The standard advice in 2026 is that any serious enterprise needs an AI Transformation Office: a dedicated unit with a senior leader, workstreams, governance and a roadmap. Major advisors sell versions of it, board papers recommend it, and many large companies are staffing one. The case looks airtight.
I think it is the wrong answer when the office is meant to be permanent. In my experience, creating a lasting central office to own a technology shift is one of the more reliable signs that the transformation is already in trouble. I have watched versions of this across Digital, Cloud and Agile Transformation Offices, and the arc repeats: a capable executive, a large mandate without underlying authority, a glowing first review, and a restructure or quiet rename within two years. When the work that mattered happened, it happened in the line.
This article explains why a permanent AI Transformation Office repeats that pattern in its worst form, and what a credible replacement looks like.
The pattern is older than AI
Look at the arc of three earlier transformation offices before getting to AI.
Digital Transformation Offices spread between 2014 and 2019. Consultancies published the playbook and recruiters built practices around the Chief Digital Officer seat. The promise was clear and the execution was uneven, and by 2021 many of the original CDO roles had been removed, restructured or merged into the CIO or CMO role. The digitization of operating models eventually happened in the line, while the office mostly absorbed political risk during the transition.
Cloud Transformation Offices followed between 2018 and 2022, with the same blueprint, vendor partnerships and five-year roadmaps. The companies that actually moved to the cloud were rarely the ones with the largest office. They were the ones whose CIO and CFO decided that the data centre lease would not be renewed. Where an office existed, it often slowed the move by building governance around a decision that had not yet been made.
Agile Transformation Offices followed between 2020 and 2023, on a smaller scale. Companies got faster when product groups received P&L authority and the freedom to staff their own teams. In many cases the office became the place the line went to ask permission, which is the opposite of agility.
Each office started from the same logic: the shift is too important to leave to the line, so centralize it, appoint a senior leader, build a roadmap and govern carefully. Each tended to produce the same result. The office became the bottleneck it was created to remove, and the slowness was mistaken for governance and the governance for transformation.
Not every transformation office is harmful. The good ones I have seen were small, short-lived and explicitly disposable from the start. The harm came from making them permanent.
Four functions a transformation office ends up performing
Watch a typical transformation office for a year and four functions fill its calendar.
The first is optics. The office produces what the executive committee shows the board and what the board shows analysts: strategy decks, capability heat maps, value trees, North Star metrics and two-by-two matrices. None of them changes a workflow, and all of them serve the political purpose of the role.
The second is absorbing blame. When AI underperforms, the post-mortem needs somewhere to land other than the CEO or the line, and the office is a convenient target. I described the executive version of this in Most Chief AI Officers Are Hired to Fail.
The third is vendor management. Many offices end up running procurement for AI vendors, with RFPs in flight, a preferred-vendor framework and a monthly governance committee. That is purchasing, and the build, buy or partner decision belongs with the unit that owns the commercial outcome. The Build vs. Buy vs. Partner tool is designed for that unit-level decision.
The fourth is controlling speed, the function nobody puts on a slide. Every initiative routes through the office and gets queued, prioritized, governed and softened. The obvious wins land later than they would under unit ownership, and the harder ones may not land at all.
These functions can have value, and none of them is transformation. In practice the office often contains the ambition the rest of the organization was supposed to deliver.
Why AI makes the mismatch worse
AI does differ from the earlier shifts, and the difference makes the problem worse. The reason is clock speed.
A transformation office runs on the enterprise calendar: annual planning, quarterly steering committees, eighteen-month roadmaps and multi-year programmes. Those cadences exist for good reasons, because they allow alignment and accountability across a complex organization.
AI capability does not follow them. METR's January 2026 update fitted a doubling time of about three months for the length of tasks frontier models can complete, vendor pricing moves quickly, and open-source releases land within weeks. Architectural patterns that were right in February can be obsolete by August.
An eighteen-month roadmap cannot govern a curve that moves that fast. In practice, the office wins the political argument while the technology moves on. The office produces careful artefacts about capability that has already been superseded, and the line either deploys without the office or stops trying. The pattern MIT NANDA documented, in which about 95% of generative AI pilots return no measurable P&L impact, is partly a symptom of this mismatch.
I described the pilot version of this in AI Pilots Are a Form of Procrastination. A permanent transformation office is the organization-design version of the same problem.
The centre-of-accountability trap
A second structural problem can do even more damage. When a CEO appoints a senior leader and an office to "own AI for the group", every line executive reads the same signal: AI is the central team's problem, and my problem is the quarter. If the central team's work helps, the line accepts it. If it does not, nobody in the line is on the hook.
That signal undermines distributed AI. AI rarely delivers value as a horizontal capability added to the organization. It delivers value when operators change how work is done in their own workflows. The change is local, so the accountability has to be local too, and a central owner makes it harder for anyone local to feel responsible.
I have seen this in several groups: a strong central AI team, well-written playbooks and real platform investment, with little measurable change in the line, because the line was waiting. The central team had absorbed both the credit and the responsibility. The problem was structural rather than a failure of leadership, and the organization chart told the line the truth even when the communications said something else.
The dissolution model
If a permanent central office fails for this shift, the alternative is what I call the dissolution model. It has three structural elements and one cultural one.
The first is a 90-day catalyst team with a hard sunset. You do need a small central team in the first quarter, and its job is to start work rather than to govern it. Five to ten of the company's best operators join on full-time secondment, with explicit authority to bypass normal governance for that period. The team has one deliverable: three production AI workflows in three business units, each tied to a real P&L outcome, by day 90. On day 91 the team disbands and its leader returns to a line role with a P&L. There is no extension clause and no phase two.
The second is a P&L-embedded AI lead in every business unit, named on day one. These are operators with a number, for whom AI is one lever among several, rather than part-time "AI champions". They report to the unit head, their budget comes from the unit's P&L, and they are accountable for unit economics rather than AI activity.
The third is a board scorecard tied to reshaped roles and changed unit economics. The most useful thing a board can do is refuse to accept "capability built" as a metric and ask for two numbers by unit and by quarter: how many roles have been retired or reshaped because of AI, and how far the unit economics have moved.
The cultural element is that the central team must be visibly disposable. The announcement of the catalyst team should include the date it dissolves, it should be funded for that period only, and its leader should accept the role on that understanding. Without that, the team will find reasons to extend, expand and entrench, as earlier offices did.
A useful starting point for a board considering this path is the AI Readiness Assessment, which shows whether the operating model can absorb distributed AI or whether underlying gaps would force centralization anyway.
The strongest objection
A CIO or chief risk officer will say that decentralization without a centre produces shadow AI, duplicated tools, inconsistent risk controls and a scramble for scarce talent. Someone has to own the platform, the security standards and the model governance.
They are right about those functions, and they should have a permanent home. A shared platform team, common security and risk standards and a small pool of scarce specialists can and should last. My objection is to a permanent office that owns the transformation itself, meaning the targets, the use cases and the credit. That mandate has to move to the business units, with the platform and standards supporting them.
A diagnostic for boards and CEOs
If you are being asked to approve an AI Transformation Office in the next 90 days, run the proposal through six questions first.
- Does it include a named dissolution date for the central office, in writing, before anyone is hired?
- Are the AI leads in each business unit named, accountable for a P&L and reporting to the unit head from day one?
- Is the central office's budget separate from the units' AI budgets and capped at the dissolution date?
- Is the board scorecard tied to reshaped roles and changes in unit economics, rather than to capability built or use cases identified?
- Does the office have explicit authority to bypass normal governance during its life, or will it route through the committees that created the bottleneck?
- Will the proposed leader accept the role on the understanding that it ends?
If the answer to any of these is "we will work that out later", send the proposal back until the answers are yes.
The harder conversation for a board is whether it will require every business unit to reshape roles, change unit economics and absorb AI into its P&L over the next three years, on a quarterly scorecard. If it will not, no transformation office will make up for that. If it will, much of what the office was meant to do becomes unnecessary.
Consider the head of one AI Transformation Office in Madrid. The office has grown, the deputy now has a deputy, and the workstream count has risen from twelve to nineteen. The executive committee still likes the readouts. Two of the original three lighthouse use cases have been quietly de-scoped, none is in production, and the processes they were meant to replace are still running. The failure lies in a structure designed to look like a transformation rather than produce one, more than in the person running it.
Monday move
If your company already has an AI Transformation Office, write down its dissolution date. If there is none, draft one, together with the three business-unit leads who would take over its mandate, and bring both to the next executive committee. If you are about to approve an office, answer the six questions above in writing before you sign.