what their FDEs learn back into the platform. Everyone else is doing the latter in varying degrees, and hiring FDEs anyway.

Sankar, watching the spread of an org chart most companies were adopting without understanding its mechanism, named it: “It’s just mimetic.”

Three numbers. One matters now.

“2026 will be the year of reckoning for AI-enabled services.” Emergence Capital’s reckoning is about the failure to encode technology learnings into platforms across the many companies now running high-touch AI delivery models with venture-backed balance sheets and services-level margins.

Standard enterprise metrics cannot detect it. Revenue growth, retention, and churn all read the same whether the AI is doing the work or the headcount is. Emergence’s framework for detecting Mirage PMF was built to measure what those metrics miss, specifically whether the platform is actually doing more of the work each quarter, or whether growth is simply tracking the headcount. The framework runs on three tiers.

The leading indicator is task-level AI leverage. Emergence insists on granularity at the step level: the specific tasks, the individual handoffs, the precise operations where a human hand is still required, rather than a macro claim like “we’re automating 40 percent of the workflow.” Crosby Legal, an AI-native legal services firm, tracks HURT: human review time per document. That single metric, owned by a named person and reported quarterly to the board, is the leading indicator of whether the AI is gaining ground. As HURT falls, margins rise. As HURT stays flat, growth is coming from headcount.

The middle indicator is ARR per service FTE relative to the legacy-provider baseline. If the company is not generating at least 2.5 times more revenue per service-delivering employee than the incumbent it is displacing, the AI is not creating real leverage yet. This ratio must climb each quarter. A flat ratio means the platform has not moved. Hiring is doing the work the AI should be doing.

The lagging indicator is gross margin, calculated to include delivery labor and inference costs in full. Above 70 percent commands software multiples. Between 50 and 70 percent means the business is better than legacy services but priced like one. Below 50 percent is, in Emergence’s precise phrasing, “a services business with a venture cap table.”

“The first metric is the leading indicator. The third is the outcome. Most founders only watch the third, which means they find out they have a problem 12 months too late.” The gross margin tells you what happened. The task-level metric tells you what is happening.

The staircase

Emergence holds one AINS portfolio company as the standard-bearer for method alone, running the approach that makes every other company uncomfortable. Every other quarter, the CEO stops selling entirely. The team builds out the AI platform, trains new human operators on top of it, and deploys them. Then it reopens sales. The resulting revenue curve is a staircase, a step-function that rises each time the platform becomes more capable. The company has maintained gross margins in the low 70s while scaling.

A board looking at two consecutive quarters of flat revenue growth will ask whether the strategy is working, whether the build quarter was spent on the right things, and whether the staircase is real or a rationalization. The CEO has to hold the line: the flatness is the investment, and the next revenue step will run at higher margin because the platform absorbed more of the delivery than the previous step did.

The threshold at which that discipline becomes sustainable varies. Crosby Legal found it at roughly 80 clients, enough to tell a rule from an exception. At that point, the company could start saying no to non-roadmap requests. When the platform is doing enough of the work, saying no becomes operationally viable. That shift is the clearest sign compounding has started.

Four moves

Most AINS founders who recognize the trap are already inside it.

The first move is to run the diagnostic honestly. The gross margin has already told you your engineers are doing the work the platform should be doing. The task-level breakdown tells you where: which steps the AI completes at acceptable quality, and which still require human judgment to close. Go one level deeper than the macro answer. If you cannot produce that breakdown right now, the gross margin will eventually tell you what happened. The task-level metric tells you what is happening now.

The second move is harder because the revenue is real. Stop taking non-roadmap work, even when it costs you a client. That instinct is strongest inside the trap. The requests feel small, the relationships feel real, and the platform work they displace has no deadline. But each non-roadmap engagement is the platform investment not made.

The third move is to stop selling before the platform forces you to. Pause selling for sixty to ninety days, build the platform, train operators, and reopen. The flatness is the investment. Each step up should run at higher margin than the one before. Flat margin after each build cycle means the engineering team carried the delivery the platform was built to absorb.

The fourth move is to assign a named person to the leading indicator. One person, responsible for task-level AI leverage. Their number climbs each quarter. That is their job.

The recovery looks like contraction before it looks like anything else. Revenue flattens. Are your FDEs discovering something, or absorbing something?

What got built into the product the last time one of them came back?

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

Buying the best legal AI tool on the market right now doesn’t give you a competitive advantage. It gives everyone else one too.

Kirkland & Ellis (K&E), the highest-grossing law firm on earth, just bet $500 million on that premise.

The platform K&E is building will be owned outright and never sold to competitors. More than $100 million of that will land in 2026 alone, funded from the firm’s $10.56 billion in revenue. Profits per equity partner reached $11.1 million last year, up 20%. That’s the margin that makes a non-billable infrastructure build survivable, even as it eats into partner distributions in the short term. No other law firm has it.

Jon Ballis, K&E’s chair, told the Financial Times that widely available AI tools are “raising the floor for everyone.” “We don’t get hired for the floor.”

The correct institutional response to a rising floor is to build something above it that isn’t for sale. The question the $500 million is supposed to answer is whether that ceiling holds.

In June, at Palantir’s AIPCon conference in Miami, Erica Berthou gave the room a live demonstration of the attempt.

Berthou co-leads Kirkland’s global fund formation practice. Her group helps raise hundreds of billions a year for well over a thousand fund formation clients, across virtually every asset class, at a scale no other law firm matches. The demonstration was built on thirty years of that volume.

She opened with the diagnosis. Off-the-shelf tools handle discrete tasks well enough. What they can’t do is hold a transaction together, and they’re trained on the same widely available market knowledge every other firm has access to, which means the output converges on the common denominator. “That does not work for us,” she said. “At the absolute beating heart of what we offer our clients is our differentiated expertise.”

The demonstration centered on a fictional fund called Vulcan Industrial Partners. A $3 billion target, 300 investors, six months to close. The Fund Formation Engine ran the scenario. Built on Palantir infrastructure, it holds thirty years of K&E fund formation judgment in operational form.

The Fund Formation Engine runs the full fundraising lifecycle as a connected system. A lawyer drops a commercial term sheet into the engine and four specialized agents work simultaneously: one captures insights from senior partners’ prior art, one extracts and structures language from model documents, one drafts new provisions, and one folds feedback into each revision. A draft surfaces in minutes. “What used to take days,” Berthou said, “now happens in minutes.”

A month into preparation, Vulcan’s managing partner wants to expand the geographic mandate to include emerging markets. The engine assesses what changes the governing documents need, pulls market data, surfaces examples from comparable transactions, then reads the current geographic restriction and offers a menu of alternatives. What used to be a research problem is now a lawyer’s judgment call.

Investor comments on the geographic change arrive next. The engine organizes every deal point, surfaces solutions, and assembles the firm’s prior art alongside relevant market data. The K&E lawyer still applies the judgment, but now with everything she needs already in front of her. “Her speed and expertise are vastly multiplied,” Berthou told the room.

When an investor negotiation produces a bespoke bilateral agreement, a separate part of the engine takes over. It ingests every agreed term and generates language from a hierarchy of potential solutions, drafting three options for each point. The lawyer chooses. The document closes around the choice the lawyer made.

Eighteen months after close, a time-sensitive deal requires checking every investor commitment for Sub-Saharan Africa obligations. The engine pulls every relevant obligation, generates a checklist, flags the right people, and drafts the reminder emails.

What makes the sequence work is the ontology underneath the agents: the underlying data architecture linking funds, market data, obligations, and transaction history into a single operating system. Berthou called it “knowledge that historically lived across a vast number of attorneys, documents, spreadsheets,” now “structured and applied directly in the flow of execution.” Ryan Taylor, Palantir’s chief legal officer, called it an operating system that “compounds every transaction, relationship, decision and obligation” with each use. The useful memory is the platform.

Most firms that attempt this encoding step get it wrong: they staff the work with people who can describe what senior partners do but not how those partners actually decide. The system learns the vocabulary of expertise without the substance, producing automation that looks right and says the wrong thing. K&E answered it with a firm-level mandate. Roughly 250 attorneys, including about 100 equity partners billing at $1,500 to $2,000 an hour, are doing the knowledge-encoding work. Unreplicable by design. Non-billable by necessity.

The litigation build starts from the same premise. An exclusive partnership with Syllo, a platform for case management from filing to verdict, gives Kirkland custom-build rights that sit above a base product Syllo continues selling to the rest of the market. Elizabeth Hess, a litigation partner and Executive Committee member, described Syllo as “decisively ahead of every other solution” the firm evaluated following a multiyear market review. What Kirkland builds on top belongs to Kirkland.

The stack underneath both builds is commodity by design. Palantir’s AIP and Syllo are commercial products. Azure AI services are standard infrastructure. The proprietary element is the ontology, the institutional heuristics, and the judgment encoded into the workflows. Replicating three decades of fund formation know-how, encoded by the 100 equity partners who accumulated it, is measured in years.

No consensus has formed on the right move.

Fried Frank built FundAssist on OpenAI and is weighing whether the AI build is an internal tool or a product, a different kind of bet than the others. A&O Shearman is building agentic tools with Harvey and licensing them to other firms, converting the technology investment into a revenue stream. Freshfields signed with Anthropic for early model access and co-development rights, betting that privileged proximity to the frontier model is the position worth owning. Latham & Watkins is building a dedicated AI practice with mandatory associate training and AI-native workflows, rather than a proprietary platform.

Own and wall off. Build and sell. Partner and access. License and adopt.

All four firms are working from the same premise: AI will reshape how elite legal services compete. What they disagree on is direction. The K&E theory is enclosure: build something unreplicable and wall it off. The A&O theory is distribution: build something valuable and sell it. One of those theories has a structural problem.

A&O Shearman’s is the theory with a structural problem. If the tools A&O builds with Harvey are genuinely differentiated, licensing them to other firms transfers that differentiation to competitors who can now buy what A&O spent to create. If the tools aren’t differentiated enough to confer real advantage, the licensing revenue doesn’t compensate for the strategic cost. A&O is either selling its moat or selling mediocrity.

Client pressure has a predictable shape: when one firm’s output is visibly faster and more consistent, clients start asking the firm across the table why it isn’t matching that. What’s less visible is the financial threshold underneath it. Non-billable encoding work at that scale is only survivable at K&E’s profit margin. Most of the firms that will face that pressure don’t have $11.1 million in PPP to absorb the opportunity cost. K&E’s bet is that the firms facing the most client pressure are the ones least able to afford the answer.

The billable hour makes the financial pressure deeper than the build alone. A platform that cuts a three-day job to three hours generates less revenue under hourly billing, not more. Ballis has already called the direction: “We already do a number of matters on value-based pricing, and that trend will only continue and it will accelerate.” The logic runs in sequence: encode thirty years of judgment, compress the time, then price on the outcome rather than the hours. The platform is a repricing instrument.

Ballis hasn’t described a commoditization tripwire. No stated threshold for reassessment, no mechanism for detecting when a competitor closes the gap. Vendors who accept exclusivity today have roadmaps that shift. The platform’s flywheel logic is an argument for why the lead will hold, not a mechanism for detecting when it doesn’t.

Vulcan Industrial Partners doesn’t exist. The platform that ran it does. Every legal matter that closes on it makes it harder for a late entrant to replicate from scratch. But the firms that will own the ceiling aren’t necessarily the ones who built earliest. They’re the ones who’ve answered the question K&E hasn’t: what would it take for a well-funded competitor to replicate what we’ve built, and who has the authority to surface that evidence when they find it?

K&E has shown the scale of what it would take: $500 million and a hundred equity partners doing non-billable work. The second half is the only one that determines whether the investment was worth it.

Investment banking faced a version of this when Long-Term Capital Management, a quant hedge fund that used Nobel-winning models to predict bond markets, collapsed in 1998. The Chief Risk Officer role was invented to give one person genuine authority to say “the model is wrong” before the market demonstrated it.

The banks that survived the 2008 financial crisis had CROs who reported outside the revenue chain, with access to every position and the mandate to challenge the consensus even when the consensus was generating the bonuses. The banks that didn’t had risk functions reporting to the people whose compensation depended on the models being right. Any firm reading this in 2026 still has what the banks in 2008 didn’t: time. The firms that will own the ceiling are the ones who name someone with the authority to say “the model is wrong” before a client forces the question. The role wasn’t new in 2008. The authority was.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

Tokens fell 10x. Enterprise invoices rose 500%. If you understood what was actually moving on your invoice, you’d be paying roughly half of what you’re paying now.

Coinbase figured that out. By June, a regulated fintech running on legacy core banking rails had cut its AI spend nearly in half while its token usage exploded. It paid market rate per token. It just stopped buying the wrong ones.

Four changes followed. Routing defaults shifted to open-weight models. A gateway began reserving frontier capacity for planning while sending execution to cheaper alternatives. Per-task context trimming went in. The cache hit rate went from 5% to 60%.

The measurement changed before anything else did. CEO Brian Armstrong named it directly: “Humans shouldn’t be choosing models at all. The system can.” Coinbase had been managing the metric on the invoice rather than what the tokens were actually purchasing. Switching the measurement unlocked everything else.

What the invoice doesn’t say

Enterprise open-source AI spend fell from 19% to 13% last year. One number, three populations, no labels.

The largest population is the newest cohort in enterprise AI spend. Code generation, copilots, and agentic pilots that default to frontier APIs during discovery. Anthropic’s Claude alone captured 42% of code generation workloads last year. These are exploration-stage workloads. They inflated the denominator. The open-source share fell. Its position didn’t.

The second population migrated successfully to open-weight models and became invisible. Spend surveys read API billing, not self-hosted compute running Llama, Mistral, or Qwen inside private cloud environments. Migrations that succeeded are absent from the number.

The third population is held in place by friction. Licensing uncertainty, cybersecurity requirements, missing fine-tuning pipelines, and incomplete domain data keep mature workloads on frontier past the point where the economics would justify moving.

Three signals, one number, no labels.

Frontier models dominate discovery. Open and sovereign models increasingly power the mature, high-volume production layer. Enterprise AI is early enough in its arc that the discovery layer fills faster than established use cases can graduate to production. Capability gaps are keeping the graduation rate low. Your workflows are already positioned in that split.

The manufacturer’s dilemma

I’ve been watching the same pattern across teams building at this layer for the past 18 months. The ones that figured out the measurement first, shifting from tokens consumed to cost per outcome, moved faster than everyone else in the stack. The ones still managing the invoice metric are subsidizing the research that will eventually price them out of the workflow they built. The logs already know which side each team is on.

Netflix watched which studio content got rewatched, paused, and shared. House of Cards wasn’t a creative bet. The studios funded the market research. The invoice called it a licensing fee. When enough signal had accumulated, the license renewals became optional.

Anthropic ran the same sequence on a faster clock. Enterprise API invoices funded 13 vertical products across finance, legal, and healthcare inside 15 months. Every workflow rebuilt three times on the API was a product specification. Every call to the logs was a focus group. Millions of focus groups. They called them customers.

The workflow your team stabilized is already a product specification. The lab has it. You are not the intended customer.

You didn’t buy inference. You bought visibility rights on your own business logic. The fee was labeled per token.

The AI Flight Plan Matrix

The four quadrants are four versions of the same problem with different timing.

Test Flights. Immature use cases, frontier models.

The most expensive AI mistake isn’t paying frontier prices. It’s fine-tuning too early, before the workflow knows what it is.

At week two, a contract review tool doesn’t know what it is yet. A clause extractor, a multi-document reasoner, or something the team hasn’t named. Fine-tuning commits to one answer before the question is closed. Frontier APIs hold the answer open. Decagon’s team could test any new model release without standing up infrastructure first. The cost of a wrong architectural bet outweighs the per-token premium.

Safety alignment and red-teaming investment are a second reason this quadrant runs on frontier. Test Flights operate in domains where the failure mode hasn’t been documented yet. In undocumented territory, governance costs and failure-mode liability are real. Frontier labs have already priced that governance cost in.

Chartered Jets. Mature use cases, frontier models.

Maturity usually predicts migration. This quadrant is where it doesn’t.

Harvey runs the sharpest version deliberately. Claude Opus handles multi-step legal reasoning. GPT handles long-form sourced output across case law and contracts. Sonnet and Gemini Flash cover document volume. When a hallucinated citation is malpractice exposure, per-turn economics never force migration.

A second population stays for agentic reliability, not liability per turn. Open-weight models are measurably behind frontier on agentic benchmarks. The failure mode is not a single wrong answer. It’s a workflow that reaches step seven and cannot recover. That is the arithmetic that keeps these workloads on frontier infrastructure.

A third is held by ecosystem gravity. Coding tools with two years of IDE-embedded, model-tuned production history don’t migrate on a benchmark comparison. Two-thirds of Anthropic’s API revenue flows through two of these clients. The choice was made at the integration layer. The model was incidental.

Chartered Jets is where the manufacturer’s dilemma is most exposed. The workflow is stable enough to copy, valuable enough to justify it, and dependent enough to make leaving painful. Harvey is already redirecting engineering capacity into fine-tuned open weights to reduce exposure to any one lab’s pricing leverage. Kimi K3’s weights went public last week, open-weight and self-hostable, benchmarking second only to the two best frontier models available. The agentic reliability gap that keeps the second population here is narrowing fast.

Homebuilt Aircraft. Immature use cases, open or sovereign models.

Some workflows are too valuable to prototype in public. A frontier API call is an IP transfer in the form of an inference request.

Pfizer’s molecular data and a hedge fund’s trading signals share the same constraint. The IP that makes discovery possible is the IP you cannot share. Pfizer runs molecular binding on real compound data because synthetic molecules teach you about synthetic molecules.

BharatGen reached a different version of this quadrant without any frontier API having existed to skip. No commercial model existed for 22 Indian languages at the scale India’s government deployment required. Today it runs the citizen services layer for 1.47 billion people.

Three cohorts reach this quadrant without passing through Test Flights. AI-native companies with ML infrastructure already built never needed a frontier API’s generality. Sovereign environments where frontier APIs are architecturally blocked never had the option. Some domains had no frontier model to skip because no lab had entered the market.

None ran the IP transfer. The manufacturer’s dilemma doesn’t apply to what the lab never saw.

Scheduled Routes. Mature use cases, open or sovereign models.

Every mature, high-volume workflow that ran long enough on a frontier API accumulated two things as a byproduct. Stable business logic and a training dataset. Those are the two prerequisites that make fine-tuning viable.

Decagon’s customer support voice agents ran on frontier infrastructure until the task was understood well enough to teach a smaller model the same work at one-sixth the cost. The migration didn’t require new data. It required recognizing what had accumulated in the logs.

Decagon now runs 90% of inference on proprietary open-weight models through a dedicated NVIDIA Blackwell partnership with Together AI. Three properties drove the migration. Volume made the per-turn cost gap decisive. Latency ceilings a frontier round-trip can’t clear settled the voice question before the financial argument was assembled. Domain boundedness meant the general capability they’d been paying premium prices for was no longer the capability the workflow needed.

Flying the diagonal

The path from Test Flights to Scheduled Routes is the diagonal, and four readiness gates determine whether a workflow can complete it. Workflow stability, data readiness, volume economics, and governance maturity. They open in sequence. Missing one means the migration completes on paper while the cost finds a new address.

Phase 1. Discovery. Committing to one answer before the question is settled is the wrong call. Prompts rewrite weekly, success criteria shift, and no production data pipeline exists because there is nothing stable enough to build one around. Fine-tuning has no signal. The frontier premium here is the cost of staying uncommitted long enough to get the question right.

Discovery ends when the workflow knows its own failure modes. That requires enough repetition to surface edge cases, enough volume to make quality measurable, and enough usage to register on a budget. Most enterprises with active AI programs are still accumulating that signal. The ones further along stopped guessing at it and started reading it from the logs.

Phase 2. Hardening. Hardening is the phase where the organization is operating a system it doesn’t fully own yet. The workflow has moved to production, generating real cost, real failures, and the first training signal worth collecting. Routing architecture gets its first real test, sending sub-tasks to smaller models and reserving frontier capacity for decisions where a wrong step can’t be unwound. Coinbase ran hardening until pre-processing prompts, task-type routing, and a cache hit rate that went from 5% to 60% were all in place.

Hardening ends when edge cases are documented and labeled examples make fine-tuning viable. That threshold also forces the sovereignty question: security, compliance, and data residency requirements determine whether inference runs inside the firewall or outside it. Who controls that layer is a different decision than which model runs on it.

Phase 3. Industrialization. This is where the cost gap between frontier and open-weight models moves from benchmark to budget line. At that threshold, model fungibility stops being architectural. It becomes arithmetic. Four gates determine whether a workflow can leave frontier infrastructure. The workload repeats at volume, open or sovereign models meet quality thresholds, fine-tuning yields measurable lift, and the team has engineering depth to operate the stack.

Decagon is the earliest enterprise beachhead where model fungibility crossed from thesis to practice. Customer support ran at the volume where per-turn cost becomes the only number that matters, in a domain narrow enough that a smaller model could learn the full job from logs. The migration question answered itself.

Getting to Scheduled Routes is a rebuild, not a replacement. The workflow has to be redesigned from outside a conversational window, around new training data, new orchestration architecture, and a continuous eval loop. None of that is in the model.

The migration ceiling

The orchestration layer that makes model-switching viable at agentic depth doesn’t exist yet in mid-2026.

Nikesh Arora, Chairman and CEO of Palo Alto Networks, articulated what doesn’t move when the model does. Data retrieval, API routing, and memory are genuinely portable today. What isn’t portable is the harness. Prompts and tool schemas tuned to one model’s quirks, evals that require full re-verification on any switch, cached workflows, and cache pricing that covers 75 to 90% of input costs on agentic tasks. Each of those is quiet model lock-in. Arora’s current read is that 75% of enterprises will realistically end up building around one model stack unless the orchestration problem is solved in the next cycle.

Three operating model constraints determine whether the migration question is even askable.

Data readiness. Decagon reached 90% open-model inference because it accumulated millions of labeled customer support conversations as a byproduct of operating at scale. Most organizations don’t have equivalent structured domain data. The binding constraint is training signal, and it only accumulates if you’ve been running production long enough to generate it.

ML and inference capability. Fine-tuning pipelines, MLOps infrastructure, inference serving with defined latency budgets, and continuous evaluation all require deliberate investment. Most teams don’t build them until a pricing event forces the question.

Governance maturity. Most security functions don’t have a framework that answers who the model is acting as, what it’s authorized to do, what happens when it fails, or who is accountable. Late governance fails because the infrastructure has already made, by default, the decisions governance was supposed to make.

The only question worth answering now

The manufacturer’s dilemma has a resolution. Self-host and the lab’s visibility ends. What remains is a dependency the operator chose rather than inherited, with switching costs measured in engineering effort rather than a renegotiation with the party that already knows exactly what the workflow costs to replace.

Decagon reached that position. Coinbase reached it. Decagon’s migration took 18 months after the training signal was there to act on. The signal accumulates whether you’re reading it or not. Starting the read now puts you 18 months ahead of the pricing event, not 18 months behind it.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

Every enterprise AI contract places your organization inside a tax regime. Most sign without reading the constitutional clause about exit. When Alex Karp went on Squawk Box July 1st to channel “the fury of American business,” he was right about the problem. The collection mechanism he built is the exhibit.

“The basic view among enterprises in this country is: I’m going to chillax and waste my time with tokens, I’m going to get no value, and they’re going to get my IP.”

Underneath the token pricing debate sits a compute shortage: costs have risen 2 to 4x in two years, and demand is still ahead of supply. More than half of that constrained capacity runs as a consumer subsidy. Enterprise workloads are the funding mechanism. When half the capacity is a loss leader, the other half carries the full freight. OpenAI’s API input pricing moved from $1.25 per million tokens for gpt‑4o to a range of $1.75 to $5.00 across the GPT-5.x family in roughly a year, with GPT-5.5 at the top of the range. Anthropic didn’t need to raise the price. In April 2026, it closed the consumer subscription workarounds enterprises had used to route agentic workloads around the meter. The exit was the announcement.

Karp is selling you a different kind of tax, one that looks like sovereignty and functions like annexation. The vehicle changes. The toll doesn’t.

The First Regime: The Swipe Fee and the Surveillance Fee

Visa and Mastercard collect their percentage of gross before you know whether the transaction made money. The retailer’s P&L is not their problem. Visa also sells transaction data back to issuers, advertisers, and category analysts. The swipe fee is the visible line item. The data product runs on the same rails and bills separately. Two taxes, one card reader.

Frontier labs charge a toll on every token, every API call, every agentic task completion. Visa’s swipe fee is a known quantity. The AI version runs on a consumption curve that looks predictable until agentic workloads hit it: one task fans out into dozens of API calls, and the meter runs on all of them. The unit economics that felt transparent at the single-query level become structurally opaque at scale.

Per-developer token usage multiplied 18.6x in nine months as agentic workloads scaled, according to one FinOps audit across major enterprise deployments. The FinOps Foundation is now building a Tokenomics Foundation just to govern AI token spend. When your cost category requires its own governance body, that’s a regime.

The prompt goes through external servers every time an employee runs a query through a frontier model API. Your workflows, customer data, pricing logic, internal processes, all of it transmitted through a third party whose business model depends on knowing which categories are worth owning next.

Your API logs are a due diligence report. Frontier labs know which verticals to own next because you just filed it.

Microsoft revoked enterprise Claude Code licenses months after enabling them, drawing the line at letting a competitor’s model retain employee prompts. If the most AI-implicated company in the Fortune 10 pulled the licenses after running that calculation, someone at your company should be running the same one before the next renewal.

The Second Regime: The Workflow Sovereign

In November 2023, the UK’s National Health Service awarded Palantir a £330 million contract to build the NHS Federated Data Platform, integrating health data across NHS trusts that previously couldn’t talk to each other. Palantir built capabilities the NHS didn’t have. The value is not in question. The exit is.

Palantir’s own documentation describes its Ontology as “an operational layer for the organization,” a “digital twin of the organization.” Every entity the NHS cares about (patient record, clinical trust, diagnostic pathway, procurement workflow) gets mapped as an Ontology object. Nothing in the system acts without consulting the Ontology. Alerts, approvals, decisions. Not one of them fires without going through the Ontology.

The exit tax here is simple. You reconstruct the entire data architecture of a national health system from scratch. A mature Palantir deployment across major operational domains represents 18 to 36 months of reconstruction effort. Default contracts provide for raw data export in CSV and JSON formats, not portability of the Ontology configuration, the workflow definitions, or the model training pipelines.

By June 2026, a UK parliamentary Science, Innovation and Technology Committee described Britain’s reliance on Palantir as an “unacceptable point of weakness” and recommended invoking the contract’s break clause. The committee’s concern wasn’t the price. It was the exit cost.

Switzerland ran a 20-page federal risk assessment on a comparable proposed deployment before signing anything. Exit dependencies were structurally incompatible with Swiss data governance requirements. They declined at the procurement stage rather than the parliamentary one.

The workflow sovereign is the operating system. The enterprise is the application running on top of it.

The Third Regime: The Sovereign Stack Empire

Karp’s counter-offer is constitutional. Own the hardware, own the model weights, control what the model sees. The vehicle is the Palantir-Nvidia sovereign AI stack. Karp didn’t go on Squawk Box to vent. He went to sell Nvidia Blackwell Ultra GPUs for compute, Nemotron open-weight models for inference, and Palantir’s full AIP/Foundry/Ontology/Apollo operating stack, all deployed on customer infrastructure. No token meter. No IP flowing to a frontier lab. The buyer owns the silicon and the model.

Karp pitches exit from all three tax regimes at once. Two of those he delivers. An air-gapped deployment stops IP leakage. Your workflows don’t flow to a frontier lab studying your category before they build the vertical to displace you. Running Nemotron locally exits the token meter. The lock-in claim doesn’t hold.

Sovereign stacks don’t abolish the AI tax. They nationalize the tax.

The U.S. Army’s NGC2 program made Palantir’s Ontology the common data layer for Army AI-driven command and control. The Army got genuine control over weights, data, and inference. Every downstream contractor, allied agency, and software vendor that needs to interoperate with that system now conforms to Palantir’s architecture. They didn’t negotiate it. They inherited it. The sovereign stack that liberated one buyer created a new AI empire for every participant who didn’t write the rules.

Hosting it yourself changes the address. Palantir’s control plane doesn’t move with the hardware. The Ontology, the layer encoding what your organization knows and how it acts on that knowledge, stays with Palantir whether it runs on their infrastructure or yours. The zip code changes; the lease terms travel with the building.

Ownership is the distraction. The question that matters is how much of Palantir’s architecture you’d have to rebuild to leave it. Which raises the obvious question: if the sovereign stack still locks you in, who actually sits at the top of this arrangement, and what do they collect?

The Capital Structure Nobody Shows You

Nvidia pulled $75 billion in profits in 2025 at 65% operating margins. Hyperscalers spent close to half a trillion in capex over the same period. In any capital structure, someone sits at the top and collects regardless of whether the operating business performs. Someone else absorbs the losses when it doesn’t. The AI stack has exactly that shape. Senior creditor, junior creditor, equity, in that order, from silicon to services.

Nvidia collects on every GPU cycle whether the model above it succeeds or fails, which is the only position in the entire stack that doesn’t require the AI to actually work.

Frontier labs sit in the junior creditor slot, pulling material revenue at margins not yet demonstrated at scale. The swipe fee is real. The durable profit isn’t, yet, which is exactly why Anthropic isn’t waiting around to find out.

Frontier labs don’t need to breach your contracts or read your logs. They can see which categories are exploding on their own APIs, map your cost structure from consumption patterns, and send PMs to interview your customers. The intelligence operation runs inside the billing dashboard. The operating platform rung above it doesn’t need to surveil you. You already handed it the constitution.

Above frontier labs sit operating platforms like Palantir, collecting at the operating layer through something far stickier than a token meter, the Ontology and workflow logic they own. Their Forward Deployed Engineers (FDEs) physically embed inside client operations as a hybrid role that compresses the strategic consultant, the software engineer, and the operational translator into one, with the handoff between diagnosing the problem and building the fix removed. FDEs can run a client workshop in the morning, map your operational workflows, and have a working application running by that evening. On your site, on your data, on your clock.

The value FDEs create isn’t in question. The diagnostic question most enterprises skip: where does the finished capability actually live? In a fine-tuned model with weights you own, that IP doesn’t leave the building. In a model trained against Palantir’s Ontology layer, the IP stays embedded in an architecture you don’t control. The asymmetry is the product.

The services firms are the implementation layer: first call for the enterprise, last in line for the margin. Each one is mid-pivot, building proprietary accelerators and productizing offerings as part of a broader transition toward services-as-software. In capital structure terms, it’s still the equity tranche. First to be diluted, most exposed to compression from above.

Nvidia collects regardless. Frontier labs collect while they can. Operating platforms collect on the workflow logic they’ve embedded. Services organizations absorb whatever margin is left. The capital structure isn’t neutral. For now, it’s a ranked ledger, and most enterprises don’t know their place in it.

The Real Risk: What They Learn While You’re Paying Them

The same mechanism that makes your workflows more productive also makes your expertise legible to the model vendor and the workflow sovereign. The actual risk isn’t visibility. It’s what they do with what they learn.

AI models scale private expertise into institutional leverage. The same sentence contains both the pitch and the warning label.

Cursor had built the coding assistant category from scratch, on Anthropic’s models. Those enterprise deals were Anthropic’s early API revenue. When Anthropic launched Claude Code directly into that category, its best customer became its most direct target. Then came Claude Science, Claude Legal, Claude Security, Claude Financial. Each one a category its own API logs had already mapped.

Anthropic’s Chief Product Officer had served on Figma’s board and didn’t resign until three days before Claude Design launched, a new vertical application competing directly in the design category. Figma’s stock has fallen roughly 50% this year while Anthropic’s valuation has surged. The CPO seat is reconnaissance with a product roadmap attached.

Frontier labs are scaling domain training data to build exactly this capability. The same model learns judgment patterns across hundreds of firms simultaneously, and no single company could replicate that signal volume independently. The enterprises running the deepest structural risk are the ones whose proprietary judgment, the operational heuristics that took decades to compound, is flowing into a shared model that their direct competitors also use. Which is why Anthropic’s roadmap for life sciences is framed as “tools for researchers today, more autonomous discovery over time,” ending in a unified Claude for Life Sciences workbench.

Your proprietary workflows give them something the API logs alone can’t. The operational blueprint for building the vertical that displaces you.

The contracts you’ve negotiated address data rights: retention, access controls, confidentiality, training opt-outs. Those cover Layer 1. What they don’t cover is Layer 2: the derived artifacts. Tasks, evaluation suites, workflow traces, correction signals, failure libraries, decision policies.

Even if your raw data never trains the model directly, once the vendor understands your hard problem and its solution patterns, they can recreate those training tasks with external experts. The model captures your intellectual capital in vendor-controlled form without technically violating your data terms.

Which Workflows Are Actually Worth Protecting

The question none of the regimes want you to ask is the simplest one: which of your workflows is actually worth protecting?

The procurement process is optimized for the vendor. Capability first, price second, exit terms in Appendix D that nobody reads until they want to leave. Before capability demos and price negotiations, someone should run an honest inventory of which workflows encode your edge and which ones don’t.

Most workflows are commodity. Any competitor running the same tools gets the same output. Route those anywhere: frontier APIs, shared platforms, whatever is cheapest and fastest. The regimes don’t care, and neither should you.

The dangerous workflows are the ones where the judgment is the product. Alpha generation. Underwriting nuance. Bespoke claims decisions. Drug discovery strategy. The operational DNA your competitors would pay anything to sequence. One test before any of this touches a shared model: would it matter if every competitor could see exactly how you handle it? If the honest answer is yes, it belongs on infrastructure you own. Off the frontier API, off the Ontology, on your hardware.

That pattern shows up most visibly in customer service: well-understood workflows, massive conversation volume, measurable quality bars. Also, the highest latency sensitivity of any AI use case in the enterprise. Which explains why the most mature deployments in this category run on small, fine-tuned, fast open-source models. The frontier model did its job at the exploration phase. The SLM inherited the production line.

The emerging playbook in mature enterprise deployments is frontier models to explore, open-weight SLMs to scale. The first phase is reconnaissance: you don’t yet know what you’re optimizing for, so renting broad intelligence makes sense. Once the use case is fully specified, inputs mapped, failure modes documented, behaviors locked, you’re paying for intelligence you’ve already extracted. A fine-tuned SLM does the same work at a fraction of the cost.

The verdict on companies still defaulting to frontier APIs for differentiated work is settling. Chamath Palihapitiya, founder and CEO of Social Capital and CEO of AI coding startup 8090 Labs: “If you are a reasonable company, why are you not finding an independent way to access this intelligence in a way that doesn’t leak your edge away?” He gives the ignorance defense one more cycle before pulling it. Back then you didn’t know. Now the data exists. To continue making the same decision anyway is, he says, “really insanely dumb.” That framing now has a number attached. Swap GLM 5.2 into the 8090 Labs harness on a modernization pilot: costs drop 16.4×, with a 3× speed penalty. You run the slower model and still come out 5.5× ahead per unit of work. First datapoints in a collapsing cost curve tend to look conservative in hindsight.

Multi-model stacks are arbitrage, not escape from the lock-in problem. Teams get optionality at the model layer, then watch switching costs accumulate in the orchestration layer as agents deepen into memory, tools, and state. The routing logic is sound. Route high-complexity work to frontier models, IP-sensitive work to sovereign deployments, commodity volume to fine-tuned SLMs. The orchestration layer below all of it is where the next regime quietly installs.

AI sovereignty means being able to swap what runs underneath without rebuilding what runs on top. The regimes can build the interface for institutional judgment, but they can’t supply decades of it. When you route that judgment through their rails anyway, you’re paying the tax and funding the next product that displaces you. The operational heuristics that took decades to earn compress into a training signal. That signal ships as a feature. Eighteen months later, your competitor demos it to your client.

A forward-deployed engineer works for the vendor. For most workflows, that arrangement is fine: the IP was never in the workflow. For the ones where the workflow is the IP, the person encoding that method into a sovereign open-source model deployment should work for the enterprise, not the vendor. Sooner rather than later.

Karp understands this, which is why he spent a CNBC interview telling enterprises to own the means of production behind their AI rather than rent them. Karp is right about the problem. The collection mechanism he built is the evidence. The three regimes describe the same capital structure from different angles. Token meters, Ontology layers, sovereign stacks: the instrument changes. The ledger doesn’t. Know your line on it.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

The most valuable thing most systems integrators own isn’t on their balance sheet. It lives in the relationship between a senior architect and a CTO who trusts them enough to say “you’re solving the wrong problem” at the start of a nine-figure AI program. That relationship is the advisory entry point. It’s how a services firm earns the right to architect, implement, and manage the stack that turns a single engagement into a decade of expanding revenue. It’s what gets a firm in the room before the AI program is scoped, not contracted after the vendor is chosen. PE-backed deal structures have been systematically underpricing that seat since the AI cycle began.

PE-backed acquirers accounted for 69 of every 100 disclosed North American SI and cybersecurity transactions in 2025, a 12.3% year-on-year rise across an SI and cybersecurity M&A market that otherwise contracted 11.1%. PE ownership is deepening across the services stack just as AI advisory relationships are becoming the scarce asset in enterprise technology. Those same deal structures are, by design, depreciating the advisory capability that generates the value PE paid to acquire.

Why the math works against advisory

From the outside, the advisory defunding decisions PE-backed SIs make year after year look like they’re sawing off the branch they’re sitting on. From inside the capital structure, they’re following the incentives exactly as written. Managed services revenue carries four to five times the enterprise value of advisory revenue in a sponsor’s financial model. At 10-plus times leverage, that ratio settles the annual advisory investment decision before the advisory case gets heard.

Architecture relationships in AI take 18 to 36 months to convert to downstream managed services pull-through. A PE fund on a five-year hold, with two years of post-acquisition stabilization already consumed, has roughly 24 months before exit preparation begins. Twenty-four months is not enough for AI advisory cycles to translate into the managed services ARR that justifies the original multiple.

Agentic AI breaks the math entirely. Early-stage agentic deployment produces variance. Failed iterations, architectural restarts, and months of experimentation accumulate before any recurring pattern holds. A typical agentic deployment runs six to eighteen months of negative or flat cash contribution. A lender’s covenant clock runs on a different calendar. So the engagements don’t get funded. The team shrinks through attrition. People who can run a board-level AI governance conversation in the morning and debug an agentic deployment in the afternoon don’t stay at firms that can’t fund the work.

The outcome is atrophy rather than any single decision.

The flywheel, and why it matters more now

Most sophisticated SI and cybersecurity firms operate through some version of the PEM flywheel, Omdia’s services capture model.

The flywheel has three components. P is the advisory and point-of-view layer: strategy engagements, risk assessments, architecture reviews, the board-level conversations where a firm establishes its perspective and earns trust with a CTO or CISO. E is execution: transformation programs, implementation projects, deployments. M is managed services: ongoing operations, platform contracts, and the recurring arrangements that lock in durable revenue year over year.

The sequence is earned. Advisory earns the right to architect. Architecture earns the right to implement. Implementation, executed well, earns the right to manage. Each loop through the cycle deepens the client relationship, adds proprietary operational context, and makes displacement progressively harder. In cybersecurity, for example, a well-run advisory engagement can anchor eight to twelve years of attached managed services revenue.

The flywheel still turns in AI. Missing the first rotation is now permanent. That conversation happens in board steering, where CEOs and CISOs set the architecture direction and governance mandate. The RFP documents the outcome. The advisory relationship shaped it. Firms in that conversation set the terms. Firms outside it respond to them, relegated to compete downstream on mandates defined by others.

Frontier labs noticed. Anthropic, with the most safety-trusted frontier model by enterprises, still built a separate deployment company to sit in client governance conversations. AWS committed $1 billion to match it. OpenAI built a $10 billion deployment vehicle to hold the advisory relationship from the start. Three separate bets, one shared premise: the advisory seat is worth paying for before the RFP exists.

The demand multiplier that doesn’t appear on the P&L

The Kyndryl trajectory makes the underlying economics visible from the opposite direction.

Kyndryl was spun out of IBM in November 2021 with a dominant managed services base and almost no independent advisory capability. IBM kept the consulting brand; Kyndryl took the contracts. The structural result was predictable: overall Kyndryl revenue has declined 8 to 11 percent annually since the spin-out. Managed services without advisory pull is a contracting business in the current AI transition environment.

What has grown is Kyndryl Consult, the deliberately rebuilt advisory unit, achieving double-digit revenue growth year-over-year while the base contracts. In November 2025, Kyndryl launched agentic AI advisory and implementation services covering workforce transformation, business process redesign, and AI experience design. The firm is spending to recreate something IBM took with it in 2021, because the managed services base degrades without the advisory pull that refreshes and expands it.

The recovery cost is real and the window is closing. Kyndryl is burning capital to rebuild what IBM’s demerger removed. That capital expenditure, tracked against the acceleration of OpenAI DeployCo’s forward-deployed engineer program across 2,000 PE portfolio companies, makes the timing problem visible and urgent. Advisory positions that were vacated by leverage-driven atrophy or formal divestiture are being filled right now by frontier-lab engineers doing exactly the work that atrophied SI advisory teams used to do.

The Optiv case: the flywheel unwound

The Optiv transaction, which closed two weeks before this essay was published, is the documented case of that vacancy happening at speed.

KKR acquired Optiv in February 2017 for a reported $2 billion, assembling one of the largest pure-play cybersecurity services firms in North America. The original value proposition was a working PEM flywheel: advisory credibility at large enterprise accounts, architecture relationships that led to implementation mandates, followed by long-term managed security contracts. At its peak, Optiv described itself as a firm that moved clients from “point of advice” through deployment into managed services. The full cycle.

By early 2026, the debt load had deteriorated to the point where lenders were controlling the exit timeline, not management.

S&P downgraded Optiv to B-minus in August 2023, explicitly citing lower demand for cybersecurity consulting as a driver of EBITDA underperformance alongside the leverage burden. By June 2025 the rating had reached CCC, with a primary loan of more than $1 billion maturing in August 2026 and no cash to repay it. Lenders extended the deadline to August 2028 in March 2026. Three months later, on June 1, 2026, Optiv completed the sale of its Advisory, Consulting and Transformation unit to Vobis Ventures, rebranded as Optiv Consulting, with approximately 500 consultants, access to Optiv’s six-thousand-client base, and a one-year exclusive services partnership.

Shedding project-based advisory creates a cleaner EBITDA base, improves revenue quality metrics for lenders, and removes the operational complexity of running a high-touch consulting practice on a leverage-constrained balance sheet. The firm that exits this restructuring holds a managed security services platform with advisory capability relocated to a separate legal entity under a different PE owner.

The divestiture is the visible event.

What the financial logic doesn’t price is the flywheel consequence. The exclusivity period sets a clock on Optiv’s access to the advisory conversations that historically originated its managed services demand. That access isn’t peripheral: it’s the mechanism through which the managed book gets renewed and grown.

The honest read is that this looks more like a well-described forced asset sale than strategic value creation. Vobis’s thesis is plausible on its own terms: Optiv Consulting may grow faster with focused capital than it could inside a debt-stressed parent. But competing for AI governance advisory mandates against Accenture, Deloitte, McKinsey, and OpenAI’s DeployCo once the exclusivity clock runs out requires capital depth and institutional weight that a carve-out under these circumstances hasn’t demonstrated.

The liability moat that advisory carve-outs destroy

Before June 1, 2026, a large enterprise client working with Optiv had a single counterparty: one P&L, one liability envelope, one relationship that ran from governance framework design through managed security operations. After the divestiture, the same client has two separate legal entities under different PE owners, bound by a one-year exclusivity arrangement that was negotiated under KKR’s debt constraints rather than the client’s risk preferences. Optiv Consulting designed the AI governance architecture. KKR-owned Optiv runs the managed security operations. When an agentic AI deployment fails a governance test under the EU AI Act, the question of which entity accepts contractual responsibility doesn’t have a clean answer. The EU AI Act’s high-risk system requirements, active as of August 2026, create director-level personal liability for exactly these failures. The advisory firm that designed the framework is one company. The managed services firm running the operations is another. The enterprise buyer absorbs the liability either way.

Advisory carve-outs fragment the liability chain connecting governance framework design to managed operations. The enterprise buyer inherits that gap without being consulted about it.

A firm offering both advisory and managed services under a single liability envelope, accountable for governance framework design and for the operations running against that framework, holds a structurally differentiated position relative to carved-out advisory units competing for the same mandates. Optiv Consulting, however capable its 500 consultants, cannot offer that package to a prospective client. The managed security operations that validate the governance framework are now a different company. The flywheel coming apart at P ends with this: the firm can no longer hold both the governance framework and the operations running against it inside a single liability promise. Integrated competitors still can.

The question worth asking now

The capital structure doesn’t care about the advisory seat. It cares about the multiple. Every SI and channel firm under PE ownership is navigating that gap right now, and many of them are resolving it the same way Optiv did.

PE capital funds infrastructure, tooling, and platform buildout. The advisory seat, the relationship that earns a firm the right to say “you’re solving the wrong problem” before the architecture is set, is the one asset it systematically underprices. That seat drives a decade of downstream AI-native services revenue. Which is exactly why it keeps getting cut.

The $267 billion that Omdia forecasts for partner-led AI and agentic AI services by 2030 has two versions. One accrues to firms with durable advisory positions at the architecture and governance layer, supported by managed services that those advisory relationships generated. The other accrues to whoever holds the FDE relationships and the model contracts. Most of the choice between those versions is being made right now, in deal structures and operating priorities, by people for whom the next margin program looks like a routine annual exercise.

You maintain and grow the capability that earns the advisory seat, or you become a highly efficient execution platform inside someone else’s architecture. Both are viable businesses, for now. Only one of them has a credible path to the $267 billion when it actually materializes.

The most valuable thing most systems integrators own isn’t on their balance sheet. The question is whether the structure they accepted to compete is the same one quietly repricing it.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

The AI at UnitedHealth Group reads aloud.

As nurses drive to patients’ homes across the country, an artificial intelligence system summarizes their medical charts out loud through the car speakers, so they can absorb the patient history while navigating suburban traffic. A separate AI listens to millions of customer service calls, identifying patterns in complaints before a human analyst ever sees them. A third is calling doctors’ offices to schedule appointments on patients’ behalf.

A $3 billion AI program looks like this from the inside, according to Bloomberg’s account from the company’s headquarters outside Minneapolis last week. The most striking thing about it is where the technology is deployed, not the technology itself. The largest, most expensive, most committed enterprise AI program in American healthcare is doing administrative work. Paperwork. Routing. Scheduling. Data reconciliation. Chart summarization. Almost ninety-nine percent of the company’s more than 1,000 AI applications in production are administrative rather than clinical. The company has committed publicly to staying out of diagnostic AI entirely. That is the back-office-first sequence at $3 billion.

Sandeep Dadlani, CEO of UnitedHealth’s Optum Insight division, crystallized the posture in a single line: “We only approve using AI. We never deny using AI.”

It’s a product decision, a governance commitment, and a crisis strategy compressed into ten words. Packed inside it are two enterprise AI lessons most organizations learn in reverse. The first: back-office work is bounded, process-mature, and auditable, conditions where returns are fastest and most attributable, generating the margin and the institutional credibility that harder territory demands. The second lesson is less obvious and more consequential: starting in the back office is the earned route into territory where mistakes can’t be walked back and the error rate doesn’t stay internal.

UnitedHealth is demonstrating both lessons simultaneously, one by design, the other by litigation.

The Misallocation

In 2026, fifty percent of enterprise generative AI budgets flow to sales and marketing. That’s from MIT’s NANDA study, 150 senior-leader interviews, 350 employee surveys, and 300 public AI deployments. ISG’s analysis of 1,200 enterprise use cases is unsparing. Revenue growth is simultaneously the most-funded AI outcome category and the most underperforming one. Compliance, cost reduction, and quality improvement consistently outperform expectations while receiving the smallest share of investment.

In 2026, the gap between where AI investment flows and where returns actually land is the defining misallocation. Capital follows excitement. The P&L follows process maturity, cost baselines, and recoverable failure modes, and that gap is the misallocation.

The structural ancestor for this misallocation predates every technology company in the Fortune 500. It’s a World War II trucking circuit.

After D-Day, the Allied advance stalled not for lack of material. American factories were producing tanks, ammunition, and fuel faster than any army in history. The bottleneck was distribution. The Red Ball Express, a round-the-clock trucking circuit running between the Normandy beaches and the front lines, consumed 300,000 gallons of fuel per day just to move fuel. The distribution infrastructure couldn’t keep up with the supply. The front starved while the ports overflowed.

Enterprise AI is caught in the same circuit. Model capability has decoupled from what the P&L reflects. In 2026, distribution infrastructure means three things: clean data at production scale, processes structured for AI, not stitched around it, and the organizational will to govern before deploying at scale. That last one is the hardest to build and the easiest to underfund.

The back-office advantage isn’t intuitive until you look at the structural properties that drive it.

Four structural properties explain why. First, they run at high volume with low variance: the same data structures, the same decision logic, the same exception patterns, processed at a scale that generates training data in abundance. Second, they have documented operational benchmarks and cost baselines before any AI program starts: FTE counts, cycle times, error rates, all already in the operational record. Third, attribution isn’t contested at the end. An invoice either matches the purchase order or it doesn’t. A claim either satisfies the adjudication rules or it doesn’t. No human judgment is required to know which. And fourth, errors are internal, recoverable, and low-legitimacy-risk. A miscategorized claim triggers a reconciliation flag. A miscategorized patient does not.

Those four structural properties explain the back-office AI ROI premium. They don’t explain why the sequence itself isn’t elective today.

The constraint for reliable agentic outcomes is runtime context. Agents need durable enterprise decision memory (what was decided, why, under what conditions) to reason across process and system boundaries without re-deriving context at each handoff. That decision traces and context graph infrastructure doesn’t yet exist.

Back-office schemas have accumulated business logic, exception handling, and operational history in structured form. Front-office workflows hold that institutional knowledge still in tacit human judgment, unstructured and unindexed. That asymmetry is what makes the back-office-first sequence a design preference today, not just a risk preference. It relaxes as the architecture catches up.

The $3 Billion Program

UnitedHealth is the world’s largest health insurer and fourth on the Fortune 500. It’s spending $3 billion on AI between 2026 and 2027. UnitedHealthcare CEO Timothy Noel has put a $1 billion operating cost reduction target on 2026, most of it AI-driven.

These are management targets, not audited results. The H2 2026 and full-year 2027 earnings calls are the first external stress test of whether the economics actually moved. The operational infrastructure behind those targets is unusually specific.

The healthcare administrative stack is a 1970s fax infrastructure running inside 2026 cloud billing. Prior authorization happens by phone. Pharmacy approvals take eight hours. Claims settle over days. Optum Real has converted multi-day settlement into synchronous system-to-system transactions: one billion processed since launch, 2.5 billion targeted by year-end 2026, manual contact costs down seventy-six percent on the workflows it covers. Digital Auth Complete, connected to more than 250 systems, achieves a ninety-six percent first-pass prior authorization approval rate, with more than ninety percent of requests resolved within one business day. PreCheck MyScript cuts pharmacy authorization time from over eight hours to under thirty seconds.

The production application count doubled in a year, from 500 to more than 1,000. That pace matters less than the composition.

Tier 1, the dominant share, is administrative and claims work at low legitimacy risk and high volume: Optum Real, Digital Auth Complete, PreCheck MyScript, fraud detection, billing code assistance, call routing. Tier 2 is clinical-adjacent, covering chart summarization for nurses, AI transcription, population health analytics, and customer call analysis, where clinical data is processed but clinical judgment isn’t replaced. Tier 3, near-zero share by design, is the avoided territory: diagnostic AI, AI-driven coverage denials. That distribution reflects a governance posture that drew a hard line between administrative optimization and replacing clinical judgment, and held it. The result: 99% administrative. The budget confirms the posture: two-thirds internal operations, one-third external products and platforms.

This is also a company in turnaround mode. The $3 billion AI commitment runs through 2025’s collapsed profits, the violent death of a CEO that crystallized public frustration with denial practices, ongoing class action litigation, and active Senate investigations. For an institution in active legitimacy repair, the back-office-first posture is the architecture. The margin strategy and the trust strategy are the same program.

Make no mistake: this is also a platform play, not just a cost reduction program. UnitedHealth is laying operations infrastructure that healthcare competitors and providers cannot ignore without incurring a widening cost gap. Administrative efficiency, at scale, is the edge.

The Trust Constraint

In 2024, insurers denied 4.1 million Medicare Advantage prior authorization requests. Twelve percent of denied patients appealed. More than eighty percent of those appeals won.

Medicare Advantage is the US government’s privatized senior healthcare program, in which private insurers like UnitedHealth administer federally funded coverage for adults over sixty-five. Prior authorization is the requirement that a doctor obtain insurer approval before delivering certain treatments, a gatekeeping mechanism that processed roughly 53 million requests in 2024 alone.

The math is devastating. The eighty percent appeal win rate applies only to the twelve percent who appealed. The rest weren’t correctly denied. They were the people who lacked the capacity, the time, or the professional support to contest an algorithm’s judgment. They just didn’t get counted.

That algorithm is nH Predict, UnitedHealth’s AI tool for processing prior authorization requests across its Medicare Advantage population. A federal inspector general report found above-peer denial rates among companies running the algorithm. Most of those denials reversed on appeal. A federal judge found that serious enough to allow core claims against nH Predict to continue. UnitedHealth maintains the algorithm was used appropriately, in line with regulatory guidance at the time. The guidance has since hardened. CMS now requires individual clinical grounding for any coverage decision: the patient’s record, physician recommendations, documented history. Population-level statistical output isn’t, by itself, authority to deny.

nH Predict is a high-legitimacy-risk application deployed at scale, on the population least capable of exercising the appeal rights that would have surfaced the error rate. The governance architecture that should have been built and tested in bounded back-office territory, where errors cost reconciliation flags rather than wrongful denials, was never there. The federal case is in Minnesota. The $3 billion program is running now.

The sequence is the strategy. Getting it wrong doesn’t just cost returns; it costs the right to operate in the territory at all.

The Order of Operations

The order of operations is the constraint, not a preference. Two questions drive the sequence.

The first question is about ordering: which use cases make the sequencing cut? Business value sets the floor; legitimacy risk determines the ceiling on how fast you can move. The second is about composition: are you funding the margin story or the valuation story? In 2026, most enterprise portfolios are overweight on board-visible, customer-facing AI while the back-office margin opportunity sits underbuilt.

Each matrix is a separate tool: the first sequences, the second reallocates.

The sequence below is the operating plan the matrix implies.

1. Months 0–12: Workhorses only. Portfolio composition runs predominantly Workhorses: Optum Real, Digital Auth Complete, PreCheck MyScript anchor the tier. No Strategic Bet programs until governance infrastructure is proven and the Workhorse layer is generating early ROI signals. Table Stakes are worth running in parallel; they just don’t anchor the economic case.

2. Months 12–24: One or two Strategic Bets, under separate governance. Independent validation, mandatory human-in-the-loop decision points, and appeal or overturn rate tracking as a leading indicator of legitimacy failure. These programs shouldn’t share a budget cycle with the Workhorses, because doing so creates pressure to cut the governance corners they require.

3. Month 24+: Balanced portfolio. Workhorses as the steady-state margin engine, selective Strategic Bets as future-facing investments. The governance track record earned in stage one is the credential for what comes next.

That sequence is the argument. It’s what makes the harder territory possible rather than aspirational.

The second question, about composition, has its own geometry.

Most organizations in 2026 are over-invested in the Showroom quadrant: fast-tracked, customer-facing AI with high narrative appeal and board visibility that demos better than it performs financially. The use cases enterprises actually need to saturate first sit in the Margin Engine quadrant: high-volume, rules-governed back-office workflows where structural savings are immediate, verifiable, and track directly to cost takeout and operating profit.

Front-office AI can work, but only in the Commercial Engine quadrant, where revenue attribution is measurable through controlled experiments, pricing optimization, recommendation engines, intelligent cross-sell. Outside that specific zone, front-office AI lives in the Showroom quadrant or in Perceived Theater quadrant: broad copilot deployments, typically internal productivity tools, with high adoption metrics and low P&L movement, where the value is largely reputational rather than financial.

The second matrix is also a reallocation tool. CFOs and CIOs can map the current AI portfolio onto the four quadrants, calculate the share of budget sitting in the Showroom quadrant and Perceived Theater quadrant, and build the case for moving it toward the Margin Engine quadrant. The exercise is clarifying in ways that budget conversations rarely are.

Most leadership teams assume their portfolio skews toward the Margin Engine quadrant. The map rarely confirms that assumption, which is why the exercise is worth running before the CFO asks.

Deploy where failure modes are recoverable first. The Red Ball Express worked because the distribution infrastructure was built and stress-tested on the supply lines before the front lines depended on it. Back-office AI is the same logic: bounded territory where a mistake costs a reconciliation flag, not a wrongful denial, and where you learn what governing AI actually feels like before the stakes turn irreversible. The financial returns arrive faster than most executives expect. The organizational muscle that comes with them is what makes the next stage possible.

Earn the track record before asking for harder territory. A thousand bounded applications that caught their own errors, logged the overrides, and produced real before-and-after baselines. That’s the credential. CMS, the EU AI Act, and the CFPB are all asking the same question when they arrive: did you govern this responsibly in recoverable territory first?

Approach the harder territory with the track record in hand. Clinical decision support. Coverage automation. Credit underwriting. Each regulatory body governing those domains is asking the same question in different vocabulary: can you demonstrate responsible governance in recoverable territory before deploying where consequences are irreversible?

Arrive with that history and you get a negotiation between parties who both have evidence. Arrive without it and you get nH Predict.

The Beachhead

Andy Grove, describing strategic inflection points, observed that organizations face the most acute internal resistance precisely when the evidence for change is clearest and the emotional case for continuity is strongest. The established capability has a track record. The boring new capability has data pointing in its direction and no history of being right yet, which is exactly what makes defending the reallocation feel unreasonable even when the numbers clearly support it.

UnitedHealth AI’s strategic inflection is unfolding in the back office, one claims reconciliation and one invoice match at a time. The organizations that recognize this and redirect capital accordingly will look unremarkable for the next eighteen months. Then, quietly, decisive. The ones that keep funding the showroom will have excellent product demos and increasingly difficult conversations about why the economics haven’t moved.

That window is architectural as much as organizational. The back-office-first sequence is a necessity while decision traces and context graphs remain an unsolved infrastructure problem. When the architecture catches up, the sequence graduates from constraint to design preference. The organizations that spent the intervening period building the beachhead will hold both the governance track record and the production infrastructure to move first into the territory it unlocks.

UnitedHealth is spending $3 billion to prove the beachhead thesis by 2027. The nurses listening to chart summaries in their cars are what AI operations look like right now at a company that has decided to earn trust the only way trust can actually be earned: through administrative accuracy, operational transparency, and the disciplined restraint to stay out of decisions that belong to physicians until the governance architecture is genuinely ready for them.

I keep coming back to Dadlani’s ten words. We only approve using AI. We never deny using AI. Read as policy, that’s just a permissive stance on adoption. Read as sequencing commitment, it’s something else entirely: a promise to build the governance infrastructure before entering the territory where approval and denial carry real consequences.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

Uber’s annual AI tools budget was gone by April. When asked to explain it, the COO said he could not. Four months into a twelve-month cycle, five thousand engineers, per-engineer monthly costs between $500 and $2,000.

Pricing infrastructure that runs 26 million trips a day, audit infrastructure that cleared a 2019 IPO and seven years of public-company reporting. None of it could name a single outcome the budget had bought.

The invoice arrived. The explanation didn’t.

What Uber’s COO put on the record, dozens of finance and technology leaders are saying in rooms that don’t make the news.

It was built before anyone asked what it was for

The environment had an architect. The architect had a cap table. Labs with S-1s in progress built it. Metered consumption meant every mandate-driven deployment was a revenue event for the lab, regardless of what it produced for the customer. Every token consumed appeared on a revenue run-rate chart that supported a pre-IPO valuation story. The enterprise P&L absorbed it as overhead. The lab prospectus counted it as traction. The lab wrote the contract. Governance, cost controls, and value attribution appeared nowhere in it.

Underneath the revenue event, a training event. Every time a developer fixed what the model got wrong, that correction went back into training data. The model learned from the mistake at the enterprise’s expense. Multiplied across tens of thousands of engineers, that signal was worth more than the subscription. Acceptance rates, edit patterns, what engineers consistently fix: the model learns from all of it. Enterprise tiers contractually protect proprietary code from retraining datasets. They don’t protect the behavioral signal. The enterprise absorbed the cost of the mandate. It also paid for the next version. The lab wrote that contract, too.

Product bundling made the lock-in contractual. Microsoft embedding Copilot into M365, Salesforce embedding Einstein into CRM, Google embedding Gemini into Workspace: none required a new purchase order. The quarterly business review filed it as adoption. The CIO reports 80% adoption because 80% of users have no path around it. From the boardroom, it reads as progress. From the invoice, it reads as a cost with no outcome attached. Switching isn’t a product decision. Renegotiate cloud credits, redo the security review, rewrite compliance documentation, retrain thousands of engineers. “If you’re already on Azure, AI credits are a line item approval. A new vendor is a procurement program,” said a senior account director at a major GSI. The decision didn’t happen at deployment. It happened at renewal, two years earlier.

Frontier model API costs fell 67% year-over-year. The decline pushed AI pilots below the procurement thresholds that trigger governance review: business cases, vendor assessments, projected ROI. Lower costs produced less oversight, not more. Legal ran into the same gap from the compliance side: no framework covered AI-generated outputs, so reviews cleared deployments they hadn’t been designed to evaluate.

The mandate accelerated into a vacuum that was already there.

The mandate, the lab, the rational employee

The first engine was the mandate machine. SaaS had trained leadership to read adoption as evidence. It left out the part where adoption is what runs the meter. At near-zero marginal cost per seat, high volume and high adoption were the same claim. AI never had that physics. Every SaaS seat you don’t use still costs the same. Every AI query you do use costs more. Running a frontier model on a task an open-weight model handles adequately is commissioning a surgeon to apply a bandage.

“Your competitor is already doing this.” The conference circuit ran that message on a loop. A vendor relationship became a pilot. The pilot became a metrics deck, reported upward as evidence of transformation. The word “outcome” rarely sat next to the number it was supposed to justify.

The mandates, when they arrived, were coherent within their own logic. Duolingo tied AI tool usage to performance reviews in April 2025. Set a target and people will hit it. Usage went up. Twelve months later, the distinction the mandate had quietly erased became unavoidable: using AI is not getting value from AI.

The second engine was the lab pipeline. Frontier labs got paid per token. Their enterprise customers counted seats. Neither side had an incentive to name it. Pilot, announcement, renewal: three gates, zero outcome requirements. Every stage checked for usage. None checked for value. That asymmetry was designed in, not accidental.

The third engine was the rational employee. Tokenmaxxing. The behavior had earned a name. Amazon put AI usage on a leaderboard. Engineers gamed the leaderboard. Amazon shut it down. Engineers routed tasks through AI tools just to generate the log entry. The log entry was the point. Code quality wasn’t the metric. The flag was. The numbers didn’t lie. The scorecard did.

In McKinsey’s 2025 survey, 88% of organizations reported regular AI use. Fewer than four in ten reported any EBIT impact. That’s a 49-point gap between reported use and reported impact. Three engines running simultaneously, nothing to interrupt any of them. The invoice told a different story. Ramp’s May 2026 data put AI spending at 13x its January 2025 level. Not 13%. Thirteen times. 42% of companies abandoned most of their AI projects in 2025. In 2024, 17% did. The 13x spend growth and the abandonment spike are the same story from two angles.

Uber’s story got out. Most haven’t.

The invoice arrived without options

Eighteen months ago this meeting wasn’t on anyone’s calendar. Now it’s in every large company. “AI spend review.” “Productivity initiative alignment.” The AI leader arrives with a deck of green metrics. The CFO arrives with one number.

Token spend is metered compute. Adoption drives it up. Poorly scoped prompts compound it. Agentic workflows without cost controls run it off the chart. Average monthly enterprise AI spend hit $85,521 in 2025, per CloudZero. A year earlier it was $62,964. The teams signing those checks had no mechanism for reading whether they were paying for outcomes or paying for activity. A faster meter is not evidence of a better result.

A poorly scoped agent doesn’t make one bad call. It makes a second call that trusted the first, and a third that trusted the second. The cost multiplier is real: one agent call can spawn dozens of downstream API calls. The cost error is recoverable. The logic error isn’t. A wrong assumption at step one becomes the foundation every subsequent call reasons from. Most agentic deployments have no circuit breaker, no cost ceiling, no stopping condition. The chain runs until it finishes or fails.

The mandate could be framed. The usage metric could be explained. The vendor relationship could be managed. The invoice arrived with none of them.

The Adoption Mirage maps where most organizations are versus where they say they are. High usage with low measured value isn’t a failure of individual teams. It’s the predictable output of an environment where the meter runs and the scorecard doesn’t.

How the invoice became architecture

Nobody had authority to turn it off. More important: nobody wanted to. The discipline governance couldn’t produce, the invoice eventually forced.

Cost pressure built its own governance. Engineering teams started routing different task types to different models based on complexity, consequence, and cost. Three tiers, in practice. Tier one: routine, high-volume tasks routed to cheaper open-weight or smaller proprietary models. Tier two: reasoning-intensive, high-stakes work that stays on frontier. Tier three is where bigger loses: domain-specific models fine-tuned for legal, financial, or clinical work outperform frontier models on the tasks those sectors judge most carefully.

The tiered routing capability is a leading indicator. Most enterprises are still early-stage or at the stage the Uber story describes: spend spiked, no mechanism to attribute it, no architecture to respond.

One bank’s infrastructure team discovered they’d been running a frontier model on document classification, routing loan applications to the correct processing queue. An open-weight model fine-tuned on six months of their own labeled data handled the same task at a fraction of the cost, with a lower error rate on their edge cases because it had seen them before. The tier didn’t exist until a VP of Engineering built a spreadsheet to answer a question the CFO had asked in a meeting the week prior. The spreadsheet became policy. The policy became architecture.

Every routing decision turned out to be a governance decision.

How much reasoning does the task actually need, and what does a wrong answer cost downstream? What error rate is the workflow allowed, at what per-outcome cost? Is the output a trigger or a recommendation, and does a human sit between it and the consequence? They were governance questions. Finance inherited them. Finance derived the answers under budget pressure. The discipline arrived. It arrived through the wrong door.

The Klarna story, read correctly

In early 2024, Klarna announced its AI system was doing the work of 700 customer service agents. Eleven months later, it was rehiring. Everyone asked whether the AI worked. Nobody asked how they’d know. The measurement didn’t.

No baseline, no observability, no protocol for when the handoff reversed. A measurement architecture would have caught the drift at 5%. At 22%, you’re calling the agents you laid off.

Every postmortem blames the model. None of them audit the measurement.

What that infrastructure looks like when it’s missing: no ticket resolution baseline from the six months before the AI launched, no drift threshold that triggers review, no defined protocol for when the system hands back to a human and under what conditions. The AI ran. The numbers moved. The definition of “working” was the one thing nobody had written down. So the drift had nothing to register against.

If the problem is the technology, the response is vendor selection or capability improvement. If the problem is measurement, the response is infrastructure that existed before the deployment rather than after.

The Klarna rebuild that followed illustrates the difference. Human agents handle the edge cases AI still manages poorly; AI handles volume where it performs well. Routing logic determines the allocation. Measurement infrastructure tracks whether it holds. The rehiring wasn’t a reversal. It was a more careful version of the same strategy. It arrived through failure, but it arrived.

Three habits that close the gap

The gains are real. They’re also narrow: concentrated among power users and a thin band of high-fit workflows. When you add review time and error-correction labor to the cost model, generalist deployments frequently go flat. Some go negative. The 40% of senior agent time consumed by escalations doesn’t appear in the productivity report. The 70% ticket resolution rate does. The productivity gain on code commits is measurable because someone built a baseline before the work started. Defect rate reduction requires a baseline that didn’t exist before deployment.

The practical question is whether to act on the measurement or honor the announcement.

Three habits separate these firms from the ones still managing the invoice.

The first: they decide before they deploy. Which tasks justify frontier compute, which route to cheaper models, and what cost-per-outcome the workflow can absorb, written down before rollout, not reverse-engineered from the invoice.

The second: the metric is always the outcome, not the event. Cost per resolved ticket, throughput per workflow, error rate by task type. Weekly active users and AI-flagged commits measure that AI ran. The firms across the gap measure whether it was worth running.

What the absence of this looks like: a team running an AI-assisted contract review workflow quietly retires the error-rate metric after the first quarter because the numbers are inconvenient. The deployment continues. Time-saved replaces it. Time-saved doesn’t measure whether anything improved. Nobody notices that the quality signal disappeared. Not until a clause gets missed and the question of what the system was actually doing for nine months becomes very hard to answer.

The measurement debate skips a person entirely. Not the CTO who owns the deployment. Not the CFO who holds the invoice. The team lead who sits between them, approving AI-assisted outputs every day without a quality baseline to measure against. They review outputs they can’t fully verify, sign off on workflows they weren’t trained to audit. The gap between the deployment narrative and the task-level reality lands on them. The Klarna drift happened in that layer, month by month, because proximity to the output is not the same as visibility into it.

The third is harder to build than either of the first two: an environment where a bad AI result is data, not a career problem. The reason it’s hard to construct is a power problem presented as a measurement problem. Where AI runs as mandate, honest measurement reads as dissent. In the organizations that clear the bar, one thing holds: ownership of AI ROI sits with the business leader responsible for the outcome, not the technology leader responsible for the deployment. Not the CTO. The person whose P&L the AI is supposed to move. When technology leaders own AI ROI, the metric is deployment. When business leaders own it, the metric is the P&L line the deployment was supposed to move.

What these firms are assembling is an AI FinOps discipline. The firms that build the practice now get to choose when the question gets asked. The ones that don’t will wait for the invoice to ask it.

The 2×2 shows why the FinOps gap is designed in, not accidental. Switching the pricing model without switching the measurement model moves the invoice, not the problem. The number climbs. The explanation doesn’t.

The measurement problem is what’s left

A Fortune 50 company with sophisticated finance infrastructure spent its entire annual AI tools budget in four months and couldn’t name the business value it had bought.

The measurement environment couldn’t distinguish genuine productivity from its theater. The invoice didn’t wait.

Discipline follows the invoice. It always has. The first FinOps practitioner wasn’t a cloud architect. It was a CFO holding an AWS bill nobody could explain. It took a budget crisis to produce a budget discipline. AI governance is following the invoice. Finance teams are building it inside an environment vendors had no incentive to make governable, because no one else had a number that forced the question.

The organizations that build the shortest chain from an AI decision to a business outcome, and develop the earliest capacity for reading when that chain breaks, win the phase that follows. The adoption question is closed. The measurement question is what’s left.

The organizations that reframed it as a measurement problem are making progress. The ones treating it as a communications problem are managing the story, not the problem.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

The enterprises outperforming their peers on early AI returns didn’t earn their advantage in the current AI cycle. Years of investment in product-centric operating models had already shifted their unit of accountability from function to workflow.

For decades, the function was that unit. Finance optimized Finance. Supply chain optimized supply chain. It was the right design. AI exposed its limits. AI creates value at the crossings: the order-to-cash workflow that runs across sales, finance, operations, and supply chain; the hire-to-retire cycle spanning HR, technology, and finance; the fraud detection loop no single department owns. The function-led model was never designed for the crossings.

IBM’s 2026 study of 2,000 CEOs quantifies that mismatch: 97% of executives say AI benefits them individually. Twenty-nine percent see significant organizational ROI. The 68-point gap between those two numbers is what the function-led operating model costs.

The structure failed before the AI arrived

Jacek Olczak, CEO of Philip Morris International, framed it in IBM’s Global C-suite Series, 2026 CEO study: “Trying to take AI tools and squeeze them into the existing organization is extremely likely to be the wrong approach. The technology is not the limiting factor. We are the limiting factor.”

Three patterns explain Olczak’s constraint. Each one is a product of the same boundary the function-led model was designed to hold. The design that organized the work also organized the failure. That’s what keeps the gap opaque by design.

The handoff problem. Every high-value AI use case crosses at least one departmental boundary. Most cross three or four. A supplier failure signal that sits in procurement cannot automatically trigger inventory reallocation in operations, rerouting in logistics, or a demand forecast recalibration in planning. The operating model treats each of those four as a separate accountability. That’s the design. That’s the problem. A pricing decision that has no visibility into real-time inventory. A customer resolution that cannot access billing history. Each of those handoffs is a governance negotiation. AI agents cannot orchestrate across boundaries the operating model was built to enforce.

The continuity problem. Project-based investment funds deliverables, not capability. Teams assemble for a project, deliver to a milestone, and disperse. They accumulate knowledge no handoff survives: data structure, model behavior at the edges, failure modes the training missed. When the project closes, the AI model is mothballed at the point of delivery and begins to drift. What gets handed over is the system. What doesn’t is the context to run it. The operations team is monitoring production logs for failures they were never briefed to expect.

The measurement problem. Output measurement is the function-led model’s native language. AI made it cheaper to produce more of it. Seat licenses became tokens consumed. Productivity reports became hours saved. Individual output metrics run in real time. End-to-end outcomes mostly go unmeasured and unattributed because no function owns them. The underperformance is invisible by design.

“82% of leaders say siloed functions block value autonomy delivers. Implication: businesses that keep them in place will stifle performance and limit the value autonomy is designed to deliver”– IBM IBV’s ‘The blueprint for agentic operations’ study

The gap deepens quietly. By the time the numbers surface, the investment cycle has closed.

Quiet doesn’t mean uniform. The failure has a position. Where it sits in the operating model determines what it costs and what closes it. That’s the map.

Where you sit determines what the gap costs you

Picture a grid. Two axes. Four quadrants.

The horizontal axis measures Accountability Span. On the left, accountability stops at the function. Finance owns Finance. HR owns HR. On the right, accountability spans the full workflow. Someone owns the outcome from start to finish.

The vertical axis measures AI Embeddedness. How deeply AI runs inside your enterprise’s actual work. At the bottom, AI is a personal productivity tool. Individuals use it. The enterprise does not. At the top, AI is running inside workflows, coordinating across functions, and accumulating value over time.

Each quadrant is a position. The fit is usually obvious. What to do about it is not.

Siloed Executor

Bottom left. Function-led structure. AI at the margins.

Work is organized by department. Budgets follow functions. Someone in marketing uses AI to draft copy. Someone in finance uses it to summarize reports. No one is accountable for AI delivering a business outcome.

Most enterprises in this quadrant have approved AI tools, run awareness campaigns, and published responsible AI policies. Nothing in the operating model creates the conditions for AI to move beyond the individual.

AI in this quadrant is additive at best. It makes individuals faster. Value captured here does not persist. It evaporates when the individual closes the laptop.

The priority for Siloed Executors is a governing decision: which two or three workflows matter enough to assign real outcome ownership. That is the first move.

Tool-Enabled Silos

Top left. Function-led structure. High AI activity.

The priority here is identifying the workflows where handoff failures are costing the most, and building the ownership structure to bridge them.

These enterprises have invested in AI. Agents are running inside finance, inside HR, inside IT. Dashboards show impressive adoption numbers. Token maxing is widespread. The AI budget is growing.

And yet the ROI remains stubbornly local. Each function has optimized its own slice. Finance is faster at closing. HR is faster at screening. IT is faster at resolving tickets. But the order-to-cash workflow is still slow. The hire-to-retire process still breaks at the handoffs. The customer experience is still fragmented.

This is sophisticated fragmentation.

The trap for Tool-Enabled Silos is mistaking activity for progress. More tools will not bridge the handoffs. A workflow-led model will.

The priority here is identifying the two or three cross-functional workflows where the handoffs are costing the most, and building the ownership structure to bridge them.

Workflow Aspirants

Bottom right. Workflow-led structure. AI still finding its footing.

This is the most promising quadrant on the map. These enterprises haven’t solved AI yet. They’ve solved the harder problem first.

The cross-functional collaboration that defeats most AI programs is not a deliberate effort here. It is the operating model. These enterprises have already organized around value streams rather than departments. Someone owns the outcome end to end, not just their functional slice.

AI is in the room but not yet in the walls. Someone owns the outcome. The workflows are clear. The AI is still finding its place inside them.

The risk in this quadrant is impatience. The AI activity looks sparse. The investment looks smaller than peers. They are not behind. They are better positioned. Once AI gets embedded here, the structure does what the function-led model never could. No functional boundaries to stop it. No ownership ambiguity to slow it down.

The priority for Workflow Aspirants is acceleration, not redesign. Find the workflow with the highest decision frequency and embed AI there first. The operating model will do the rest.

Product-Centric Orchestrators

Top right. Workflow-led structure. AI deeply embedded.

Most enterprises pick one. Structure without AI, or AI without structure. This quadrant requires both. Workflow ownership gives AI a container. AI gives the team leverage no function can match.

AI agents handle the repeatable, high-volume decisions inside guardrails the workflow team owns. The role is ownership. The workflow accumulates. Every decision adds to what the system knows about the next one.

That accumulation is the structural advantage. No function-led model can replicate it. The function owns outcomes up to its boundary. The workflow team owns them end to end.

The risk in this quadrant is not falling behind. It is outrunning the governance that makes the advantage durable. The enterprises that sustain it will be those that treat governance as a capability, not a constraint.

The priority here is to compound deliberately. Protect the structural advantages that got you here. Define the boundaries of AI autonomy before speed pressure defines them for you. The other three quadrants are stages, not destinations. The path here runs through the product-centric operating model.

The design decision that collapses the 68 points

What separates Product-Centric Orchestrators from the three quadrants beneath them is a design decision about the operating model.

One large global energy company redesigned its digital technologies function around a structural inversion: business and IT no longer sat on opposite sides of a request. They sat on the same platform team. Digital platforms were joint ventures. Product Line Managers and Product Owners came from the business. The technical depth came from capability chapters: engineers, architects, and delivery specialists embedded into each platform. Business drove the “what.” IT owned the “how.” The connection between business priority and technology execution was not a negotiation. It was a team.

They also own it after deployment. They monitor model behavior, catch distributional drift, and retrain against production data. There is no handoff event. The accountability that begins at conception runs through the full lifecycle, and the people who built it are still the ones running it when it breaks.

How ownership replaces negotiation

Customer Insight Groups were standing forums connecting business leaders, functional managers, and end users directly to the platform teams building against their priorities. They operated at three levels: twice-monthly at the product team level, quarterly at the product line level, and twice-annually at the platform level. Business did not submit requests and wait. It participated continuously in shaping what the platform worked on next.

Teams did not assemble and dissolve. Platforms evolved. Investment was allocated at the platform level, across cycles, rather than to discrete projects. The teams that built the platforms also ran them. Build and operations were not separated. The people who understood the model’s training assumptions were the same people monitoring production behavior when it began to diverge. The data accumulated by a persistent platform team running an AI system for 18 months is not comparable to the data accumulated by a project team that delivered to a milestone and moved on. Performance was not measured by output. It was measured by business impact, captured in Objectives and Key Results at the platform level.

That structure kept delivery priorities and capability development organizationally separate. Digital platforms owned delivery priorities. Discipline-based chapters owned capability development and talent standards. That separation meant an AI deployment decision didn’t require sign-off from a function that didn’t own the outcome. Quality standards held consistently across platforms through the chapter structure. Each platform kept its own product autonomy. Investment decisions followed the same logic: a three-layer prioritization architecture, with platform-level teams governing their own allocation, mission curation teams shaping strategic opportunities, and a work prioritization team approving the largest cross-platform investments. Competing investment decisions had a home. No governance negotiation required.

That governance architecture matters more as the volume of AI decisions grows. What the energy company built was not an AI program. It was the operating model that makes an AI program work. The results follow the structure.

The decision that precedes every other one

AI-first CEOs, IBM’s top quartile on integrated AI maturity, report 17% higher revenue growth over three years. That top quartile shares a common profile: cross-functional collaboration, AI embedded across end-to-end workflows, and accountability structured around outcomes.

That profile isn’t a management initiative in the product-centric model. It’s the structure. No one negotiates alignment. The structure makes it automatic. AI embeds across workflows because the team that owns the workflow also owns the AI running inside it. Accountability sits at the outcome, not the function. Value compounds because the team persists after delivery.

The function-led model produces the 68-point gap. The product-centric model produces the 17. Most enterprises are making every other AI decision first. That’s the gap.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

Eliyahu Goldratt spent decades making the argument. Every system has a binding constraint. Resolve it and the next one becomes visible. The throughput of the system is determined by that constraint, not by anything else. He called it the Theory of Constraints and wrote The Goal in 1984 to illustrate it through factory physics: a production line where improving every station except the bottleneck produces no improvement in output.

The framework has a second law that gets less attention than the first. Once you find the constraint, everything else must subordinate to it. Improving throughput at a non-bottleneck stage doesn’t help. In a factory, it creates inventory that piles up in front of the real bottleneck. In a software market, it creates product categories that expand impressively and then contract the moment the underlying constraint shifts. The constraint is the only thing worth optimizing. Everything else is noise.

Marketing technology is a 15-year proof of this. The software ecosystem governing how companies find, engage, and retain customers has been running a live experiment in constraint migration since 2011, when the landscape contained 150 products on a single slide. It now contains 15,505. The State of Martech 2026, published by Scott Brinker and Frans Riemersma this spring, has documented every iteration. The report makes visible a constraint the market is only beginning to articulate.

The First Constraint: Execution

For most of the martech era, the binding constraint was execution: the cost and capacity of producing marketing at scale. You couldn’t publish without producing. You couldn’t produce without people. And people ran out of hours.

Writing content cost money and time. Personalizing it at scale cost more. Testing variations required engineering. The bottleneck was production, and that bottleneck became an industry. 150 products to 15,000 in 15 years, the bulk of them promising to make execution faster, cheaper, or more automatable.

The execution constraint hasn’t resolved. It’s resolving. Generative AI broke it open; platform bundling is doing the rest. Content marketing is the leading indicator.

575 products in 2023. 1,102 by 2025, after two years of generative AI enthusiasm flooded the category with writing tools, creative platforms, and content optimization engines. 2026 reversed the trend: 176 removed, 139 added. Content marketing is the category where the shakeout hit first. 176 exits in one year. That’s what a rising floor looks like. Foundation models absorbed the primary use case. Writing a blog post, producing ad copy, and generating social content at scale are now default features in ChatGPT, Claude, Gemini, and every major productivity suite. Adobe, HubSpot, and Salesforce moved quickly, embedding generation into existing enterprise contracts. The survivors ran into a harder wall: proof that the content worked, not just that it shipped.

There is a pattern here that predates AI. In the 1990s, anime studios raced toward volume while Miyazaki did the opposite: kept Ghibli deliberately small, drew thousands of frames by hand, and tore apart finished scenes when the feeling he was after hadn’t held. Miyazaki’s Spirited Away grossed more than any Japanese film in history when it released in 2001. The studios optimizing for volume that year are mostly gone. So are the films.

The lesson is not that craft beats volume. Miyazaki’s judgment, which frame to hold, which feeling to chase past what the script required, lived in him and nowhere else. Execution abundance collapsed the market for execution. What didn’t collapse was judgment that couldn’t be transferred. That’s where the value migrated.

That pattern is visible in the martech numbers. Execution is getting cheaper and more replicable from two directions at once: foundation models dropped the production cost, and platform bundling put the capability inside contracts organizations already own. Capabilities don’t disappear when they commoditize. They become the rent, not the return.

Constraint resolution has a signature. The tools built to relieve the bottleneck get bundled into the platforms they once competed with. The categories thin.

This is the moment Goldratt described. The bottleneck shifts, the next one surfaces, and the system’s throughput belongs to whoever sees it first.

The Second Constraint: Relevance

The execution constraint resolving does not reduce pressure. It relocates it.

Producing at scale was the problem. Mattering at scale is the next one. This is the relevance constraint, and it is generating the most active category growth in the 2026 landscape.

CMS and web experience management grew 21.4%. E-commerce platforms grew 19.9%. Mobile and web analytics grew 11.3%. SEO tools posted positive net product movement for the third consecutive year, each year with a different definition of “search.”

The mechanism is consistent across all four. Websites now answer to a reader they weren’t designed for: agents that don’t browse pages but parse them on behalf of principals who never visit. CMSs are being rebuilt to structure content for machine consumption, not just human reading. Analytics grew because AI-mediated research leaves fewer fingerprints. The trackable portion of the journey shrank; the signals that survived got more valuable. E-commerce platforms grew because AI intermediated the product comparison stage. When customers evaluate options through AI assistants rather than search results, structured catalog data is the price of entry.

SEO should have contracted. AI disrupted the core use case, keyword-driven search, and the category added products anyway. The discipline that began as keyword stuffing for search crawlers is evolving into Answer Engine Optimization (AEO) and Generative Engine Optimization (GEO): infrastructure for making brands findable and accurately represented inside AI-mediated search, conversational interfaces, and every surface where customers form opinions about products they haven’t yet searched for. The relevance constraint is a harder problem than the execution constraint, and no foundation model carries what solving it requires: your customers’ actual signal. No model trained on it. No competitor can buy it.

The Third Constraint: Orchestration

Relevance is the constraint of knowing what to say and to whom. Orchestration is the constraint of making those decisions flow through connected systems at the speed and volume AI makes possible.

For years, martech integration was a craft industry. Every connection between tools required custom API work, custom data mapping, and ongoing maintenance. The integration layer of the martech stack was real engineering, and its cost set a practical ceiling on how connected any organization’s stack could become.

The Model Context Protocol changed this. Open-sourced by Anthropic in late 2024 and adopted by OpenAI, Google, and Microsoft within months, MCP handed AI agents a standardized interface to external tools and data sources. A universal connector for the agentic layer. According to the State of Martech 2026 report, independent registries now index more than 29,000 MCP servers. The commercial martech landscape took 15 years to reach 15,000 products. MCP covered comparable ground in 18 months.

What that number means in practice: the connectivity problem, the problem the integration layer spent years solving, has a solution now. Most agents can reach most tools in principle; in practice, the wiring is still being built. The constraint that justified years of integration engineering is no longer the primary bottleneck, though the standard has outpaced adoption.

The harder question resides one level down. Integration platforms grew 8.0%. Marketing automation grew 5.9%. The question those categories answer couldn’t be asked while connectivity itself was the bottleneck: which connections matter, what should flow through them, and who owns the decisions that result. The category is attracting builders who recognize that agentic orchestration is architecturally different from rule-based campaign workflows. Campaign logic doesn't port. The governance model, the org structure, and the accountability chain all have to be rebuilt.

Brinker and Riemersma name those roles. The value engineer identifies where company objectives and customer needs produce the highest-leverage intersection and concentrates resources there. The context engineer determines what the AI can see, what data it can access, and what constraints it operates within. Neither role existed as a formal discipline five years ago. Both are now job descriptions in forward-leaning marketing organizations. The judgment each role exercises, which connections matter, which constraints to surface, which decisions to escalate, is particular to the institution running the stack. The orchestration constraint defines the org chart of the AI-native marketing team.

The Fourth Constraint: Coherence

The offer goes out. The retargeting sequence fires. The cadence triggers. The customer receives three contradictory signals in a week. None of the agents involved knew what the others did. Every one of them was working as built.

The coherence constraint emerges from abundance. Any team can spin up agents in days now. The constraint is what happens when they all run simultaneously. An email personalization agent optimizing for open rates and a retargeting agent optimizing for conversion can send contradictory signals to the same customer in the same week. Neither agent knows what the other decided. Neither can explain its reasoning to the human reviewing the outcome. Scale that pattern across a modern martech stack and the organization hasn’t governed its decisions. It has made accountability unlocatable.

The failure mode is invisible until it cascades. The agents are individually doing their jobs. Narrow optimization. Systemic incoherence. An offer that tested well for a segment. A retargeting sequence tuned for conversion probability. A cadence optimized against open rate benchmarks. The customer receives contradictory signals. The organization can’t explain who sent them. The organization has traded traceability for throughput.

The 7.1% growth in governance tools is a lagging indicator. The incoherence arrived first. That growth has two sources. Some of that pressure is regulatory: EU AI Act compliance deadlines, FTC scrutiny of automated decision-making, expanding platform disclosure requirements. Some is operational: organizations that have deployed multiple agents discovering that shared infrastructure without shared constraints produces outputs that are technically correct and organizationally incoherent. What both pressures expose is the same gap: a layer of institutional knowledge, governance rules, and decision constraints that every agent in the stack can access at the moment of decision.

Brinker and Riemersma call it the context layer. What belongs in it: which customer segments received which offers, under which constraints, and what tradeoffs the business decided to weight. How many contacts is too many for a high-value account. Which commitments were made in the last sales conversation. Which escalation logic a human already reviewed and approved. Not a database. Accumulated judgment, made legible to machines.

The coherence constraint is the market discovering that the most valuable thing an organization can build is shared institutional memory, encoded in a form that every agent downstream can use.

Where the Constraint Lives Now

Goldratt was particular about what to do once you find the constraint: subordinate everything else to it. Not improve everything else. Subordinate it. Everything else is a local optimization that looks like progress and isn’t.

The trap isn’t misreading of the constraint map. It’s treating all four stages as priorities.

At the execution layer, the equation is resolving, not resolved. The capability is moving towards table stakes: arriving as a line item in renewals, not a purchase decision. Standalone generative tools for content production are becoming infrastructure. Treating them as a strategic priority is a category error.

Relevance is a different stage: the infrastructure is being built now, and the signal that makes it valuable is already accumulating at competitors who started early. AEO, GEO, structured content pipelines, first-party signal architecture: these are the tools of this constraint. The advantage hardens when built on real customer signal, not borrowed model output. That’s what separates proprietary relevance infrastructure from a capability any competitor can replicate on the same timeline.

Orchestration is the active frontier. MCP is a solution for a connectivity bottleneck in principle, though the standard has outpaced adoption; what remains is the harder question of which decisions should flow through it, and who owns them. Agentic orchestration platforms are defining the architecture now: which agents own which decisions, how human judgment stays in the loop without becoming the bottleneck, which connections matter enough to build for.

Coherence is the constraint most organizations haven’t yet accounted for in their roadmaps. Build this layer now and the lead doesn’t hold steady. Every documented decision, every structured escalation, every constraint a human locked in and a downstream agent inherited widens it. Context graphs, governance frameworks, institutional memory in machine-readable form. They accumulate.

The bottleneck just migrated. Resolve one station, and the throughput the system was hiding becomes visible at the next. The marketing stack is the factory now. Each move left behind a category of tools that looked like progress and became infrastructure.

Find the constraint. The rest is inventory piling up in front of it.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

For twenty years, mid-market managed services has been the market GSIs couldn’t enter. Not for lack of trying. The math never worked. Mid-market, for our purposes, means the 250–2,000 employee company that has always been too big for a local IT guy and too small for PwC.

A 400-person company generates maybe $180,000 in annual managed security revenue. Continuous analyst coverage across L1, L2, and L3 costs more than that. Every GSI that modeled the mid-market hit the same wall. Managed services revenue doesn’t absorb delivery cost.

PwC’s April 2026 managed security service made a credible run at clearing that test: tiered per-asset pricing, 250-plus AI agents on Google SecOps, backed by PwC’s reported $400M Google Cloud collaboration investment. Security is the beachhead, the contract where the math showed signs of finally moving.

What looks like a coordinated push is actually three unrelated capital decisions landing on the same segment at the same time. A GSI the post-pandemic enterprise surge built delivery headcount for, then stranded. A PE rollup that produced a pool of dissatisfied clients through exit-optimization incentives. Hyperscalers whose co-selling programs treat managed services partners as a cloud consumption multiplier.

Three separate capital structures. Each with its own incentive, its own exit clock, and its own definition of success. One response to all three misses all three.

The GSI’s calculus

Global system integrators (GSIs) built delivery headcount for peak enterprise digital transformation demand in 2021 and 2022. The pipeline contracted. The headcount didn’t. The rational move is redirecting that headcount toward managed services: a retainer that doesn’t require a new SOW each quarter, revenue that doesn’t reset when the project calendar does, a contract for the people the pipeline left stranded.

GSIs have redirected headcount toward managed services in prior demand cycles. None of those pushes held mid-market. This cycle is different for one reason. The economics moved.

The variable that killed every prior attempt was delivery cost. A conventional managed security contract is built around a Security Operations Center: monitoring, detection, triage, containment, recovery, with analyst handoffs at each layer. Delivering that for a 400-person company requires L1, L2, and L3 analysts on continuous coverage. Each analyst covers a fixed number of client environments before triage accuracy degrades. A mid-market contract doesn’t generate enough revenue to absorb the coverage it requires.

Agentic architecture may have finally moved that floor. PwC Managed Services Leader Tim Canonico said the April product doesn’t need the people, time, or infrastructure the conventional model requires. AI agents absorb the bulk of L1 and L2 work: alert ingestion, context enrichment, severity scoring, triage routing. PwC’s April product runs 250-plus agents on Google Security Operations across the layers where mid-market gaps have historically been most expensive to close. Per-client cost collapses. The analyst team that once covered four or five environments now oversees AI-managed workloads across dozens. The 400-person company’s contract can finally cover its own delivery.

Cost compresses. Domain depth doesn't. A healthcare company isn’t shopping for managed services coverage. It needs a provider who already knows its HIPAA audit cycle, its EHR compliance obligations, and the OCR investigators it might face. That context takes years to accumulate inside a delivery team. GSIs are solving it through targeted M&A. Accenture bought Cabel Industry for mid-market banking and NeuraFlash for Salesforce-heavy healthcare. Capgemini took Cloud4C for regulated financial services cloud. Tech Mahindra bought Avant Techno Solutions for ISO 20022 infrastructure in mid-market banking. Wipro paid $375 million for Mindsprint, Olam Group’s IT arm, and its 3,200 engineers of agribusiness domain context. Precise bets, with domain depth that doesn’t transfer easily and a pace organic build cannot match.

The PE rollup’s incentive structure

Between 2019 and 2024, PE rollup platforms paid four to six times EBITDA to acquire mid-market MSPs. Exit targets were eight to twelve times. The aggregation worked. The economics of owning them didn’t.

AI stack modernization, triage automation, service desk rearchitecture: each depresses the EBITDA the exit multiple rides. Every dollar spent modernizing the delivery stack depresses the exit multiple. The incentive structure knows this. Eighteen months from exit, the decision is already made. EBITDA today beats delivery quality tomorrow.

Poor CX is a predictable downstream consequence. Clients who joined a MSP for its delivery quality and local responsiveness arrive inside a PE rollup platform optimized for exit, not retention. Delivery infrastructure ages. Local relationships convert to ticket queues. Churn concentrates in the twelve to eighteen months post-acquisition.

The dissatisfied cohort forms through the incentive structure operating as designed.

That cohort was already forming when GSIs arrived with a cost model capable of serving it.

The hyperscaler’s consumption multiplier

Hyperscalers have one definition of a successful managed services partnership: growing compute consumption on their infrastructure. Co-selling is how that interest becomes GSI deal flow.

Google, Microsoft, and AWS route certified mid-market clients to preferred managed services partners because managed services volume is a reliable cloud consumption multiplier. The hyperscaler absorbs deal acquisition costs from the cloud revenue the relationship generates. That math is what changed the mid-market entry problem for GSIs. Customer acquisition cost was the barrier every prior GSI mid-market attempt hit. Co-selling programs transfer that cost to the hyperscaler’s consumption arithmetic, not to the GSI.

The co-selling advantage isn’t GSI-exclusive. Regional and independent managed services providers (MSPs) qualify on the same terms. Demonstrated delivery competency, documented references in the relevant segment, maintained certification standards. MSPs that meet that bar get the same routing advantage.

That's today. AWS Managed Services, Azure Lighthouse, and Google Cloud are each building direct managed services paths. They're routing clients through partners with one hand and building around them with the other.

Uncoordinated intent, coordinated compression

The GSI is solving a stranded headcount problem. The PE rollup is solving an exit problem. The hyperscaler is solving a consumption problem. Each is moving toward mid-market for unrelated reasons, on unrelated timelines, with unrelated definitions of success.

What none of the three structures solved was the AI stack gap inside the PE rollup portfolio. The rollup incentive structure left mid-market operators with aging delivery infrastructure and no path to modernize it. The Anthropic/Blackstone/H&F/Goldman JV, announced in the first week of May, was designed for that gap: a capital vehicle that routes frontier AI investment and embedded engineering directly into PE-held portfolios, without waiting for individual sponsors to prioritize it.

The vehicle runs three tiers. The PE sponsor holds board mandate across the portfolio. Anthropic supplies the intelligence layer and embeds Applied AI engineers in client work. A third tier delivers under the terms the first two set. The JV overlays the rollup platforms, upgrading the AI stack the exit incentive structure never had reason to build.

For GSIs, the JV structure forces a positioning choice: are you in the vehicle, or are you on a subcontract? A position in the vehicle means recurring economics across the portfolio. A subcontract means one engagement, one client, one set of terms.

Three capital structures converged on mid-market independently. A fourth was designed to fix the AI stack gap the third left behind.

Where each player actually sits

Plotting the May 2026 mid-market managed services field reveals four distinct positions, and the pressure each player feels is different. Two variables determine where each player sits: delivery scale and AI investment depth.

The low scale / low AI investment quadrant is the most exposed. Independent MSPs serving local markets with labor-intensive delivery models and no material AI investment in the stack. Their defenses are relationships and proximity, real advantages, but ones that erode when a well-priced alternative enters with a lower cost to serve.

The PE rollup platforms sit in the high scale / low AI investment quadrant. Evergreen Services Group (via Lyra Technology Group) and Nexus IT are examples of a wider pattern: PE-backed rollups that grew through acquisition and never modernized the delivery stack. The Anthropic/Blackstone/H&F/Goldman JV is the first capital structure explicitly designed to fix that, routing AI investment into PE-held MSP portfolios without waiting for sponsors to prioritize it.

One quadrant earns a different kind of attention. Low scale / high AI investment: Sourcepass, Thrive, Integris, and comparable MSPs with a regional focus that rebuilt delivery around AI before the GSIs arrived. They built the delivery infrastructure the rollup exit clock never had reason to fund. They have local relationships the GSIs haven’t built. Clients leaving rollup platforms are looking for what these operators built. The race is whether they find each other before the GSI reference base in those verticals matures.

The GSI entrants sit at high scale with high AI investment. Global brand equity, enterprise-grade infrastructure, hyperscaler alliances, AI-augmented delivery compressing the cost of serving smaller clients below the threshold that kept mid-market off-limits. What they haven’t done yet is run this at mid-market volume. The delivery motion at this scale and this segment is unproven.

Everyone is repositioning. Scale without AI investment and AI investment without scale are both exposed. The variable is execution.

For independent MSPs, execution matters more than the moves.

Five moves, in order

For independent MSPs caught in this convergence, the five moves determine the outcome.

The immediate priority is the delivery stack. The labor-intensive model is the vulnerability GSI entrants exploit. The tools are accessible. The constraint is execution, not capital. AI investment strips out commodity delivery costs, breaks price parity with GSI offerings, and redirects capacity toward advisory work standardized GSI models cannot replicate. For shops that can’t build it: white-label partnerships with AI-native providers are the fastest path to cost structure relevance.

Within twelve months, lock vertical specialization. Generic managed services is the most exposed market position. Vertical depth (healthcare, financial services, manufacturing, legal, construction) drives pricing power, referral density, and defensibility that a standardized GSI offer cannot match. The GSI reference base in many verticals is still thin. The window to establish credible vertical depth before GSI references in that vertical mature is measured in months, not years.

Alongside that: reposition contracts around outcomes, not labor. Per-device and per-user pricing converts an MSP into a commodity line item evaluated on cost. GSIs with AI-compressed delivery win that comparison. Outcome-based contracts eliminate the comparison. The MSP stops competing on coverage volume and starts competing on results the client measures against its own risk or compliance objectives, a much harder target to replicate from a standardized GSI playbook.

Build or join a hyperscaler partnership ecosystem. The co-selling programs that give GSIs distribution advantage are open to any partner that meets the certification bar. MSPs that reach meaningful partner tiers access co-marketing, technical enablement, and deal flow they can’t generate alone. The routing advantage inside hyperscaler ecosystems is the GSI advantage but it’s not exclusively theirs.

Position for M&A, as buyer or seller. For MSPs with steady recurring revenue, demonstrable vertical specialization, and manageable client concentration, the next 24 months present a favorable exit window before GSI competition mutes mid-market valuations. For those with appetite to scale, acquiring vertical specialists or AI-native platforms is faster than organic build.

Two clocks, one window

Prior GSI pushes stalled on unit economics. PwC is betting that variable cleared: 250-plus AI agents on Google SecOps, live production workloads, per-asset pricing a mid-market company can absorb.

For GSIs, mid-market security is the proof, not the business. Renewals fund the next product cycle. Product cycles build the client reference base. The reference base earns the full managed services conversation with the next client. Security is the first rung, not the ladder.

The dissatisfied PE rollup cohort goes to whoever reaches them first with a credible delivery model and a local relationship. That’s the regional specialists today that rebuilt early. GSIs need twelve to eighteen documented renewals in a vertical before that reference base is real. The specialist wins the current cycle. The GSI wins the next.

The GSI reference base is the slow threat. The faster one: AWS Managed Services, Azure Lighthouse, and Google Cloud are each building direct managed services paths that don’t need an MSP in the chain at all. Track both clocks, and track which one moves first.

The math looks different in 2026. The real question is whether your operating model did too.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

A $750 million announcement is also a confession.

It tells you what Google cannot do on its own timeline: activate the enterprise AI contracts it has already signed.

Google carries $460 billion in committed enterprise cloud spend. A substantial portion of that contracted revenue sits in AI projects that are live in contracts but not yet live in production. Every month a project does not go live is a month of recognized revenue that does not clear. An organization with Google’s financial sophistication runs that calculation precisely: what GSI delivery capacity costs to rent versus what deferred recognition on $460 billion in commitments costs.

The $750 million cloud partner fund, announced in April 2026, is the output of that arithmetic.

The fund is real. The revenue is real. GSIs are the most efficient way for Google to protect both, until a cheaper alternative is ready. The firm building that alternative controls when the rent stops.

The perimeter was set before you arrived

The hyperscalers did not arrive at the infrastructure layer of enterprise AI by accident. They built it deliberately, over years, before most of the market recognized that a structural position was being constructed.

The pattern of investment makes this clear. Google invested in Anthropic. Microsoft invested in OpenAI. Amazon invested in Anthropic alongside Google. These were not yield plays chasing model company valuations. They were structural choices made while the market was analyzing the model race rather than examining the counterparties behind the compute contracts.

Every foundation model trains on hyperscaler compute. Every inference run at enterprise scale sits on hyperscaler hardware. If the frontier labs win their respective competitions, the infrastructure that hosts their training and inference participates in that value. If the labs fail commercially, the compute bills have already cleared. The infrastructure owner participates in the upside and is insulated from the downside by the same contract. The position was designed to have that property.

The Q1 2026 earnings reports made the design visible. Google’s parent, Alphabet, posted $62.6 billion in profit for the quarter. Approximately $28.7 billion came from revaluing its Anthropic stake. Amazon disclosed $16.8 billion in Anthropic investment gains in the same period, more than half of its pre-tax income for the quarter. Neither figure was incidental. Both were the return on structural positions established before the model race had a declared winner.

The distribution layer followed the same logic. Copilots, proprietary models, and AI capabilities are now bundled into enterprise renewals that GSIs did not negotiate and cannot replace through procurement. The infrastructure owner and the application-layer competitor are the same legal entity. The firm whose tools GSIs are deploying is also the firm they are losing deals to.

Satya Nadella described the topology plainly when characterizing Microsoft’s relationship with OpenAI as “below them, above them, around them.” The description applies equally to any firm operating inside a hyperscaler’s economic perimeter. It describes a location, not a relationship. The perimeter was set before anyone else knew there was one.

GSIs hold the client relationship the infrastructure owner excluded from its own model by design. That is the role of the issuing bank in a payments network, and the economics follow the same logic. They manage accounts, run delivery organizations, absorb execution risk, and take responsibility when enterprise AI deployments do not perform as projected. They access the ecosystem by paying a form of interchange, accepting terms they did not set and cannot unilaterally renegotiate.

The issuing bank can also be repriced. The network cannot.

The curve always looks obvious in retrospect

Every major platform implementation cycle ends the same way.

It reads, afterward, as obvious. It always does.

The Salesforce implementation market is the most legible version of the pattern. Early implementation work in the mid-2010s commanded 35 to 40 percent EBITDA for the firms doing it. By 2022, the same category of work had settled at 15 to 25 percent for most of the market. The compression was not caused by firms executing implementations more expensively. Certification supply grew. Tooling matured and reduced the skill threshold. Offshore capacity entered the market at labor cost differentials that onshore generalist practices could not match. Procurement teams learned to treat certified implementers as interchangeable on rate card.

The margin compression was the predictable result of a commoditization curve visible to anyone who had looked at comparable cycles in earlier platforms.

The firms that retained margin were not the ones that ran implementations faster. They built proprietary products, vertical accelerators, and industry-specific assets the market could not source from a competitor’s certification roster. They stopped selling implementation and started selling something the market couldn’t certify its way into. The firms that stayed in the implementation economy experienced the compression as an external event rather than a structural selection.

The AI implementation cycle is running the same sequence. The timeline is compressed, and the Partner Margin Collapse Curve maps where the pressure has already arrived. It moves through four phases: Scarcity → Scale → Tooling → Marketplace. Each has a distinct margin profile. Each selects against a different set of capabilities.

Agentic AI sits at multiple points on this curve simultaneously. Deep multi-agent orchestration in complex brownfield environments (production deployments with legacy system integration, exception handling at scale, contested data environments) remains Scarcity-phase work. Few teams can credibly execute it. Standard productivity agents, GenAI copilots, and single-function automation are accelerating through the Scale and Tooling transitions as hyperscaler native tooling matures and certification supply grows. The miscalculation is pricing the entire agentic portfolio as Scarcity-phase premium work when the majority of visible pipeline sits in engagements the Tooling phase will reprice within the next twelve to eighteen months.

The pool of defensible positions narrows at each transition. The firms that survive to Marketplace with pricing power intact are the ones that used the Scale stage to build three things the tooling cannot certify: institutional knowledge of how specific enterprises fail in production, measurement infrastructure that makes outcome contracts operable, and marketplace presence that opens shortlists before the first conversation begins.

The fund is also a clock

Google disburses the $750 million against go-live dates and consumption curves, not certifications, not relationship scores. The mechanism is calibrated to convert stalled projects into recognized revenue, which is precisely the gap in Google’s revenue schedule the fund was sized to close.

Google is not alone in the calculation. AWS, citing Omdia research, puts the multiplier at 7.13x: partners delivering services across the full customer lifecycle reach that revenue multiple on every dollar of AWS revenue they enable. Microsoft has extended Copilot partner incentives and its Frontier Transformation framework. All three hyperscalers are running the same calculation: externalize activation cost, internalize cloud revenue on the contracted timeline.

The information structure embedded in the fund is important to note separately. A GSI learns about a Gemini Enterprise roadmap revision, new native capabilities in a competing platform, or structural changes to partner tier economics at the public announcement. There is no advance notice tier. Every firm in the market receives the information simultaneously.

A GSI that structures its revenue substantially around a single vendor fund ends up optimizing for that vendor’s incentive calendar, not its own. It is a conflict that most planning cycles decline to name.

Naming it is necessary. The harder move is separating the planning cycle into two explicit tracks: one that optimizes for fund activation metrics on the vendor’s timeline, and one that builds the organizational assets that survive the fund’s contraction. They require different investment criteria, different ownership, and different definitions of what success looks like. The mistake is running them as a single track, because the near-term model shows positive variance either way. The divergence only becomes visible when the window narrows and the preparation period for track two has already closed.

The third feature of the terms is less frequently discussed: the window in which the fund operates is determined by the same entity building the alternatives that close it. Internal engineering capacity, forward-deployed engineers at enterprise clients, and productivity tooling that reduces the activation gap without GSI involvement are all under active development. GSIs hold organizational capability the hyperscaler cannot manufacture on a quarterly timeline. The fund is what the hyperscaler pays for that capability while the alternative matures.

The renewal economics on the fund change as the alternative matures. The hyperscaler controls that timeline.

Three assets Google cannot manufacture

Three things have no equivalent inside Google’s organizational structure. They are not evenly distributed in their strategic importance.

The most important is institutional memory. Not the general knowledge that a manufacturing client has legacy ERP debt. The institutional knowledge of how that specific enterprise’s systems fail in practice. What its regulators examine in actual reviews, not in published guidance. Where exception-handling logic resides in the operational memory of supervisors who have managed a process for years and documented nothing.

No hyperscaler has that knowledge. It accumulates through years of operating inside specific enterprises during the messy middle of transformation: the brownfield integrations, the undocumented exception logic, the regulatory nuances that separate published guidance from what auditors actually examine. That experience lives in GSIs at a depth Google’s organizational structure was not designed to hold. GSIs are working to make it durable. IP productization is the mechanism: converting delivery experience into owned accelerators, vertical playbooks, and structured context that survives the engagement and compounds across subsequent ones. The work is active and unfinished. The firms completing it are building the moat. The firms still treating each engagement as a contained delivery event are generating the same knowledge and letting it leave with the team.

The second is outcome measurement and attribution architecture. AI-native competitors pricing outcomes in narrow workflows today built their measurement infrastructure before they signed outcome contracts. They selected outcome pricing domains deliberately: narrow workflows where outcomes were clearly measurable, controllable, and fully attributable without contested causal claims.

Most GSIs are building measurement capability and commercial contracts simultaneously, under client pressure. Treating the shift from time-and-materials to outcome pricing as a commercial decision conflates the announcement of a new model with the capability to execute it. Revenue structured around client outcomes compounds independently of headcount decisions.

The difference is not a pricing structure. It is what the firm can actually measure, control, guarantee, and handle as liability exposure on its balance sheet.

The third is marketplace presence that precedes the relationship. Enterprise procurement for AI solutions increasingly runs a catalog evaluation before initiating vendor conversations. The shortlist is assembled from what is visible in the marketplace, considered complete by the procurement team, and then used as the basis for outreach. An unlisted practice does not appear on a shortlist formed before the first call.

Marketplace presence is not a listing. It requires published solutions with verifiable consumption metrics, clear pricing structures that procurement teams can evaluate independently, and client-validated outcome data that stands on its own without a relationship to interpret it. The procurement team running the catalog evaluation does not know the GSI partner who would otherwise open the door. That relationship does not exist at the stage where the shortlist is formed.

Client relationships retain genuine value in this environment: a contact inside the buying organization opens a conversation that would not otherwise begin. It does not reopen a shortlist the buyer has already closed.

The risks already in motion

The three risks most consequential for GSI strategy are already in motion. Planning cycles have largely treated them as future exposures.

The first is the timing asymmetry in the transition economics. The margin trough is a structural feature of any move from utilization-based to platform-based economics: the period in which the operating model has changed but platform revenue has not yet reached the scale that covers the transition cost.

GSIs that began moving toward platform economics earlier are at or approaching the inflection where platform revenue covers the trough and compounds from there. GSIs beginning the transition now arrive at the trough under different conditions: fund support has contracted relative to the initial cycle, tooling has commoditized further, and earlier-moving firms have compounded their IP lead during the interval.

Delay does not eliminate the trough. It selects the conditions inside it, and determines who has already settled into the positions above it by the time the latecomer arrives.

The second is the liability structure in outcome contracts, accepted ahead of the infrastructure that would make the liability manageable. A failed outcome in agentic enterprise delivery today spans model behavior, orchestration logic, enterprise data quality, and human decision points in sequences that preclude clean causal attribution. The infrastructure required to make outcome contracts operable at enterprise scale (event-sourced decision trails, auditable orchestration paths, attribution frameworks across multi-agent interactions) is under active construction by the firms building the next pricing era.

GSIs signing broad outcome guarantees in advance of that infrastructure are not distributing risk across a new commercial structure. They are absorbing liability that the current state of attribution infrastructure renders unresolvable in a contested case.

The third risk is subtler than the other two. Most enterprise AI delivery organizations carry a significant portion of their operational knowledge in individual practitioners rather than in systems. The knowledge exists in the firm. It does not exist in a form the firm can use after the practitioner moves on, bring to a second engagement in the same vertical, or productize into an accelerator that captures the engagement economics in recurring form.

The informal knowledge base is a realized asset. The institutional knowledge base is the compounding one. The gap between them is where most GSIs are bleeding value they cannot see on any balance sheet.

Two tracks, one window

Track one: optimize for fund activation metrics. Deploy capacity against stalled projects. Convert contracted-but-not-live inventory into recognized revenue on Google’s timeline. The incentive calendar is clear, the payment terms are real, and the revenue is available now.

Track two: use the window to build the organizational assets that hold pricing power in the market that comes after it. Proprietary IP built from delivery experience. Outcome measurement infrastructure built before the outcome contracts that require it. Marketplace presence established before procurement teams close their shortlists.

The two tracks are not mutually exclusive. They require different investment criteria, different ownership, and different definitions of what success looks like. The mistake is running them as a single track because both show positive variance in the near-term model. The divergence only becomes visible when the window narrows, by which point the preparation period for track two has already closed.

The entity that controls how quickly track one closes is the same one building the internal alternative that makes it unnecessary. GSIs do not set that timeline.

The position GSIs occupy in enterprise AI holds real and durable value: client relationships, accumulated delivery knowledge, exception handling at production scale, and the accountability the infrastructure owner deliberately excluded from its own model. The terms on which that value is compensated will narrow as the market matures. The entity setting those terms is the same one building the alternative.

Firms that productize delivery experience into owned IP, price outcomes within the envelope they can actually guarantee, and establish marketplace presence before the relationship begins will carry pricing power into the mature market.

The ones still optimizing for the construction window need to have thought carefully about who built it, and why.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

Professional services surfaced this question first because the arrangemen…

Read more

Between 1991 and 2001, the decision arrived in the form of a procurement question: build an internal ERP competency center, or engage Andersen Consulting, IBM Global Services, or SAP SE’s own teams to own the implementation. It was not framed as a governance question. The frame was operational: speed, cost, talent scarcity. Most enterprises, under genuine pressure, chose the external party.

Ten years on, the choice had sorted the field. Internal centers held design authority: when the business needed a different workflow, the team redesigned it. Outsourced implementations left the enterprise able to operate the system but not change it. The external party had built, through the engagement, a knowledge position the client had funded and did not own.

The pattern predates enterprise software by decades. In upstream oil and gas, it became the organizing principle of returns: the operator holds design authority and determines what happens in the domain; the non-operator contributes capital and receives proportional results. Authority and economics follow the same line.

The mechanism is identical. The timeline is not. Two comparable Global Capability Centers (GCCs), examined in 2026 and again in 2028, make visible what the ERP wave took a decade to reveal. One ended as an Operator GCC. The other remained a Delivery GCC. The assets were identical. The governance was not.

The assets were identical

Both enterprises ran multi-function GCCs of comparable scale and maturity. One covered nineteen legal entities across a European and Asian footprint within a large industrial manufacturing group. The other covered twenty-four entities across North America and Southeast Asia within a global consumer goods company. Both operated across finance, supply chain, procurement, HR, and IT services. Both held SAP S/4HANA as their platform of record. Both had been investing in AI, with their deepest investment in finance operations, where automation in accounts payable and order-to-cash was producing genuine, measurable results.

The structural assets each organization held were equivalent and substantial. Each had accumulated institutional process knowledge through more than a decade and a half of operating the full complexity of its entity portfolio: the exceptions, the regional variants, the undocumented workarounds that experienced practitioners navigated without instruction. Each controlled the S/4HANA environment on which every other business function ran. Each held operational data reflecting what its AP and AR processes actually did in practice, including the failure modes and edge cases that only become visible through years of live operation.

An external operator entering either organization in early 2026 would have spent months reconstructing what each GCC held as ordinary working knowledge. That knowledge was not documented. It was institutional.

The unasked question

The AI program in each organization was producing early results on the metrics it tracked. Automation rates in accounts payable had improved. Cost per invoice was trending down. Neither program was far enough along to produce the kind of outcomes that redraw a domain, but both were further along than most comparable programs, and the board conversations reflected that. The gains were real. They were also narrow.

Read more

Pick a claim. Let’s say it is a prior authorization denial from a mid-size regional health system. The hospital submitted the claim. The insurer rejected it. Someone, somewhere in the revenue cycle, has to figure out whether to recode, appeal, resubmit, or write it off.

This claim is healthcare. The structure underneath it is not. The same liability architecture governs a disputed invoice in AP automation and an unresolved support ticket in enterprise CX. The example is specific. The problem is not.

That decision will pass through at least three commercial entities before it is resolved. The EHR vendor provided the platform where the claim originated. A system integrator configured the revenue cycle workflow on top of that platform. A business process outsourcer staffed the team that is now looking at the rejection queue.

Each of these entities has a contract with the health system. Each contract has a liability cap. The EHR vendor is capped at fees paid. The integrator is capped at deliverables. The BPO is capped at SLA performance. Stack the caps. Whatever falls below them all, the financial cost of the denial, the write-off rate, the uncollected cash, belongs to the only party that negotiated without one: the health system.

The enterprise technology liability stack was constructed deliberately and for defensible reasons. The case for dismantling it begins with the case for building it.

The logic that produced the gap

ASPs learned it the expensive way: run software for a failing business and you become the explanation for why it failed. SaaS vendors read that lesson into every contract that followed. We run the application. You run the business. Liability stops at the application layer.

GSIs learned from the ERP wars. When FoxMeyer Drug’s trustee sued Andersen Consulting for $500 million over a failed implementation, the industry drew the obvious lesson: no clause should ever require the business to survive. The obligation terminates at the delivered system. Liability stops at the implementation layer.

BPOs priced what they controlled and nothing more. Headcount. Transaction volumes. SLA performance. Error rates spike: they add labor. The downstream P&L consequence of that error rate is not their problem. It never appeared in the contract because it was never meant to. Liability stops at the process boundary.

Three layers. Three liability caps. The denied claim belongs to the health system, not to any vendor who touched it. Every party performed its contractual obligation. No party’s obligation required the outcome to exist. Call it the execution-to-outcome gap: every input delivered, no output guaranteed.

Now run the same claim through a different set of hands. Except there is only one set.

The seam is where the liability lives

Evolution IQ did not enter as a platform vendor or a services firm. It entered as a vertical operator and took control of the end-to-end claims management workflow. Prior authorization, coding, submission, denial management, underpayment recovery: no handoffs, no seams. The contract followed from that.

Previously, the application layer handed off at the EHR boundary, the workflow obligation terminated at the GSI, and the headcount absorbed the process. Now: one entity, one outcome, one number.

That is the shift the outcome-based pricing (OBP) conversation keeps missing. Control came first. The contract followed. Running the workflow continuously generates what no project-based model can replicate: a growing corpus of decision traces, exception libraries, and edge cases that make the next cycle faster and the one after that faster still. GSIs restart context at each engagement. AI-native operators carry it forward. That asymmetry is irrelevant in year one and decisive by year three.

When the claim does not resolve, the vendor does not get paid. There is no other party absorbing the gap. Evolution IQ entered, owned the work surface, and exited for $730 million. That is the market’s answer to the question: what is outcome control worth when you can actually prove it?

The workflows where outcome pricing is already working

Healthcare revenue cycle management (RCM). Evolution IQ is one entry point. The broader RCM BPO market runs the same pattern at scale: headcount priced to absorb a failure rate, with every rework cycle passed through as labor cost. Operators priced against resolution do not absorb that cost. They eliminate it. The wedge is the same; the addressable base is orders of magnitude larger.

B2B customer experience. Sierra AI, Intercom Fin, and Zendesk arrived at the same pricing unit independently: charge per resolved ticket, nothing for unresolved ones. Sierra reached $150 million in ARR. Intercom’s revenue quadrupled year over year. The convergence is not coincidence. It is what outcome pricing looks like when the vendor actually controls the loop.

Finance back-office. Loop automates 99% of freight audit invoices and configures new carrier contracts in under 12 hours, against a two-week baseline for legacy systems. Salient is doing the same in auto lending collections. The signature is the same: binary outcome metrics, data that lives in the vendor’s system, and a cost base that was built around BPO labor rather than resolution rate.

That is the displacement in the narrow workflows where it has already happened. The harder question is what determines where it can.

Not all outcomes are priceable

Not every workflow or deal type can bear outcome-based pricing (OBP). Some only support the language. The Outcome Fit Matrix is the tool for telling them apart.

The matrix runs two axes, each running low to high. On the x-axis: outcome measurability and attribution clarity. Can you measure the result and prove who produced the result? On the y-axis: vendor operational control. Does the vendor run the loop, or just touch it? Running means the vendor controls sufficient workflow scope, data access, and execution levers: enough that a miss ends up on their balance sheet, not the buyer's. Where both are high, outcome pricing is viable. Everywhere else, it is theater.

Each quadrant has a different answer to the same question: who absorbs the miss?

Q1: High Control / High Measurability. OBP is genuinely viable. The vendor runs the full loop and charges only when the result is in. Contact center resolution, AP automation, healthcare RCM. Sierra AI does not collect on unresolved interactions. Intercom’s Fin does not charge for abandoned tickets. HighRadius's outcome‑based pricing for the oCFO suite is explicitly tied to measurable financial outcomes. No resolution, no revenue. The risk transfer is genuine, but so is the dependency. Outcome contracts run on information sharing and a level of operational transparency that most enterprise vendor relationships were explicitly designed to avoid.

Q2: High Control / Low Measurability. OBP is attempted but contested. The integrator controls the pipes but not the result. GSIs control significant operational infrastructure but can’t close the attribution loop. Revenue lifts after an ERP modernization. Whether the GSI produced it, or the sales team did, or a favorable demand environment, is far removed from the benefits delivery chain to price cleanly. OBP is attempted in Q2. Disputes follow.

Q3: Low Control / High Measurability. Outcome is visible but vendor can’t drive it. SaaS platforms and AI copilots measure everything and control nothing. A CRM knows how many deals closed. It doesn’t control whether the sales org is well-run, the product has market fit, or the pricing is competitive. Usage is fully instrumented. Control over the outcome is not. Q3 has the most OBP language. It has the least risk transfer.

Q4: Low Control / Low Measurability. OBP is pure theater. No vendor controls the outcome here, and no one can measure it either. Multi-year cloud migrations, ERP transformations, enterprise AI programs. The delivery chain runs through four parties and three years before anyone can call the result. What gets signed are phase gates and value realization reviews. The enterprise holds the outcome. The vendors hold their invoices.

The categories where OBP works are real and growing. They are not where enterprise tech budgets concentrate. Transformation programs, platform modernization, enterprise AI programs: the money pools in Q3 and Q4, where attribution is contested or absent. Better AI agents narrow the execution-to-outcome gap. They do not close the attribution gap. Buyers who conflate the two will sign outcome language backed by nothing and call it a win until the miss surfaces.

Who holds what, and what it is worth

The Outcome Fit Matrix tells you where the opportunity is real. The 2026 map tells you who is actually in position to take it, and what they are risking to try.

The 2026 map plots every major archetype on two axes: outcome ownership and AI nativeness. Moving right means owning more of the miss. Moving up means replacing human judgment with machine judgment. The top-right is the collision zone. Every player in this market is trying to get there by 2027-2029. Not all of them will.

GSIs. The squeeze is coming from both ends. AI-native operators are winning outcome-based mandates in the vertical workflows. Model vendors are pulling margin upstream through alliance structures that GSIs cannot afford to leave. The legacy delivery pyramid sits in the middle, exposed to both, moving fast enough for neither.

What GSIs carry into this fight: thirty years of enterprise relationships that survive leadership cycles, regulatory credibility that moves procurement committees in industries where credibility is the entry ticket, and change management depth that matters when the client’s real problem is not the technology. AI-native operators are acquiring all three. The GSI advantage in each category is real. None of it is durable by default.

Three moves, none of them deferrable. First, productize accumulated workflow knowledge before the next operator enters the same workflow without the context cost. Second, shift the highest-value contracts to shared-gain outcome models while the enterprise relationship still creates the opening. Third, acquire or build genuine AI-native operator capacity before incumbents replicate the language and operators replicate the trust.

Process BPOs. BPOs priced around the one thing incumbents could not touch: the labor pool. AI-native operators bypassed the argument. They do not compete on headcount. They do not compete on SLA performance. They price against resolution in the same workflows, CX, RCM, collections, where BPOs absorb failure as a cost of doing business. The displacement is not incremental. It is categorical.

BPOs carry something operators cannot synthesize on a short timeline: exception libraries from years of high-volume edge cases, the failure taxonomies of regulated workflows, domain context that no AI system builds from first principles. Operators spend their first operating year acquiring what BPOs have already accumulated. The clock starts the moment operators begin running cycles in the same workflows, because every cycle generates the corpus they don’t yet have.

Three moves, roughly in order. First, productize what exists: turn exception libraries and domain context into licensable IP before operators replicate it from their own decision traces. Second, acquire or partner into AI-native vertical delivery in the one or two workflow clusters where the BPO already has the deepest data. Third, reprice those workflows against resolution. The BPO that waits for all three conditions to be perfect will find that operators satisfied two of them without asking permission.

SaaS incumbents. Every enterprise will need someone to govern how agents from multiple vendors hand off work, resolve conflicts, and produce an audit trail. SaaS incumbents are already inside the processes where that governance has to happen. That adjacency is the right starting position, if they treat it as a mandate rather than a default. The risk is not a cheaper competitor. It is that agents don’t need seats, and enterprise buyers running their own AI stacks don’t need a SaaS layer to intermediate workflows they now run directly. The per-seat model is not being disrupted. It is being made irrelevant.

The required moves can run in parallel. First, bundle agents as native platform capability, not premium SKUs. Second, establish the agentic orchestration and governance layer while the integration advantage still holds. Third, sign one managed outcome contract priced against a metric the customer already tracks. That contract is the proof of concept the rest of the portfolio cannot provide: that the per-seat era is over and the SaaS incumbent chose which side of that line to stand on.

AI-native Operators. The incumbents retrofitting their models are not standing still. GSIs are buying in. BPOs are productizing their exception libraries. Model vendors are acquiring enterprise access through PE JVs and GSI alliances. SaaS incumbents are bundling agents into renewals and calling it transformation. Enterprise buyers are getting better at pricing the gap between outcome language and outcome risk. AI-native operators’ lead has a shelf life.

The advantages are three, and they compound. No inherited delivery pyramid consuming margin before a single agent runs. Decision traces that accumulate with every cycle rather than evaporating when the engagement closes. A contract structure where the vendor gets paid when the buyer wins, not when the SOW is signed. Each advantage widens the longer the operator runs the workflow.

Three moves, none of them optional. First, own the outcome contract in the workflows already won before incumbents replicate the language without the risk transfer. Second, expand horizontally before other players buy their way into the vertical. Third, build the audit and compliance infrastructure that regulated enterprise buyers require before they expand. The enterprise accounts that matter will not expand an operator they cannot put in front of a regulator.

Enterprise buyers. For thirty years, enterprises had no leverage on outcomes because no vendor could control them. That asymmetry produced the liability stack, and the liability stack produced the execution-to-outcome gap. The asymmetry is gone in the workflows where operators have taken control. The enterprises that renegotiate first won’t just shape what the standard looks like. They will be the standard.

Three moves, in order. First, own the outcome definition: the metric, the measurement method, and the consequence of a miss belong in the contract, not in the dispute that follows when they're absent. Second, require interoperability and data portability from day one: decision traces and workflow data built on the operator’s system are the enterprise’s asset, not a switching cost. Third, pilot at least one genuine outcome contract before the category consolidates. The ones that wait will sign whatever the category decided without them

What we do not yet know

Four questions are finally testable. The answers will be more instructive than the last five years of forecasting combined.

One: Will AI-native operator unit economics hold their advantage once inference costs, compliance overhead, and human-in-the-loop requirements fully price in? The cost case is real. So are inference costs, compliance overhead, and human-in-the-loop requirements, none of them fully priced in yet. If they plateau before closing the gap with BPO labor, the economics hold. If they don’t, operators have a better architecture and roughly the same margins.

Two: Will a meaningful share of GSI managed services revenue be genuinely outcome-priced by 2029? Define “genuine.” If the metric is contracts with outcome language, the number will look large. If the metric is contracts where a miss costs the vendor real money, the number will be small. Only the second definition is worth tracking.

Three: Who holds leverage in the alliance structures forming now? The assets are distributed: GSIs hold the relationship, operators hold the context, model vendors hold the intelligence. Every partnership forming now is an attempt to consolidate without paying acquisition price. Watch which ones survive their first contested renewal.

Four: Will enterprise procurement build a category for outcome operators before incumbents replicate their language well enough to win on relationships alone? Most procurement teams have no category for an operator that owns the workflow and prices against the result. That absence is the incumbents’ window. Outcome language is easy to copy. The evaluation framework that distinguishes it from genuine risk transfer is not. Operators need that framework to exist before incumbents make it irrelevant.

The foundational questions are testable now. The next twenty-four months will be more instructive than the last five years of forecasting combined.

The Rolls-Royce standard

Call it the execution-to-outcome gap if you want. The vendors who built the liability stack called it protection.

Rolls-Royce solved this in 1962. Power by the Hour: fly the hour, pay the hour. One entity owned the engine, the maintenance, and the liability. When the Trent 1000 blade defect grounded aircraft in 2018, the $200 million repair bill sat on Rolls-Royce’s balance sheet. No base fee. No liability cap. No other party. That is what control, measurability, and contract aligned in one entity actually looks like.

That is the benchmark. The standard every contract in this stack should be measured against: one entity owns the loop, one metric defines done, one balance sheet absorbs the miss.

Some actors are building toward it. The rest are writing “outcome-aligned” into their decks and billing by the hour. The contract tells you which is which, if you know what to look for.

The denied claim is still sitting in someone’s queue. The question is whether the entity responsible for resolving it is the one that gets paid when the claim clears.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

One of them looked at a $40 million client relationship and made a proposal. Call this Firm A.

We can do this for $25 million.

Same scope. Same outcomes. Better economics for you. Less revenue for us.

The other firm ran the same numbers. The economics were clear. The engagement renewed at $40 million. Call this Firm B.

That was the fork. What follows is not analysis. It is a clock.

Why no one in the room is paid to find the $15 million

The services P&L is not a broken system. It is a precision instrument, built to answer different questions than this one.

The partner has a revenue number. The account manager gets paid on bookings, not margin. The delivery lead has headcount to place. The platform team has a deployment goal and a roadmap they are proud of. None of them is gross margin.

The $25M contract runs fifteen margin points better than the $40M engagement it replaces. That is the rational basis for the trade. No one in the room is compensated to find it.

That person exists. She is not usually in this meeting.

So the conversation doesn’t happen. The platform gets embedded as a delivery accelerator inside the existing contract structure: improving execution speed, reducing rework, making the delivery team look good on their internal metrics. The platform story stays technically true. The commercial model stays structurally intact.

The incentives are not misaligned. They are aligned to the wrong outcome.

Three decisions before anyone walks into the room

The proposal required three things to be in place before anyone walked into the room. They are sequential, not parallel.

Compensation protection for the account team. Cannibalization is a compensation decision. Right now, it is not in the comp plan. Without it, the account manager and the partner follow their comp plan to defend existing bookings and revenue. The system is wrong to ask them to choose between their comp plan and the firm’s platform strategy.

A platform with its own commercial identity. A platform without its own contract line has no leverage over pricing. The problem is not the number. It is how the platform was built into the contract. The commercial model change requires a real SKU, its own pricing logic, its own margin accountability. Not bundled into services. Not described as an “accelerator” that makes the billable hours go faster. Until it appears on an invoice the client could have declined, it is not a business. It is a cost center with a logo.

Leadership willing to hold the J-curve in front of a board and say so publicly. The revenue falls before it recovers and someone calls it underperformance. The $40M becomes $25M before NRR above 110% makes the math recover. That shape has a name: a J-curve (that i wrote about in ‘Swallowing the Fish’).

Predictable on a whiteboard. Indefensible on an earnings call. Every firm that held the line had shown the board that shape before the numbers moved. Every firm that reversed course had not. The firms that also disclosed it to the market before the revenue moved had one more thing: they couldn’t reverse course even if they wanted to.

Month twelve: platform economics vs. services renewal

The client paying less is the client that expanded.

Read more

That was Palantir’s structure in financial services: a joint venture (JV) with TWG Global, an investment firm. Palantir owned the platform. TWG brought the portfolio access. Implementation partners were brought in to execute within a framework defined by the JV. The joint venture owned the strategic client relationship, the platform controlled the data, and subcontractors were left holding the invoice: margin constrained, leverage limited, with switching costs building on the platform’s side of the ledger.

Replace TWG Global with Blackstone. Replace Palantir with Anthropic. The March 2026 term sheet negotiations become legible immediately.

This structure has been running for months. The interesting question is not whether it exists. It is whether anyone in the implementation layer noticed it arriving, or recognised which phase they are in.

Two announcements, one architecture

Two stories reached the business press in separate weeks. Read separately, they looked like a consulting story and an investment story. They are one architecture.

In February 2026, OpenAI announced the Frontier Alliance: multi-year partnerships with BCG, McKinsey, Accenture, and Capgemini. Anthropic followed with a $100 million Claude Partner Network anchored by Deloitte, Accenture, and Infosys. The frame was partnership. The structure was distribution. Labs had built remarkable models and discovered they could not cross the enterprise threshold alone.

The gap was never the model. It was everything after the model: sequencing an AI initiative across six business units in fourteen countries, navigating integrations that are political minefields as much as technical ones, explaining the deployment to an auditor, naming who answers when something fails. Labs had no institutional credibility to close that gap alone. Global Systems Integrators (GSIs) had decades of it. The alliances were a structured concession: we will own the intelligence layer; you own the program. For now.

OpenAI’s Forward Deployed Engineering team co-deployed alongside every GSI partner is the tell. Labs are not just buying distribution. They are buying proximity to knowledge they cannot train on: brownfield data estates, governance playbooks, the scar tissue from a thousand programs where something broke in ways no benchmark anticipates. Every joint engagement is a curriculum. Every engagement teaches the lab what it still needs to build.

Then, three weeks later, the second structure surfaced.

The lock-in, not the capital

OpenAI in advanced talks with TPG, Advent, Bain Capital, and Brookfield: a joint venture valued at roughly $10 billion, offering a guaranteed 17.5% minimum return in exchange for preferred access to some 1,200 portfolio companies. Anthropic in parallel discussions with Blackstone, Hellman & Friedman, and Permira for a similar vehicle, though reports suggest without the financial guarantees OpenAI is offering.

The 17.5% is not the story. OpenAI’s latest private funding round has grown to more than $120 billion, up from the $110 billion it announced in February 2026. Anthropic was valued at $380 billion in its February 2026 funding round. Neither needed capital.

What they needed is what BCG’s Matt Kropp named plainly: “a big race to lock in as many desks as possible.”

Once a portfolio company’s data pipelines connect to a model and workflows rebuild around it, the CIO takes credit. No sponsor sanctions a rip-and-replace after that. The 17.5% buys that integration moment, at scale, across a controlled PE portfolio, before a competitor does.

There is a third motive. Routing deployment costs through a separately capitalised vehicle eases near-term pressure on both companies’ balance sheets. It also produces cleaner segment reporting. Both OpenAI and Anthropic are widely expected to pursue public offerings. A well-structured enterprise unit, with contracted portfolio company (portco) deployments across a controlled PE portfolio, reads clearly on an IPO roadshow. The JV is not just a distribution channel. It is a prospectus chapter.

And here is the Palantir structure: model vendor supplies the intelligence and the platform relationship. PE sponsor supplies the portfolio relationships and the board-level mandate. Implementation partners execute within a framework the JV controls. Three tiers. Tiers 1 and 2 jointly create the JV: the commercial instrument that holds the client relationship inside the portfolio, sets the subcontracting terms for Tier 3, and determines who delivers at what margin. The JV is not a third party. It is the mechanism by which the two tiers above govern the one below:

Tier 1: The PE sponsor. Board-level mandate across the portfolio. Decides which lab wins. Controls the investment thesis.

Tier 2: The model vendor. Licenses and consumption that compound with every new deployment.

Tier 3: Implementation subcontracts at terms the JV sets, resetting with every engagement, awarded to whoever the JV chooses.

There is no Tier 4.

GSIs are Tier 3. When the JV closes, the counterparty changes without announcement. Not the portfolio company anymore, but the vehicle. In the Palantir/TWG structure, implementation partners held the invoice. In the Anthropic/Blackstone and OpenAI/TPG structures taking shape now, GSIs hold the same position.

What happened to implementation partners inside the TWG structure is instructive. They delivered. They reset. The JV owned the client. It set the terms for the next engagement. Implementation partners competed for the work on the terms the vehicle had already defined. That is the default trajectory for GSIs in the current architecture, and it is not yet locked.

The facts are right. The timeframe is wrong.

The senior GSI partner reading this sees a different picture. Alliance membership is real. The Frontier Alliance includes the most recognised names in professional services, by invitation, for multi-year terms. Revenue is disappointing across the sector right now, and yet: the Frontier Alliance, the Anthropic $100 million Claude partner network, the PE deal flow arriving at the doorstep. The optimist reads that as the turn. If labs needed GSIs enough to build formalised alliances to anchor them, that looks like leverage, not dependency.

The optimist is not wrong about the facts. The error is in the timeframe.

The alliances represent Phase 1 of the Palantir model: the proof phase, where the platform learns what it needs from the local partner before it formalises terms. Every co-deployment is teaching the lab what enterprise delivery actually requires. Every joint engagement narrows the knowledge gap the GSI currently holds. The alliance feels like partnership because it is designed to feel that way, until the lab has learned enough to negotiate from a different position. The most important moment in a relationship is not when the contract is signed, but when one party figures out it no longer needs the other on the same terms.

The question is not whether GSIs are valued right now. They are. The question is who holds the terms when the JV closes.

In the TWG model, implementation partners were structurally necessary and structurally constrained. The JV needed them. It did not need them to stay.

Three things the current JVs cannot replicate inside a fund hold period, and that did not exist inside the Palantir/TWG structure either.

Brownfield depth. The 2027 SAP ECC mainstream support deadline is not just a migration project. It is a countdown on exit multiples. Portfolio companies cannot complete an agent deployment on top of unmapped, unrationalised data. An AI agent hitting an unmapped SAP profit-centre field does not degrade gracefully. It fails at go-live. Six months of remediation follow, inside a hold period that does not have them. GSIs hold the map of that brownfield estate.

Regulated-sector precedent. Regulated sectors are not a niche of the PE portfolio. They are the majority of it. A lab deploying into financial services or healthcare without compliant data residency, documented model governance, and an audit trail does not face a penalty. It faces disqualification. That documentation takes years to build. It is not achievable inside a hold period, and it is not acquirable in a hiring plan.

Failure-mode knowledge. AI pilots fail for the same reason every time: not the model, but the six months of work nobody scoped before it. Data nobody cleaned or owned. Integrations nobody mapped. Training nobody attended. A JV offering deployment commitments across a controlled portfolio needs someone who has watched go-lives collapse and knows what to do the week before they would. That pattern library is not on any lab’s roadmap.

In a PE context, each one is multiple-compression insurance. A portco that cannot demonstrate production-grade AI impact does not just miss upside. It sells at a discount. Buyers are pricing AI-driven EBITDA capability into acquisition multiples. A stalled deployment is a multiple-compression event. The GP carries it. The LP notices at distribution. Day rates are the wrong unit of account for a problem whose value is measured in fund returns. The GSI that re-denominates its value in exit multiple terms is negotiating from a different position than the one that sends a time-and-materials invoice.

The question for GSI leaders is whether the capabilities that make them necessary become actual leverage, or whether they deliver the work and reset while someone else takes credit at exit.

The question is not whether GSIs are valued right now. They are. The question is who holds the terms when the JV closes.

Where the window is closing

Three forces are converging on the timeline simultaneously, and all three originate from the same source: the exit model.

The first is buyer-side pressure. AI execution is on the critical path to exit. Not adjacent to it. On it. PE sponsors drafted those business cases before buyers started pricing AI execution into offers. The gap between what was scoped and what acquirers now expect is where the pressure lives. A production deployment is worth turns of EBITDA at exit. A pilot deck is worth a footnote.

The second is PE’s internal build. Blackstone already runs a fifty-person data science team pushing shared models across pricing, labour, and opportunity identification into 230 portfolio companies. No GSI invoice in that stack. Apollo requires each portfolio company to map its highest-ROI AI use cases before any external vendor engages. Vista Equity built its Agentic AI Factory with Microsoft, Google, AWS, and Anthropic: thirty portfolio companies now generating agentic AI revenue, no integrator in the structure.

There is a sharper version of this threat that sits inside the second force. Carve-outs are the highest-margin GSI engagement in private equity: urgent, complex, non-deferrable, with limited client capacity to self-serve. They are also remarkably pattern-matchable. TSA scoping, IT separation, cloud migration, ERP carve-out: these workflows are the most exposed to agentic automation of anything in the GSI toolkit. The standardisation that will make carve-outs profitable is what makes them vulnerable. GSIs that do not productise their carve-out IP into replicable platforms will watch that margin thin from the outside in.

Heidrick & Struggles named the AI Operating Partner as a formal emerging role. The mandate is not delivery. It is pre-emption. The AI Operating Partner specifies what delivery should look like, what it should cost, and which standards it must meet, all before a GSI walks in. That is not a procurement function. It is a pricing function aimed at GSI margin, built into the PE operating model as a permanent fixture. The JV sets the terms from above. The operating partner sets them from inside. The GSI arrives to find the room arranged.

The third force is the 2027 cliff. SAP ECC creates non-deferrable, business-critical work. Once that cohort clears, subsequent engagements become more discretionary at the margin. PE sponsors will have options. They will use them. GSI pricing power after 2027 depends entirely on positions negotiated before it.

Default positions don’t require a decision

There is one precedent for a different outcome worth naming. In 2000, Accenture didn’t just accept preferred‑integrator status with Microsoft; it co‑founded and took ownership of the vehicle. Avanade has compounded for more than a quarter century, growing to around 60,000 professionals across more than two dozen countries. The distinction matters: Accenture didn’t just negotiate better terms inside Microsoft’s structure. It secured a position that generates revenue and changes the terms for everyone else. A channel agreement generates revenue until the platform decides otherwise. An ownership position generates revenue and sets the conditions.

The TWG structure did not exclude implementation partners from the commercial relationship. It simply never included them. Default positions don’t require a decision.

The JVs are not yet signed. The terms are not set. Most PE-backed companies sit below the floor of both the Frontier Alliance and the PE JV structures. That is not a gap in the market. That is the majority of the market, currently uncontested.

The IPO clock sharpens the timeline. Labs need GSI delivery to make the deployment numbers real and auditable before the S-1 is filed. The useful window for GSIs is not simply before the JVs close. It is before the prospectus is written.

A GSI negotiating a named position inside the JV structure earns recurring economics across a controlled portfolio. A GSI waiting for portco mandates earns a billing rate that resets with every engagement. One position accrues. One resets. The mandate that matters does not come from the portco CIO, who controls one deployment. It comes from the PE operating partner, who controls thirty. A portfolio-level AI operating agreement across an entire fund does not yet exist as standard commercial paper. The GSI that writes it first sets the price for everyone who comes after.

The Palantir playbook is legible. The difference this time is that the implementation layer can read the structure before it formalises and choose what position to negotiate. In the TWG structure, nobody offered implementation partners a choice. They discovered their position after the architecture closed.

That is still a choice here. For now.

Digital Faultlines tracks the fault lines as the worlds of professional services, software, and AI collide in the enterprise. Get no-fluff signals by subscribing.

Subscribe now

Share Digital Faultlines

Why the alliances happened

The model was the easy part. What didn’t exist was a path from the demo to the workflow that a chief risk officer would sign off on. The gap between demo and deployment became the place where enterprise AI went to stall.

How do we sequence this initiative across six business units in fourteen countries? Which of our business processes should be reengineered for the prioritized business outcomes? Who actually knows each country’s instances, integrations, upstream and downstream dependencies? Who owns the data pipelines? Who signs off when something goes wrong? Can we explain this to an auditor? Who do we name when the regulator asks who’s responsible? These are the questions that killed enterprise AI projects. None of them have anything to do with model quality. Labs had no convincing answers. So pilots stalled, due diligence dragged, and nothing reached scale.

What labs couldn’t buy, borrow, or build quickly was institutional knowledge and credibility. The ability to redesign a business process across fourteen countries and two Global Capability Centers (GCCs). Knowledge of which integrations are political minefields. Credibility with model risk committees and Architecture Review Boards that only comes from having faced one. And the relationships wired into the CIO, COO, and sometimes CEO that turn intellectual interest into a funded mandate. Those things accrue over decades of programs. None of it shows up in the term sheet.

The credential is not a capability. It is a track record. Every business and operating model redesign, every M&A integration, every ERP cutover where something went wrong and someone stayed until it was fixed: that is what makes the difference. A lab can build a remarkable agent. Only a GSI can arrive at a client with that agent governed, tested against a risk committee’s questions. AI packaged inside a transformation inherits years of that trust. AI arriving alone starts from zero.

Wipro’s Client Zero strategy is the clearest expression of this pattern. Gemini Enterprise ran through Wipro’s own operations first: finance, HR, sales, delivery. Wipro took the integration risk so the client didn’t have to. What reaches the client is not a prototype. It is a system that has already encountered production conditions and survived them.

Enterprises buy accountability, not models

From the enterprise buyer’s perspective, most AI vendors look familiar. The motion is the same as enterprise software in the 1990s: demos, benchmarks, direct deals. The assumption baked into that motion is that buyers will evaluate the technology on technical merits and sort out the rest. In regulated enterprises, the rest is the primary obstacle.

The object of an enterprise purchase is accountability. A lab offering “access to a powerful model” hands the buyer a capability and a problem: now build the governance, sequence the change, name the owner, and explain it to the auditor. A GSI offering a transformation roadmap has already done that work. The model is wrapped inside it. The buyer inherits the outcome, not the liability exposure.

The Anthropic–Infosys arrangement makes the division of labor concrete. Infosys does not resell Claude. It wraps Claude inside its Topaz platform, embedding it as a native component within domain-specific AI agents for claims, compliance, and network operations. The client never buys a model. They buy an Infosys program, governed in Infosys’s language, accountable under Infosys’s name. Claude is the engine. Infosys is what the risk committee sees.

The same dynamic runs through the Google partnerships, expressed differently by each partner. TCS brings institutional scale: 13 Experience Centers by end of 2026, practitioners who have spent years inside client environments. Wipro brings field validation: agents that cleared one risk committee before the enterprise has to face its own. The lab provides the intelligence. The GSI provides the cover.

The labs are studying while they partner

Every partnership announcement has a subtext: the labs cannot yet get into the rooms where enterprise AI is decided without someone who has earned a seat at that table. GSIs are not partners in the ideological sense. They are a pass. The alliances are a concession: we will own the intelligence layer; you own the program. For now.

What the labs are acquiring through their partners is not just distribution. It is access to a playbook they cannot buy any other way. Every joint engagement teaches them where the real friction lives, which governance questions recur, and how regulated industries actually think about model risk. Forward-deployed engineers are not just critical on these programs. They are taking notes. The GSI’s proprietary client context is the prize: the brownfield knowledge, the client relationships, the governance narratives. The partnership is how the lab gets close enough to study it.

OpenAI has stopped being subtle about where this goes. Sarah Friar, OpenAI’s CFO, has outlined a business model moving well beyond subscriptions and API usage toward outcome-based pricing: arrangements where the lab shares in the value its models create in enterprise operations, financial modeling, and scientific research. OpenAI’s ARR has tripled each year from $2 billion to $6 billion to over $20 billion, tracking compute capacity closely. Friar’s argument is that the next leg of that curve runs through deeper workflow ownership, not better benchmarks. A lab with a direct economic stake in enterprise outcomes is not a technology supplier. It is a competitor to every firm that helped it get into the room.

Two primaries and a battlefield

Seen through this lens, the competitive structure of enterprise AI collapses into a surprisingly clean picture. Two sets of players are competing for the center of gravity in enterprise AI. A third cluster owns the terrain they must both cross to get there.

Frontier labs are competing to become the default intelligence layer: the substrate that enterprise AI runs on across domains, programs, and years. The bet is on embedding, not selling. Google’s version is visible in the Wipro numbers: 200 production-ready agents across six industries, baked into a GSI’s industry solutions before the enterprise ever makes an explicit choice. The lab is the ingredient. The GSI is the shelf. OpenAI’s Frontier Alliances work the same way, albeit through a different route: embed the agent platform inside the reference architectures that McKinsey, BCG, Accenture, and Capgemini will carry to hundreds of clients. The intelligence layer travels with the playbook, not the pitch deck.

What GSIs are competing for isn't the intelligence. It's the client operating context the intelligence can't survive without. The multi-year roadmaps, the operating model redesigns, the change management, the regulatory scaffolding. The opportunity, at its fullest, is to become the orchestration layer that decides what to deploy and when at a client, across hyperscalers, SaaS vendors, and model providers alike. Their advantage is relational, process-based, and earned through years of absorbing what goes wrong on programs that matter.

The GSIs who understand the stakes are also doing something the labs cannot easily replicate: codifying that hard-won delivery experience into platforms and vertical solutions that outlast the engagement, not just with the people who ran the last program. The productization effort is already underway. GSIs are certifying hundreds of thousands of practitioners in agentic AI fundamentals and deploying them across ERP, contact centers, and commerce at a scale no lab can replicate with its own staff. The GSI that architects a client’s AI operating model decides which intelligence layer sits at its center, how it is governed, and who answers for it when something breaks. That is the difference between a vendor and a principal.

GSI earnings already reflect the stakes: leading firms expect to double bookings from AI and data ecosystem alliances even as overall legacy revenue grows modestly. The alliances are not a sideline. They are becoming the growth story, which is the kind of concentration that gives a lab leverage when it decides to renegotiate.

What the alliance announcements leave unresolved is the question that will ultimately matter most: who controls the workflow surfaces where AI has to show up and perform. The SaaS platforms, vertical systems, and AI-native startups that own the CRM, the ERP, the industry-specific process layer are not signatories to any of these deals. They have their own intelligence strategies, their own embedding ambitions, their own alliances, their own ideas about which lab is at their center. A GSI that commits to just one lab’s platform is locked, by extension, into wherever that lab sits in this fragmented landscape. That may not be where the client’s critical workflows live.

What the labs cannot buy, borrow, or build

None of this is an argument that GSIs lose. It is an argument about which GSIs win, and why.

Strip away what labs can eventually buy or build, and four things remain. None of them are glamorous. All of them are hard to replicate.

The first is institutional trust: earned over programs where something went wrong and someone had to absorb it. A regulator, a board, a procurement committee extends that trust to a firm that was already in the room when the last hard decision was made.

The second is client context: the proprietary context accumulated inside client environments and knowledge of what a client’s estate actually is, as opposed to what the architecture diagram says it is. Decades of undocumented integrations. Processes built around workarounds that became policy. Institutional memory distributed across people who have been on the account for years. Before an agent can run, someone has to translate that reality. That translation is not in any lab’s training data.

The third is accountability structure: regulators do not ask who built the model. They ask who signed off on deploying it, who monitors it, and who answers when something goes wrong. Those questions have to resolve to a named organization with an established relationship with the regulator. A lab, however capable, is not a credible answer. A GSI that has been the named accountable party on a hundred regulated programs is.

The fourth is domain and architectural judgment: the ability to decide not just what to deploy, but where, in what sequence, with what human oversight, and at what threshold the agent stops and a human takes over. A software vendor can deliver a capable agent. It cannot tell a chief underwriter which claims decisions the agent earns the right to make alone, or advise a COO on how to restructure a fulfillment operation around it, or design the escalation path that a regulator will accept. That judgment lives at the intersection of industry knowledge, functional expertise, and architectural experience accumulated across hundreds of programs. It is the reason a GSI engagement starts before the model is chosen and continues long after it is deployed. No lab has it. No lab can train for it.

The durable advantage belongs to GSI firms that productize what they know, not just to firms that know it.

If GSIs execute well, frontier models become raw material that GSIs refine into proprietary platforms, fixed-price programs, and non-FTE revenue streams: subscriptions, licensing, education. AI capabilities become recurring, higher-margin income rather than episodic projects. In that scenario, OpenAI, Anthropic, and Google are not just tools GSIs implement. They are interchangeable engines nested inside a GSI-designed operating system for the enterprise.

There is a version of this that goes badly. Treat the intelligence layer as the source of differentiation: build the methodology around one lab’s tools, train the delivery organization around one lab’s platform, and you have quietly made your competitive position dependent on a partner whose incentive structure is evolving away from yours. Friar has been explicit: OpenAI intends to capture a share of the outcomes its models enable. Follow that logic and the lab’s revenue growth runs through the same margin the GSI currently owns. A GSI that cannot swap the intelligence layer when that conflict matures has very little negotiating leverage left.

The governance architecture that prevents that dependency is in [Two Machines Cannot Share One Throttle →]

The strategic challenge is not whether to partner. It is how to keep the labs as interchangeable engines under a GSI-designed operating model, rather than becoming a highly skilled but ultimately interchangeable channel in someone else’s platform story.

Partner today, channel tomorrow

The current arrangement has the feel of an armistice. Both sides need each other, the economics work, and no one has the capacity to fight on a second front. That holds until it doesn’t.

Every joint engagement the labs run through GSI partners is also a lesson in what the integrator charges for and why. Governance tooling, reference architectures, forward-deployed engineering: the labs are not guessing at the value stack. They are mapping it in real time, one engagement at a time.

History has a name for this dynamic. The firm that owns distribution for a platform it doesn’t control starts as a partner and ends as a channel. Distribution is not ownership. The alliances are a genuine opportunity. The question is whether the GSIs treat them as a starting position or a destination.

Digital Faultlines tracks the fault lines as the worlds of professional services, software, and AI collide in the enterprise. Get no-fluff signals by subscribing.

Subscribe now

Share Digital Faultlines

Two camps. Neither is wrong about the facts. The “AI will eat Services” camp: once agents write code, configure systems, and close tickets, the case for paying GSIs to do it more slowly disappears. The “Services-as-Software” SaS camp: GSIs sit on a goldmine of codifiable workflows, a generational prize for whoever encodes delivery into agents and platforms. For every dollar spent on software, six are spent on services. That ratio is why neither camp thinks it needs to argue harder.

Spend twenty years in the messy middle of global enterprise programs at F500 companies, as I have, and the real variable comes into focus. Both camps describe real forces. Neither describes why clients actually sign.

The blind spot is the same in both camps: they count what GSIs do, not why clients hire them. That is what makes the 6:1 ratio dangerous. A market that size, with that much of its value sitting in liability and institutional trust rather than deliverables, is not a market you take by automating the deliverables.

The Numbers Are Right. The Conclusion Isn’t.

HFS Research sizes the addressable market at $1.5 trillion by 2035. Foundation Capital pushes the estimate to $4.6 trillion, pointing at total enterprise salary and outsourcing spend that “intelligent agents are now poised to absorb.” General Catalyst has committed $1.5 billion to an AI roll-up fund focused on acquiring and transforming labor-intensive services firms into high-margin, AI-driven platforms. Six out of ten enterprise leaders surveyed by HFS Research plan to replace some or all of their professional services with AI-driven solutions within three to five years. That is not a fringe view. That is the mainstream.

The operating data runs in the same direction. Accenture booked $5.9 billion in applied AI deals in FY2025, nearly doubling year-on-year. ISG’s Index Insider reports that the fourth quarter of 2025 saw the most significant divergence between revenue and headcount growth in over three years. Revenue decoupling from headcount is no longer a theory.

Stop the story there, and GSIs look like software companies with a services heritage waiting to be shed. Call this the SaS Sooners thesis. The transformation is not coming. It has already arrived.

But this version quietly assumes something large: that most of the economic value in a GSI engagement sits in the tasks and workflows you can encode, rather than in the liability, politics, and institutional trust you cannot.

The Firm That Gets Fired When It Goes Wrong

Ask a Fortune 500 CIO why they hired a leading GSI instead of a cheaper integrator. The honest answer is rarely about capability. It is about what happens when something goes wrong.

They are buying a sentence: “We followed expert guidance.” The firm that can make that sentence credible is the firm that gets the contract. As JC Bahr-de Stefano at Better Tomorrow Ventures argued in “Services Won’t Become Software”: AI can do the analytical work. It cannot do the blame-absorbing work.

A large chunk of professional services spend is CYA spend. Cover your ass. Liability transfer. Credentialed judgment you can point to afterwards.

The GSI engagement is not a knowledge product with a services wrapper. It is a liability product with a knowledge wrapper.

Nihar Bobba’s work on Vertical AI sharpens this with three diagnostic questions: Who is the principal? Who owns the liability? Who holds the relationship with regulators and auditors? In a GSI engagement, the answer to all three is the GSI. That is not a detail about how services happen to be structured. It is the reason they exist.

That structure is not accidental. It is load-bearing. Which is why the next development is so counterintuitive: the AI technology that threatens to automate services is simultaneously expanding the legal obligation to hire humans to oversee it. The EU AI Act imposes penalties up to 35 million euros or 7% of global turnover, with liability sitting squarely on the deploying firm regardless of where the AI came from. That is new. What is not new: Sarbanes-Oxley 404(b) has mandated external attestation of financial controls for two decades. ERISA has required outside advisors as a fiduciary prudence defense for longer. The CYA function was never informal. It was always statutory.

The litigation record is consistent enough to be a rule. Zimmer Biomet is not an outlier. MillerCoors sued HCL for $100 million, arguing the integrator lacked the “governance, methodology, or experience required.” Hertz sued Accenture to recover more than $32 million it had already paid the firm after Accenture 'never delivered a functional website or mobile app' following a two-year digital overhaul engagement it had been hired to lead. The common thread: enterprises hired a named firm because they needed someone to depend on. When things went wrong, that firm became the defendant. Every failed program reinforces the board-level reflex to hire an even more recognized firm next time.

The litigation archive is, in effect, the CYA function made public.

The litigation cases are the visible end of something that begins much earlier. Long before any program fails publicly, someone had to decide: delay go-live and absorb the cost, or ship and absorb the compliance risk. Scope the data migration tightly and take the business disruption, or over-scope it and take the budget hit. Interpret the regulatory requirement conservatively, or accept the ambiguity. Each call carries real financial and career consequences for the executive who owns it. What they are buying from the GSI is not just expertise. It is a counterparty who shares the downside, negotiates the compromises, and stands in front of the board when the outcomes are messy. The CYA function is not only retrospective. It operates in the room, in real time.

None of that works without institutional memory. Before the first agent touches a workflow, someone has to map decades of undocumented integrations, broker alignment with the leadership that has lived with the old system, and redesign processes built for a world without AI. That translation work, making a brownfield estate legible to agents, is not a technical problem. It requires the kind of change management experience that only comes from having done it badly before. The GSI’s institutional map of the legacy estate is not a relic of the old delivery model. It is the entry ticket to the new one.

What Gets Eaten and What Doesn’t

Picture a typical large-scale GSI engagement. A CIO presents a business case for consolidating 14 legacy ERP instances onto SAP S/4HANA, migrating 60% of workloads to cloud, and layering AI-powered planning on top. The board approves a $180 million transformation budget, plus a five-year managed services contract worth another $250 million.

The first phase, strategy and design, looks nothing like software delivery. Months of consultants in rooms, mapping processes and coaxing alignment from people who have every incentive to resist. The deliverables are SaS-addressable (meaning AI and agentic tools can plausibly automate or compress them): target operating model, migration roadmap, all increasingly AI-draftable. The politics are not. Getting a European business unit president to accept the same chart of accounts as North America is CYA-anchored work. It requires a credentialed firm. It cannot be automated away.

Then the factory opens. Hundreds of consultants configure, build, integrate, and test. This is pure SaS-addressable territory. Build and test together account for the majority of project effort, and the majority of billable hours. The productivity gains in these phases are real: 35 to 45% on code generation, 40 to 50% on documentation, based on current benchmarks. These phases will be automated heavily. The SaS thesis gets this phase right. They are also the phases where the CYA-anchored layer above them becomes more exposed, not less, as velocity increases and governance gaps widen.

Running parallel to the factory is the layer that is almost entirely CYA-anchored: governance, risk, compliance, and organizational change management. Change management is routinely underfunded in ERP programs, despite being the phase most likely to determine whether the system gets adopted. The deliverable is accountability itself.

After go-live comes the managed services tail: a five-year contract worth $250 million, which exceeds every phase of the implementation combined. Of all the phases, it is the most SaS-addressable and the most CYA-anchored at the same time. The platforms absorb tickets, patches, and incidents. But the post-go-live period is also when regulatory exposure peaks, when audit relationships crystallize, and when the client is most dependent on a named party to stand behind the system. The indemnification clause is why the client signed with a named GSI rather than a cheaper platform.

Map those phases back against the SaS thesis, and the same pattern emerges each time: AI shrinks the factory, accountability holds the contract. Three places where the thesis misreads this.

First misread: treating all GSI revenue as output-driven labor. The SaS view implicitly assumes most GSI spend is paying for deliverables that can be automated. Under the CYA lens, a large share of that spend is paying for risk ownership, political cover for internal leaders, and brand and credential signaling to boards and regulators. AI can automate artifacts. It cannot sign the SOW, sit in front of the audit committee, or get fired if things go wrong. That is the non-addressable TAM, and it is larger than the SaS story acknowledges.

Second misread: assuming automation gains translate into durable margin. They do not. As one Big Four Chief AI Officer described it, AI improved systems integration efficiency by roughly 30%, and those gains showed up “in large part as price discounts.” The GSI captures the spread for 12 to 18 months. Then procurement notices the velocity change and reprices it. What never gets repriced is the governance wrapper. You cannot put accountability on a rate card.

Third misread: assuming the CYA layer shrinks as the factory accelerates. It does not. Every point of velocity gained in build widens the governance gap above it. Programs fail and AI pilots fail in production when the humans responsible for change, alignment, and accountability cannot keep pace with the execution. That is not a technical problem. It is the problem a named, credentialed firm exists to absorb. AI makes the factory faster and the CYA layer more exposed at the same time. That gap is the opening. Most GSIs are leaving it unpriced.

So: the factory is not the argument. Agentic AI automates it, procurement reprices it, and the SaS Sooners thesis was right about all of it. The “AI will eat services” camp was right too. The factory gets eaten. What both miss: the factory is not what clients are paying for. The half it cannot touch: executive steering, change management, compliance attestation, vendor orchestration, program ownership. A board needs a named, credentialed, insured party to stand behind the work. AI cannot do that. That is where enduring margin sits. It always was.

Act One Ends When Procurement Reads the Benchmarks

Everyone agrees AI will improve GSI margins. They disagree on whether that improvement is durable. Both sides are half right.

The structure matters. Most analysts are still writing about Act 1.

Act 1 is margin harvest. AI cuts delivery costs across code generation, testing, and document handling before procurement catches up. The GSI keeps the difference. It lasts 12 to 18 months.

Act 2 is re-commoditization. The same tools that gave GSIs an edge ship to every competitor at once. Hyperscalers, OpenAI, Anthropic, tool vendors: the proprietary advantage is gone inside 18 months. Procurement teams now hold the benchmarks. They demand the discount back.

Act 3 is where durable margin lives. Outcome-based contracts with genuine risk transfer. Managed Services positions where the GSI owns the SLA and carries the indemnification clause. Proprietary IP embedded deeply enough into client operations that switching hurts.

The GSIs that reach Act 3 will look different from the ones that entered Act 1. Fewer consultants, more agent-orchestrated platform delivery, IP that accumulates with each engagement.

The economics will not reach software levels. The blended margins will still land in services territory, not 80% SaaS. But IP-led and agentic AI-led lines will carry meaningfully higher gross margins than the headcount-arbitrage model they replace. The difference is that those margins will be durable, because they are attached to accountability rather than efficiency.

Still Services

The ideal end state is AI-intensive execution wrapped in an accountability shell that no software company can replicate. Thinner factories, leaner but more valuable specialist and governance layers, and a revenue mix that looks more like platformized services than pure software. Better margins. More recurring revenue. Still services.

The GSIs that get this wrong will automate their way to a price war.

Durable differentiation in this model will come from the intersection of three things that are genuinely hard to replicate: institutional trust built over decades of client engagement, proprietary context accumulated inside client environments, and the ability to orchestrate human judgment and AI together in ways a pure software vendor cannot.

That combination is the moat. It always was. AI just made it more visible by stripping away everything else.

Services will not become software. But they will become more software-like.

The alliances between Frontier Model Labs and the major GSIs are the first real-world stress test of everything argued here. The next piece in this series reads the signals, maps the opportunities, and names the challenges.

Digital Faultlines tracks the shifting fault lines where professional services, software, and AI collide inside the enterprise.

All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

Somewhere in every Global System Integrator (GSI) strategy conversation, the same question surfaces: how do we build an agentic AI platform business without destroying the services business we already have? It is a reasonable question. It is also, in my view, the wrong one. The right question is whether those two things can coexist inside the same organizational structure at all during the transition.

I think the answer is no. And the reason is not just culture or change management or leadership commitment. It is a structural property of how the two businesses work.

Essay 1 of this series (Spin, Embed, or Buy?) mapped the three available plays: Disintegrate → Operate → Integrate, Embed and Evolve, Acquire and Protect.

This essay is about the mechanism that makes two of them fail, and what it actually takes to run the one that doesn’t.

The Incompatibility Is Structural, Not Cultural

A services business is a linear machine. Revenue scales with headcount. To grow twenty percent, you hire approximately twenty percent more people, train them, staff them, and bill them. The returns are relatively predictable. The margins are relatively stable. The whole system is precision-engineered to maximize the output of billable human hours.

A platform business is a compounding machine. Revenue scales with adoption. Once the platform is built and proven, each new customer costs a fraction of the first. Inference compute scales with usage; unlike SaaS, these costs don’t approach zero. But the workflow logic, the encoded judgment layers, and the data learning loop multiply across every client simultaneously. The economics improve as the decision traces, context graphs, and proprietary data accumulate and the agentic platforms get better. The returns accumulate. (see The Missing Layer in Enterprise AI: Context Graphs for the full framing of the context graphs problem)

The problem is that these two machines require fundamentally different governance structures, incentive systems, and resource allocation processes. The linear machine is built to optimize for utilization. The compounding machine is built to optimize for retention, expansion, and product velocity. You cannot run both optimization functions from the same control panel.

What happens when you try is not ambiguity or tension. It is a predictable, systematic result: the linear machine wins every resource allocation fight. Every quarter. By design.

The Linear Machine Wins Every Budget Fight. By Design.

In 1997, Clayton Christensen articulated why successful organizations systematically fail to invest in disruptive initiatives. The mechanism is not strategic blindness. It is resource allocation. The governance structures of successful firms are calibrated to fund initiatives that meet existing profitability thresholds. Platform investments, which require years of cash burn before producing competitive returns, fail those thresholds every quarter. So every quarter, they lose the resource fight to initiatives that don’t.

In a services firm, this plays out through the partner P&L. A senior leader’s compensation, career trajectory, and organizational standing are tied to FTE utilization, quota retirement, and attributed revenue. When transformation programs ask that person to route deals through a platform P&L they don’t control, they are asking them to take a concrete compensation hit in exchange for abstract future credit. This is not a request that an incentive system built around quarterly utilization will accommodate. The engineers get redeployed. The roadmap gets captured by client requests. The platform quietly becomes a cost center.

Revenue per employee has not meaningfully moved in a decade, despite sustained investment in automation and AI tooling. It is evidence that the linear machine is functioning perfectly. When the bench gets long and platform engineers are sitting idle, the system redeploys them to billable work. It is making the rational decision given its objective function.

The conclusion I draw from this is that the linear machine cannot be modified from inside to produce non-linear returns. The optimization function has to change entirely, and that requires structural separation, not transformation programs.

Structural Separation Changes the Fight, Not the People

Disintegrate → Operate → Integrate is the structural response to this problem. The logic is not complicated.

Spin the agentic AI platform capability into a legally and financially separate NewCo. Give it its own P&L, its own governance, and its own incentive structures built around product metrics. Let OldCo become NewCo’s first paying customer at arm’s length prices, which forces NewCo to build something worth buying and forces OldCo to behave like a real customer rather than a captive internal budget. Once the platform has established independent customers, defensible margins, and a track record, use OldCo’s distribution to expand it.

The key insight is that this structure changes what the governance fight is about. Inside a single organization, the fight is between a platform initiative and a services P&L competing for the same engineers and the same budget. That fight has a known outcome. Between two legally separate entities with separate P&Ls, the fight is a commercial negotiation. NewCo has to earn OldCo’s business. OldCo has to pay a real price. Both entities are forced to behave like the businesses they need to become.

The structural design for the Operate phase, the period when OldCo is NewCo's primary customer, is what I call hub-and-spoke. NewCo is the hub: it owns the agentic AI platform, the IP, and the data learning loop that makes it defensible. OldCo’s service lines are the spokes: domain expertise, client relationships, human judgment. The spoke manages the client relationship. The hub manages the product. This resolves the client confusion problem while keeping the platform’s commercial integrity intact.

Kyndryl’s separation from IBM in 2021 is the most recent example of this logic applied at scale. Post-separation, Kyndryl exited the legacy client relationships that IBM’s account culture had subsidized for years. It signed hyperscaler partnerships with AWS, Microsoft, and Google – partnerships that were structurally impossible inside a firm competing with those same hyperscalers. Gross margins expanded from 15.6% to 18.8% within three years. The separation did not create new technology. It created new governance. IBM couldn’t pull Kyndryl back into its orbit, so Kyndryl was forced to find its own orbit. It did.

The Atos experience is the failure mode. A separation attempted under financial distress, with a balance sheet that constrained options and a leadership team that changed multiple times, produced neither a clean services company nor a clean technology company. Both entities spent years in structural limbo. The lesson is not that separation is risky. It is that separation initiated from distress is not the D→O→I playbook. It is a restructuring that borrowed the playbook’s language.

Price the Downside. The Upside Takes Care of Itself.

I want to address the capital allocation argument, as I think most boards are receiving a version of it that emphasizes upside and underweights downside.

The upside case for all three structural plays involves the same story: platform revenue, multiple expansion, competitive differentiation in an AI-driven market. These stories are not wrong. They are incomplete.

The complete version requires assigning probabilities. Most large platform bets fail. Not because of execution failures specific to any one firm, but because the base rate of platform success is low, the competitive landscape is brutal, and the window for establishing defensible position is short. Under a realistic probability distribution, the relevant analysis is the expected value of the failure scenario, not the success scenario.

1. Embed and Evolve failure: a services business that spent a decade distracted from its core market, with scattered automation assets inside business lines that are not separable, saleable, or recapitalizable.

2. Acquire and Protect failure: a balance sheet impaired by a software multiple priced into a services-equivalent revenue stream, with the engineering talent already departed and the platform roadmap captured by enterprise client requests.

3. D→O→I failure: OldCo intact, running its linear machine, generating cash. NewCo with its own customers, contracts, IP, and P&L, available for sale to a strategic acquirer, recapitalization by an external investor, or orderly wind-down without contaminating the parent.

The third failure mode is structurally superior to the first two. A board evaluating these options should be assigning explicit value to the optionality that D→O→I preserves in the downside scenario. Most aren’t. They are evaluating the transformation narrative, not the failure mode.

There is an additional point that rarely surfaces: D→O→I is the only playbook that manufactures a financeable asset. When IBM fully separated its managed infrastructure business, Kyndryl gained the ability to attract capital, partners, and talent on its own terms. That option did not exist inside the parent. D→O→I builds it from day one. Embed and Evolve failure leaves nothing to spin off. Acquire and Protect failure leaves an asset that has been converted from software to services and can’t be unwound. The asymmetry matters more than the narrative.

The Operate Phase Is Where Compounding Either Starts or Dies

The Disintegrate is a structural and legal decision. The Integrate, if the platform succeeds, is commercially obvious. The Operate phase is where the seperation has to start. It is also where most operating model implementations I have observed lose the plot.

Three things have to be true for the Operate phase to work.

Transfer pricing has to be real. I’ve watched this fight up close during the rotation to digital operating models in the last decade. Finance proposes booking NewCo’s platform revenue at cost to the group P&L, recovering it later “once platform adoption accelerates.” It sounds sophisticated. It is a quiet attempt to turn NewCo into the same '‘internal lab’ it was before the separation: funded centrally, priced at zero in deals, cut when margins are tight. The internal platforms that major GSIs have built – the SynOps and Cognix architectures – are shadow-charged at zero and given away to win services deal renewals. They have never seen a price signal. The teams building them don’t know whether anyone would pay for them at market rates. Arm’s length pricing is the only mechanism that forces the machine to prove it can hold price.

Incentive design has to be explicit. The transformation programs I have reviewed universally include language about “permission to cannibalize.” Permission is not enough. The incentive system wins over the permission system every time. The BU head routing deals through NewCo’s platform needs to see that choice reflected in their compensation, not just in a transformation program charter. The tell is simpler than any diagnostic: where do your best people want to sit? As long as the real money and the real prestige are in OldCo P&Ls, the gravity shift hasn’t happened. D→O→I, for those firms, remains a sentence on a slide.

The productization discipline needs to be institutionalized. Agentic AI-assisted delivery generates insights that are currently evaporating at the end of every engagement. A team finds a faster way to process claims, the client notices, the engagement ends, and the insight lives in the heads of the people who did the work. The firms building advantage extract that insight systematically: which steps run in sequence, where human judgment is irreplaceable, where Agentic AI handles volume. They convert it into repeatable models that can be hired to, trained against, and sold again. One corollary: when clients want features outside the platform roadmap, that request hits the spoke’s P&L, not the hub’s. The moment NewCo starts building bespoke features for individual clients, it has become a services business. It starts with one exception. It never stays there.

The real moat here is not methodology. It is the institutional decision memory and proprietary operational data: decision traces, claims processing patterns, invoice error distributions, supply chain exception signatures. This is the context and training data for AI agents that competitors cannot replicate. Operational data and context accumulated over years of running client processes cannot be replaced. OldCo’s spokes generate this raw material. NewCo’s job is to productize it.

Scaling or Quietly Decaying: Four Diagnostic Questions

Most GSIs already have slides that say “NewCo,” “platform,” and “Agentic AI.” The question is whether the transformation is scaling or quietly decaying into internal tooling. The most reliable signals are not in dashboards. They show up in who complains, where politics cluster, and what your auditors notice.

Are your internal transfer-pricing fights painful? In a genuine D→O→I, NewCo charges OldCo near-market rates. That creates real friction. The scaling pattern: BU heads complain that “NewCo is too expensive,” but still route deals through it because they lose outcomes without the platform. The decay pattern: finance tells NewCo to “support the business” with internal discounts, and the platform becomes a pre-sales subsidy. If your internal customers aren’t upset about paying NewCo, NewCo probably isn’t real.

What happens at renewal? Early wins are easy to spin. Renewals tell the truth. In a durable transformation, the platform holds or raises prices because outcome data and integration depth make it hard to rip out. In a decaying one, renewals turn into classic services renegotiations: platform pricing gets cut to defend the services envelope and reclassified as an enablement cost. If you cannot hold price with a satisfied client, you do not have a defensible asset.

What bad news reaches your desk? In a transformation, NewCo is in real markets competing against real vendors. The bad news includes losing bake-offs to external SaaS or Vertical AI providers, missing feature commitments, under-delivering on contracted outcomes. That is painful. It also means NewCo is a vendor, not a slide. In a decaying transformation, the bad news is about low internal adoption, lack of “change readiness,” insufficient enablement. The platform is an internal change-management problem, not a product facing a market. If you are not hearing about NewCo losing or winning external deals, you are not yet in the platform game.

Can your auditors tell you are becoming a platform company? Transformation shows up in accounting before it shows up in marketing. In a transformation, finance and auditors force real debates about ARR recognition, cost allocation between platform and services, and data-driven performance and outcome contracts. Your metrics pack changes because your economics change. In a decaying one, life looks the same to the CFO: most revenue is still T&M or fixed-price services, platform contributions disappear inside “digital enablement,” and the only change is a few internal project codes. If your auditors can’t tell you’re becoming a platform company, neither can the market.

The Procurement Problem Is an Advantage Waiting to Be Claimed

One dimension that doesn’t feature enough in the D→O→I debate: neither GSIs nor clients fully understand how to contract for AI-driven outcomes yet. Neither side knows how to get there. Providers cannot reliably price outcomes they cannot fully control. Clients cannot define the outcomes they want with enough precision to hold a vendor accountable.

The firms that develop workable contracting language, risk frameworks, and pricing models first will have a structural advantage that outlasts any tool or methodology. This is work NewCo is better positioned to do than OldCo, because NewCo is building toward repeatable delivery patterns rather than bespoke engagements. Once your contract structures become the industry reference point, switching costs calcify. That is what the compounding machine looks like in practice.

Not Every GSI Deserves a NewCo

The prerequisites for D→O→I are important to name, because a significant number of firms attempting this playbook don’t meet them.

Protected capital: ring-fenced budget for NewCo that does not compete quarterly with OldCo’s P&Ls. NewCo burns cash for three to five years. If that budget is subject to the same quarterly resource fight as everything else, it will lose that fight.

Honest assessment of agentic AI platform capability. Some GSIs have neither the proprietary domain data nor the workflow depth to build an agentic AI platform that external customers would choose over a market alternative. For those firms, the right strategy is to be an excellent spoke – the domain expert and integration partner that sits on top of agentic AI platforms built by firms that had what it takes. That is a real business with defensible margins. It is not the narrative that goes on an analyst day slide. But it is honest, and it does not destroy a decade of capital on a bet the firm was not equipped to win.

The compounding machine is real. Building it alongside a linear machine, without destroying either, requires structural discipline that most organizations find uncomfortable. The ones that achieve it will define the economics of enterprise services for the next decade. The ones that run the language without the structure will find out, around 2030, that their linear machine is now a component of someone else’s self-reinforcing one.

This is Essay 2 of the series on enterprise technology services transformation. Essay 1: Spin, Embed, or Buy?

Digital Faultlines tracks the fault lines where professional services, software, and AI collide in the enterprise. Get no-fluff signals by subscribing. All essays at Digital Faultlines.

Subscribe now

Share Digital Faultlines

The agentic AI platforms launching right now (from pure-play startups with nothing to cannibalize, from hyperscalers and platform incumbents with captive distribution, and from frontier model providers building domain agents to move into the services layer) are selling performance and outcomes at a fraction of the cost of her delivery pyramid. Her clients are running the numbers. Some have stopped scheduling the QBR.

She has options. Her board calls them “the Services-as-Software platform opportunity.” She knows they are three different wagers on what her organization is capable of. Most firms narrate one and run another. There is a default. She doesn’t know it yet. That clarity is still ahead of her.

This essay is about that honesty.

OldCo vs. NewCo: The Gap

Before the playbooks, a precise understanding of the gap.

Business outcomes that used to be delivered primarily through bespoke human professional services are now codified into software platforms, AI agents + human delivery mechanisms, and automated workflows, sold on a recurring or outcome basis, and continuously improved through data feedback

OldCo, the typical technology services firm today, runs on billable headcount. Employee costs consume 54–58% of operating income. EBIT margins cluster at 15–25%, structurally capped because more revenue requires more people. A revenue-per-employee number that has not moved meaningfully in a decade, despite years of automation investment and internal AI platform development. Platforms like Accenture SynOps and TCS Cognix exist as internal tools and improve delivery efficiency. They are not sold as standalone products. They are OldCo getting marginally better at being OldCo.

NewCo, a genuine Services-as-Software business, runs on ARR, NRR, gross margin expansion, and cost-per-outcome. Its talent mix inverts OldCo’s: roughly 60% product and engineering, 30% domain and delivery, 10% sales and customer success. Revenue per employee is not a ceiling. It is a variable that rises as the platform scales, because the marginal cost of serving the next customer approaches zero while the marginal value of the accumulated operational data approaches infinity.

Seven criteria define whether a technology services business has actually made the transition to a Services-as-Software business:

The distance between OldCo and a business that meets all seven criteria is not primarily a technology problem. The distance is organizational, commercial, and structural. The three playbooks are three different answers to the same question: what actually stops transformation, and how do you get around it?

One framing clarifies the stakes: call this the Vertically Integrated Services bet. Start as a services company, then build or buy software platforms to radically transform the P&L.

Pull it off, and a mediocre services business crystallizes into a software-services hybrid with expanding gross margins, growing reinvestment capacity, and a defensible data moat. Miss, and the outcome is permanent: an undifferentiated, capital-intensive services asset with no path to non-linear economics. There is no partial credit.

The three playbooks are three different mechanisms for attempting the same bet.

Playbook 1 – Embed & Evolve: “We can transform the core from within”

The Embed & Evolve thesis is appealing in its elegance: don’t fracture the organization. Embed AI and platform economics into every service line simultaneously. Redefine KPIs across the firm. Retrain the workforce. Shift commercial models deal-by-deal. No NewCo. No brand confusion. No organizational civil war between OldCo and a competing internal entity.

The strongest proof point is in banking

DBS Bank executed this successfully, and its transformation is among the most studied and most misread case studies in enterprise technology.

In 2014, CEO Piyush Gupta made a specific decision: no separate digital bank. He would transform the whole institution. The mechanism: 33 technology platforms rebuilt from the ground up; a “2-in-a-box” leadership model pairing every business head with a technologist counterpart; 60-plus customer journeys redesigned from customer backwards rather than system forwards. No ring-fenced lab. No innovation theater. The bank became the digital unit.

Why It Fails

DBS had three things that almost no technology services company has simultaneously.

A decade-tenured CEO with complete board alignment. Gupta has been in the chair since 2009. He did not manage the transformation by consensus. He decided and executed. The average technology services CEO tenure is four to six years. That is not enough runway for Embed & Evolve, which requires sustained, consistent pressure against every organizational incentive that pulls toward the status quo.

A relatively homogeneous core product. Banking services are banking services. DBS was improving the delivery of a stable product set, not dismantling and rebuilding what it sold. Technology services companies typically span twenty or more service lines with different clients, different competitive dynamics, and different degrees of platform readiness. The coherence requirement scales with complexity, and at that complexity it exceeds most leadership teams’ capacity.

A culture that could be induced to self-cannibalize. Partner compensation in technology services is built around revenue attribution. Asking a partner to shrink their attributed FTE revenue in order to grow a platform P&L they don’t control requires institutional altruism that partner incentive structures actively punish. DBS had employment-based compensation, not partner-model attribution. That difference matters more than most analyses acknowledge.

Clayton Christensen’s core insight applies here with precision: the resource-allocation process of a successful incumbent systematically defunds initiatives that cannot meet the profitability thresholds of the existing model. Every quarter in a technology services firm, utilization must hit 80%, the bench must be filled, and the $30M SAP deal must be staffed. Every quarter, the platform initiative loses that fight.

In technology services, Accenture is making the equivalent bet. June 2025: CEO Julie Sweet collapsed five business units into a single entity called “Reinvention Services.” No structural separation. No NewCo. In October 2025, it went further: appointing its first-ever Chief Offerings and Products Officer, mandated to productize offerings and embed AI into every engagement. A CPO inside a $70 billion technology services firm is an organizational declaration that product economics now require a dedicated C-suite owner. Whether Accenture can hold the coherence that Embed & Evolve demands is still the test.

The honest verdict: Embed & Evolve works if you have a decade-tenured CEO, exceptional board alignment, a relatively homogeneous product set, and a compensation model that can actually reward self-cannibalization. For most technology services companies, at least two of those four conditions are absent. When they are absent, Embed & Evolve becomes a sophisticated cover story for incremental improvement.

Playbook 2 – Disintegrate → Operate → Integrate: “Create the NewCo you intend to be eaten by”

If the immune system will kill the new model, give the new model a separate body. But if you only separate, you have two companies rather than a transformed one. The point is not independence. It is gravity shift. D→O→I is a staged sequence designed to create a protected space where new economics can develop, prove themselves through real commercial discipline, and then pull the center of the whole firm toward them.

The Three Phases

Phase 1: Disintegrate. Carve a new entity with a separate P&L, its own leadership, distinct metrics (ARR, NRR, gross margin, not utilization and bill rates), its own talent brand, and genuine commercial autonomy. NewCo’s charter: build services-as-software products, including verticalized agent platforms, outcome-priced workflow bundles, and domain-specific AI services. The legal structure can vary: a full spinoff, a joint venture with external investors, or at minimum a ring-fenced business unit with protected multi-year capital. What it cannot be is virtual. A virtual NewCo with shared leadership and shared metrics is a cost center with a rebrand.

Phase 2: Operate. OldCo becomes a paying customer of NewCo. The structural architecture is Hub-and-Spoke: NewCo is the hub, owning the platform, the IP, and the data learning loop; OldCo’s service lines are the spokes, embedding NewCo’s capabilities into client engagements and providing the domain expertise and change management that the platform cannot supply alone. Transfer pricing is at arm’s length. NewCo charges OldCo the same rates it would charge any external client, minus a modest internal discount. This is not a courtesy arrangement. It is a commercial forcing function. If OldCo won’t pay a real price for NewCo’s platform, it means NewCo hasn’t built something worth buying. Internal subsidy produces internal tools. Arm’s-length pricing produces products.

Phase 3: Integrate. When NewCo’s economics clear the thresholds (ARR exceeding 20–25% of combined revenue, gross margins above 40%, platform adoption across a majority of major client engagements) you shift the firm’s center of gravity toward the NewCo model. Integration at this stage is more economic and cultural than legal. The firm’s identity, talent investment, go-to-market, and capital allocation all reorient toward platform-delivered outcomes. FTE-based work doesn’t disappear. It becomes the supporting layer: the human judgment, domain expertise, and exception handling that the platform’s learning loop depends on to stay defensible.

The Precedents

IBM’s 2021 spinoff of Kyndryl offers a technology services example of benefits of full seperation. Kyndryl entered independence with operating margins around 15.6%, weighed down by unprofitable long-term contracts and the governance friction of operating inside a parent with different strategic priorities. Three years post-separation: margins expanded toward 18.8%. Not because the underlying work changed, but because Kyndryl had permissions IBM would never grant. It fired unprofitable clients. It became hyperscaler-agnostic, signing partnerships with AWS, Microsoft, and Google that were structurally impossible inside a company selling its own cloud. Full separation created the conditions for margin discipline that protected separation had consistently prevented.

The construction industry offers a structural template that maps more precisely to this playbook than any tech analogy. Construction ownership transitions: a NewCo takes on new contracts while OldCo provides labor, equipment, and bonding via lease arrangements; profit shifts from OldCo to NewCo over five to seven years; if NewCo succeeds, it absorbs OldCo’s assets and capabilities. If it stumbles, OldCo’s assets remain protected. Replace “construction equipment” with “client relationships and delivery workforce” and the template is almost exact.

Why It Fails

Cannibalization paralysis. OldCo leaders resist routing revenue through NewCo because it shrinks their attributed P&L. Without explicit incentive redesign (rewarding OldCo leaders for total client value, not just FTE-based revenue) this quietly kills the Operate phase.

Capital drain without patience. NewCo burns cash for three to five years before breakeven. Public services companies with quarterly earnings pressure discover that the CFO’s patience is shorter than the timeline requires.

Atos as the cautionary tale. Atos’s attempted separation of Tech Foundations from Eviden happened from a position of severe financial stress. Debt levels made asset sales structurally difficult. The separation dragged, the debt ballooned, and the firm spent years fighting for survival. The lesson is brutal: D→O→I requires strength. It cannot be the answer to financial distress.

Integration timing. Integrate too early and OldCo reabsorbs NewCo, defaulting to FTE pricing. Integrate too late and NewCo develops its own identity and resists. The ARR and gross margin thresholds are guideposts, not formulas. This remains a judgment call.

The honest verdict: D→O→I is the most structurally honest playbook for most technology services companies, because it is the most honest about organizational physics. It does not assume the core will self-cannibalize. It does not assume an acquisition will be protected. It creates a structure designed to force the behaviors that culture cannot be trusted to generate. But it requires a CEO with a seven-plus year horizon, ring-fenced capital, explicit cannibalization incentives, and the discipline to not integrate before the thresholds are real.

Playbook 3 – Acquire & Protect: “Buy the non‑linear economics”

If building product DNA organically takes too long and transforming the core culture is too hard, buy a company that already has both, and protect it from the immune system that will immediately try to absorb it.

The Canonical Examples

Cognizant’s 2014 acquisition of TriZetto for $2.7 billion is the case study. TriZetto managed healthcare benefits administration for roughly 180 million covered lives, half the insured US population. At acquisition: approximately $700 million in trailing revenue, EBITDA north of $190 million, deep clinical and administrative domain data that no new entrant could replicate on any reasonable timeline. Cognizant kept TriZetto semi-autonomous within its healthcare practice. Product leadership retained. Engineering culture retained. Software-centric metrics retained. Cognizant wrapped its services distribution around TriZetto’s platforms and projected $1.5 billion in cumulative revenue synergies over five years.

The explicit goal was non-linear software revenue attached to a domain-defensible platform. The distribution flywheel worked because Cognizant’s healthcare client relationships were deep enough to accelerate adoption of a platform that already had proven unit economics.

Accenture Song is the second instructive example. Accenture Interactive, founded in 2009, spent thirteen years running a disciplined Acquire & Protect strategy: buying creative and experience agencies (Droga5 in 2019 being the largest and most notable) to accumulate brand, design, and creative DNA that its consulting culture could not generate organically. For years it protected individual agency identities and kept acquired leadership intact. The strategy worked. By the time Accenture rebranded the practice as Song in 2022, it had assembled the world’s largest creative agency network inside a technology services firm.

The Three Conditions

Three conditions must hold simultaneously for Acquire & Protect to work.

Genuine product-market fit, not just interesting tech. The acquired company has paying customers, measurable unit economics, and a product roadmap driven by market demand. Platforms with genuine penetration come with the operational data and learning loops that make domain defensibility real. If the technology has never competed for an external budget and won, it has not been tested.

Genuine operational autonomy. Separate leadership. Separate P&L. Separate engineering cadence. The acquisition target keeps its soul, its compensation structure, and its hiring standards. The moment the acquired company’s product roadmap becomes subject to OldCo’s governance calendar, the acquisition begins dying.

A real distribution flywheel. OldCo’s sales force can cross-sell the acquired platform into existing accounts, and the sales team is compensated for platform signings at rates comparable to services signings. Without this, you have paid a software premium for a business you will run as a services business.

Why It Fails

The failure mode here is well-documented and consistent: the mothership absorbs the acquisition. The acquirer imposes its governance cadence: quarterly reviews, utilization targets, partner approval gates. The acquired company’s best engineers leave when they realize they are now in a consulting company. The product roadmap gets colonized by bespoke client requests. Within three to five years, the platform is a tool, not a product.

The honest verdict: Acquire & Protect is the fastest path to non-linear economics. It is also the most fragile, because it depends entirely on a discipline (leaving the acquisition alone) that every process and incentive inside a services firm works against. The acquisition is worth more the less you integrate it. The more you integrate, the more you destroy what you paid for. Most services firms cannot hold that discipline for five years. The ones that can are worth studying.

The Diagnostic: Which Play Are You Actually Running?

Three playbooks. Three bets on organizational reality. The diagnostic is not which box you want to occupy. It is which box your incentive structure, your capital allocation, and your governance calendar are actually putting you in.

Strip away the language in the strategy deck and ask three questions:

1. Where is the new model’s capital coming from? If it is competing quarterly for budget against utilization-driven P&Ls, you are running Embed & Evolve regardless of what you call it.

2. Who owns the platform roadmap? If the answer involves account leaders, partner committees, or client escalation paths, you are building a delivery tool, not a product, regardless of whether you have a NewCo on the org chart.

3. How are OldCo leaders compensated when FTE revenue shifts to platform revenue? If they are penalized, the transformation will not happen. The incentive structure is the strategy.

   ---

In the next essay of this series, Two Machines Cannot Share One Throttle, I step out of the neutral analyst role and build the comparison framework, makes the explicit case for D→O→I as the default starting position, and examine the counterintuitive signals that reveal whether a transformation is compounding or quietly decaying.

Get no-fluff signals on enterprise software, AI, and professional services by subscribing to Digital Faultlines.

Share

Share Digital Faultlines