Who should own AI strategy in an organisation?

AI strategy usually arrives in an organisation attached to somebody: the transformation director, the head of process, the chief technology officer. Each of them owns a real part of it. None of them owns the part that decides what the organisation will still be able to do in five years, and that part is rarely assigned to anyone.

The ownership argument is almost never about ownership. Underneath it sits a question most organisations have not answered out loud, and often have not noticed they are answering: is this an augmentation play or a replacement play? Where AI sits in the structure follows from that, usually without anybody deciding it. So the reporting line ends up being a statement of belief that nobody had to write down and nobody can be held to.

Augmentation and automation are not two words for the same thing#

Automation removes a task from a person. Augmentation changes how a person does a task they keep. The distinction sounds academic until you notice that the two produce opposite results on the measures a board actually watches, and that most organisations are pursuing both at once without saying so.

The useful test is not what the technology does. It is what the business case counts as the benefit. If the saving is headcount, the play is replacement whatever the announcement says. If the benefit is throughput, quality or reach at constant headcount, it is augmentation, and the two require different measurement, different governance and, as it turns out, different people in charge.

The choice is not a philosophy. It has measurable, opposite effects#

Here the argument stops being a matter of taste. Autor and Thompson, in Expertise (2025), built a content-agnostic measure of task expertise and applied it to four decades of task data across 303 US occupations from 1980 to 2018. What they found is the most useful single result for anybody deciding where to point this technology.

• Automating the LESS expert tasks in a job raised wages and reduced employment. Fewer people, each doing the harder part, paid more.
• Automating the EXPERT tasks lowered wages and increased employment. More people, each doing the easier remainder, paid less.

Read that twice, because it inverts the usual assumption. Automation does not simply reduce headcount. Which tasks you remove decides whether the role appreciates or commoditises, and the two paths point in opposite directions on both pay and employment. That is a strategic choice with a distribution consequence. At present it gets made inside tool-selection decisions, by people nobody asked to make it.

The authors name one limit themselves: the data ends in 2018, so this is a lens rather than a forecast for generative AI. It gives you a question to ask about your own roles, not an answer about them.

Can an organisation do both at once?#

Yes, and most do. The combination is not the problem. Trouble comes from running the two through different functions against different targets, so nobody holds the trade. Operations is measured on cost per unit. A capability or L and D function, if it is involved at all, is measured on completion rates. Neither is measured on whether the organisation can still do the work when the system is unavailable, wrong, or repriced.

Pursuing both deliberately is coherent. Automate the routine perimeter, augment the expert core, and say which is which. Pursuing both by accident produces an organisation with two AI strategies, one budget and no account of the interaction between them.

Where an organisation files AI tells you what it believes about its people#

Structure is not neutral. Each common placement is competent at one thing and blind to another. The blindness follows a pattern worth naming.

• Technology. Excellent at capability of the system, security and integration. Structurally not accountable for what happens to the people using it, and rarely asked to be.
• Transformation or change. Good at sequencing, adoption and programme discipline. Tends to treat resistance as a change-management problem when it is sometimes an accurate signal that the work has been designed badly.
• Operations. Best at extracting the measurable saving. Put it here and the replacement play gets selected without a word being said. Organisations that file people under operations rather than treating them as strategic tend to file AI there too. Nobody then owns the capability question.
• The business unit. Closest to where the judgement actually lives, and usually without the mandate, budget or technical depth to act on it.

None of these is wrong. The point is that the placement answers the augmentation-or-replacement question administratively, before anyone debates it, and then the debate never happens because the answer already exists in the org chart.

So who should own the strategy?#

There are two parts to this. Most people dislike the first.

Ownership of the judgement allocation is not delegable, and that part belongs to the chief executive. Deciding which decisions stay human as adoption increases is a decision about what the organisation is, not about which tools it buys. It sets the risk position, the capability position and, per Autor and Thompson, the shape of the workforce. Nobody below the chief executive can make that call across functions, and no function will make it against its own metric.

The counter-argument deserves stating, because it is strong. Chief executives own everything and therefore own nothing; a subject "owned" at that level with no operating capacity attached becomes a slide rather than a programme. So the workable version is narrower: the chief executive owns the choice and the standard, and delegates the delivery. Specifically, the chief executive should personally hold two things and can reasonably hand over the rest.

• The declared position. Whether this is augmentation or replacement, by domain, in writing, with the trade accepted rather than implied.
• The capability floor. What the organisation must still be able to do unaided in three years, and who reports on whether it still can.

Everything else, tooling, integration, sequencing, training, vendor management, belongs with the functions that already do those things well. This is also the answer to whether AI is a technology strategy or a people strategy: it gets filed as the first and behaves like the second. That mismatch strands a great deal of work.

Who should own implementation, and what a good team looks like#

Implementation is a different question from strategy, with a different answer. It should sit with whoever owns the work being changed, supported by technology, rather than with technology supported by the business. The reason for that is evidential rather than political.

Dell’Acqua and colleagues, in a field experiment with 758 BCG consultants, found that inside the model’s competence AI-assisted work was dramatically better and faster, while just outside it the same people performed worse than consultants using no AI at all. The boundary is jagged rather than smooth, and where it runs is local to each domain. Nobody in a central function can know where it falls in a claims team, a ward or a drafting practice. Only the people doing that work can find it, which is an argument for putting implementation next to them.

On the composition of the team, one finding does most of the work. Vaccaro, Almaatouq and Malone’s preregistered meta-analysis of 106 experimental studies and 370 effect sizes found that human and AI combinations performed significantly worse on average than the better of human alone or AI alone, at Hedges’ g of -0.23, with the losses concentrated in decision-making and the gains in content creation. Their conclusion is the sentence every implementation team should have on the wall: adding a human is not a control, and undesigned pairing can subtract.

That gives a concrete test for whether a team is any good, and it has nothing to do with the skill mix. Ask whether anybody on it is accountable for the design of the human-machine split rather than for shipping the tool. In practice a serviceable team has four things: somebody who does the actual work, somebody who can build, somebody who can change process and permissions, and somebody accountable for capability rather than delivery. The fourth is the one nearly always missing, and its absence is why so many teams can tell you adoption and cannot tell you whether anyone got better at anything.

Why pilots succeed and rollouts fail#

Almost every organisation has seen this and few explain it. Two candidate mechanisms are worth testing against your own programme. Both can be true at once.

The first is selection. Pilots are staffed by volunteers who already had the judgement to spot a wrong answer, the population for which the tool is safest. Rollout removes that filter. The Vaccaro result predicts what follows: the pairing gained where humans beat the AI and lost where the AI beat humans, so extending it to people who cannot tell the difference reliably moves the average the wrong way.

The second is that pilots optimise a task and rollouts change a system. Humlum and Vestergaard, linking adoption surveys to administrative records for roughly 25,000 workers across 7,000 Danish workplaces in eleven exposed occupations, found precise null effects on earnings and hours two years after ChatGPT, ruling out effects larger than two per cent, alongside substantial task reorganisation and new tasks in AI oversight and integration. The work moved considerably. The numbers a board watches did not. If your rollout looks like it is failing, check whether it is instead succeeding at something nobody assigned it to do.

Augmentation is achievable under specific conditions#

It is worth being clear that this is not an argument that automation hollows out work by necessity. The cleanest counter-case in this evidence base comes from care rather than knowledge work. Lee, Iizuka and Eggleston, using regional robot subsidies as an instrument across a panel of Japanese nursing homes, found that robot adoption raised employment and improved retention, most strongly for non-regular staff, moved worker effort towards direct care, and improved quality on hard measures: less use of physical restraint and fewer pressure ulcers.

The conditions matter and the authors name them. Japanese long-term care faces an acute labour shortage, so the robots substituted for vacancies rather than for people. That is a particular situation and it does not generalise on its own. What it does establish is that the hollowing-out is a consequence of choices about which tasks to remove, not a property of the technology.

The cost of the choice usually falls on people who are not in the room#

One more result belongs here, because it changes who should be consulted. Dauth and colleagues, using German administrative worker and plant data from 1994 to 2014 with a shift-share instrument for robot exposure, found that incumbent workers largely kept their jobs and moved into new, higher-quality tasks inside their original plants. The cost fell instead on young labour-market entrants, who shifted away from vocational manufacturing training altogether.

Twenty years of manufacturing data making the missing rungs argument before anybody applied it to knowledge work. The people in the room when an AI strategy is signed off are incumbents, and on this evidence incumbents are the group it treats best. The consequence appears one intake later, in people who were never consulted and are not yet employed. Generative AI is not industrial robotics and the technologies differ substantially, so this is a warning about where to look rather than a prediction.

What to actually do#

Four things, in this order.

• Write the position down. Augmentation or replacement, by domain. An organisation that cannot state this in a sentence per domain has a procurement plan rather than a strategy.
• Check it against the business case. If the stated position is augmentation and the benefit line is headcount, the business case is the real strategy. Treat the statement as decoration.
• Name the capability floor and give it an owner. Cost has an owner and a number. Capability usually has neither, so capability is what erodes without anyone deciding it should.
• Ask which tasks are being removed, not how many. On the Autor and Thompson lens, the expert ones and the routine ones move pay and employment in opposite directions. Most reporting counts tasks automated and never asks which.

Where this sits in my own argument#

This is drift versus design applied to the org chart. My argument throughout this research is that organisations do not decide to hand over judgement; they discover afterwards that they have. Ownership is where that happens first, because the reporting line gets set before anybody frames it as a decision, and after that the question stops being asked.

The capability floor above is the same instrument as a capability audit, and the thing it protects against is capability debt: cost incurred now, paid later, invisible on every current measure. Deciding this in advance, rather than discovering it afterwards, is what I get hired to help with.

What I have observed in organisations#

In the early days of implementation I watched an accounting firm hand AI to its technology team. They used it for email systems. Nothing else had been thought through, because nothing else was their job, and nobody had asked them to think about the rest. The placement had answered the question before anybody framed it as one.

What eventually moved them was not an internal argument. It was watching other accounting firms, in the United States and in the UK, start offering services they could not match. Only then did they bring somebody in for AI transformation.

Even after that, most of the leadership team did not use the tools. Some partners were slow to adopt ChatGPT or Copilot for anything, including email. Rather than the senior team working out what it meant, it turned into a process matter and was pushed down the agenda. It moved when the chair of the board changed the ethos around AI, and not before. That is the argument on this page in one organisation: the question sat with people who could not answer it, until somebody at the top made it theirs.

What this page does not claim#

It does not claim there is one correct owner for every organisation. Structures differ, and a placement that works in a 200-person firm with a technical chief executive will not transfer to a 40,000-person group. What it claims is narrower: that the placement is currently being decided by inheritance rather than by argument, and that it silently answers a strategic question that deserves a deliberate answer.

It does not claim the evidence settles the augmentation question for generative AI. Autor and Thompson stop in 2018. Dauth is industrial robots. Lee is Japanese care homes under acute shortage. Each is a lens on a mechanism rather than a forecast, so the mechanisms are well evidenced and their application to this technology is not.

And it does not claim that replacement is illegitimate. There are roles where automating the expert part is the right call and the organisation should say so plainly, price the consequence and take it. The argument here is against choosing by default, not against choosing.