The trade
The trade many enterprises are making right now goes something like this: cut the people with twenty years of organizational change pattern recognition, and replace them with people who have eighteen to thirty-six months of AI implementation pattern recognition, on the theory that the AI knowledge matters more.
That trade only pays if the failure modes are technical. If the failure modes are organizational, which is what the data has been saying for thirty years about every enterprise technology rollout, the math gets ugly.
The argument inside the executive suite is reasonable on its face. AI is the visible technology shift of the decade. Change capacity feels like overhead. If you are going to bet, bet on the people building the future, not the people who maintained the structures of the past. The CFO likes the math. The board likes the story. The trade gets made.
The scale is not small. Across 2026 year-to-date, fifty-six percent of tracked tech layoffs (roughly one hundred and fifty-six thousand workers) explicitly cite AI or automation as a contributing factor, and roughly six in ten companies have signaled they plan to conduct layoffs this year. The headcount that disappears is concentrated in functions where AI is positioned to do the work and in functions framed as overhead, which is where change and transformation roles sit on the org chart.
The argument has a problem inside it, though, and it shows up when you look at who actually has the AI experience the trade is supposed to be paying for.
The calendar problem nobody talks about
Enterprise-grade large-language-model deployment has been a real category for about two and a half years. Serious agentic AI in regulated production environments has been deployable for closer to twelve to eighteen months. Both of those numbers are generous, and they apply to the people at the very front of the field.
There is no one with a fifteen-year track record of agentic AI in regulated industries, because the technology did not exist fifteen years ago. The most experienced enterprise AI implementer in the world today has roughly three years of post-ChatGPT work, and even that profile is rare. The median AI hire being brought in to fill the gap left by a departing change practitioner has materially less.
So what is actually being traded is twenty years of pattern recognition about how organizations adopt new technology under regulation, in exchange for less than three years of pattern recognition about how a specific new technology behaves in production. The first is deep and durable. The second is shallow and still settling.
This is not a knock on the AI hires themselves. Many of them are excellent. The point is that the calendar has not had time to produce the candidate the trade assumes exists.
What the failure mode actually looks like
The headline number from MIT's Project NANDA report in July 2025 is that ninety-five percent of enterprise generative AI initiatives are delivering no measurable business return, despite roughly thirty to forty billion dollars in cumulative enterprise investment. Gartner's late-2025 survey of seven hundred and eighty-two infrastructure and operations leaders found only twenty-eight percent of their AI use cases fully succeeding and meeting ROI expectations, with twenty percent failing outright and the remainder stuck in the middle. Gartner is also predicting that more than forty percent of agentic AI projects will be canceled by the end of 2027.
The interesting question is why. The naive answer is that the models are not good enough yet, which is increasingly hard to defend after two years of capability gains. The more honest answer is the one the NANDA authors gave directly: the failure is not in the technology, but in the approach. The bottleneck is some combination of unclear ownership, missing governance, missing change adoption, fuzzy success metrics, regulatory risk that nobody owned, customer-facing comms that nobody wrote, and an operating model that did not evolve to accommodate a new unit of work.
All of those bottlenecks are exactly the things a competent change and transformation function exists to solve. Cutting that function and then being surprised that the AI rollout stalls is the enterprise equivalent of firing the QA team and being surprised that the next release has bugs.
The change capacity numbers in the same period are telling. The share of organizations offering formal AI upskilling fell from roughly thirty-five percent in 2025 to twenty-six percent in 2026, even as AI deployment accelerated. Only thirteen percent of US workers report receiving any AI training from their employer. The trade is not just visible in who is being hired and fired; it is visible in what is being funded.
The model is not the bottleneck. The bottleneck is the gap between deploying a capability and actually changing how the organization runs around it. That gap used to have an owner. In a lot of places, it does not anymore.
What the consulting firms did
If you want to know which side of this trade is correct, do not listen to what the consulting industry says about it. Look at what it actually did with its own hiring.
Across 2024 to 2026, the major strategy and transformation houses did not cut their change and operating-model practitioners to fund AI. They kept them, gave them AI tools, and expanded their scope.
- McKinsey grew QuantumBlack to about 5,000 AI specialists integrated into its existing transformation practice. Its internal Lilli assistant is now in active use by roughly 72 percent of its 45,000 consultants. Headcount is planned to grow another 12 percent in 2026.
- BCG is expanding BCG X to around 3,000 engineers. AI-related work is projected to be roughly 40 percent of BCG's revenue by 2026.
- Bain equipped its entire 18,000-person team with internal AI tooling. AI-and-tech revenue is around 30 percent of its consulting business, with a path to 50 percent.
- Accenture committed roughly three billion dollars to its Data and AI practice and is doubling its AI workforce to 80,000 specialists.
- Deloitte, PwC, EY, and KPMG all deployed internal generative AI assistants (PairD, ChatPwC, EYQ, Workbench) to tens of thousands of existing consultants on the same 2023 to 2024 timeline, framed as augmentation of their transformation and strategy practices.
The pattern is uniform across firms whose business model depends on reading what will be valuable in the next ten years. The transformation, change, and operating-model skill set is treated as the moat. The subject matter (cloud, then digital, then agile, now AI) is treated as the interchangeable piece. The methodology travels. The technology comes and goes.
If the people whose business is reading the next ten years are keeping their change practitioners and expanding their scope, that should tell you something about the read on the next ten years.
The irony in the trade
There is a quieter dynamic underneath all of this that is worth naming.
The same consulting firms that are arming their transformation practitioners with AI tools and selling that combination to enterprises are, in many cases, helping to frame the conversation that justifies the enterprise cutting its internal transformation function. The pitch is some version of: AI is the future, you need to move fast, your change capacity is overhead, here is how we can help. The transformation expertise that just got cut on the client side does not disappear from the engagement. It comes back through the consulting firm, billed at consulting rates, with an AI label on the deck.
This is a clear-eyed bet on what skill is durable, and the consulting firms are right about the bet. The irony is that the enterprises taking the advice are absorbing the opposite lesson, hollowing out the internal capability that would let them do this work themselves, and reinforcing the dependency that flows back to the same firms.
The eighteen-month window
If this pattern plays out the way enterprise technology adoption usually does, the bill comes due in roughly eighteen to twenty-four months. The rehiring wave is already being predicted in industry analyst commentary, which is unusual; it normally takes one to two business cycles before the reversal becomes visible enough to forecast. Inside that window, a small number of organizations will quietly realize their AI bets are not converting, and the explanation is not the AI. They will look for the people who can take an early-stage capability and integrate it into the operating model the way it has to work for an underwriter, a claims handler, a relationship manager, a portfolio leader, a regulator. They will find that those people are scarce, partly because the market already cut them, and partly because the surviving ones will have learned what the work is actually worth.
The intersection that gets re-priced inside that window is not pure transformation expertise, and it is not pure AI engineering. It is the practitioner who can hold both halves at the same time: AI implementation experience, organizational change experience, and the discipline to govern engineering practice quality as velocity climbs.
That intersection is the underpriced credential of 2026. It is likely to be the premium credential of 2027 to 2029.
What to do with the trade you are about to make
If you are sitting on the executive side of this trade and have not made it yet, or have made part of it and have time to recalibrate, there are three moves worth considering.
Treat the cut as a portfolio decision, not a category decision. There is a real argument for thinning the agile-coach layer if it has drifted into ceremony work, the same way there was a real argument in 2010 for moving commodity development offshore. The mistake is treating the entire change capability as commodity. Some of it is. The senior change architects, the operating-model designers, the people who can sit with a regulator and explain how an agentic workflow is governed, are not. Those are the people you want close to the AI work, not adjacent to it.
Integrate, do not run in parallel. The AI rollout team and the change capability cannot be running parallel programs that meet quarterly. The change practitioners need to be inside the AI design conversations from the first sprint, because the operating-model implications and the governance implications are not afterthoughts. They are the work.
Build the governance layer before you need it. Most enterprises will not, because the velocity of the AI work feels like it cannot pause for it. That is exactly when it gets paid for later. The teams that get the governance layer right early are the teams that do not have to refactor their AI estate when the regulator shows up.
The frame
Cheap labor cycles in enterprise technology work the same way every time. The technology that promises cheap output is good enough for the demo, good enough for the executive readout, good enough for the first user. The failure modes that take judgment to anticipate show up in the eighteen-to-twenty-four-month window, because that is when the substitution has been running long enough for the gaps to become visible.
What is unusual about this cycle is that the substitution is happening at two layers at the same time. AI is replacing both the cheap code at the implementation layer and the cheap change at the organizational layer. Both substitutions are likely to fail in similar ways, on similar timelines, for similar reasons. The consulting firms appear to have priced this in already. Their clients have not.
The trade is still being made. If you are making it, the question is whether you are making it on the same read the consulting firms are, or the opposite read.
If this maps to a decision you are sitting with
If you are working through an AI rollout that is not converting, an operating model that has not evolved to match the technology, or a transformation function that has been thinned to the point where the AI work is running without a clear change owner, I am glad to compare notes.