Part of our guide: AI transformation for COOs
There is a gap in most mid-market businesses that nobody has named, and in regulated firms it has a rule number attached.
The FCA's AI Update sets out that there is no dedicated senior manager function for AI, so existing accountability applies instead. It says that technology systems are normally the responsibility of SMF24, the Chief Operations function. Under the overall responsibility rule, any use of AI falls within the scope of a senior manager's responsibilities whether or not anyone has assigned it.
Now set that against how AI is actually owned, because a 2026 survey of AI and data leaders by Heidrick and Struggles mapped the functions owning AI strategy. The CIO, CTO or chief digital officer held it at 30 per cent, the chief data and analytics officer at 23 per cent and a chief AI officer at 21 per cent. The remainder sat with an AI centre of excellence at 7 per cent, the CEO at 6 per cent and the CFO at 4 per cent.
The COO owned it in zero per cent of firms.
So in a regulated business the accountability lands on the Chief Operations function by default, and the strategy sits somewhere else by design. That gap is not a governance nicety, and it is the specific structural reason AI programmes stall between a technology function that can build and an operations function that owns the consequences.
Why this is not just a financial services problem
If you are not FCA regulated, the rule does not bind you, and the shape of the problem still holds.
The COO owns the processes AI is supposed to improve, the people whose work it changes, the service levels it affects and the operational risk when it misbehaves. When an AI system produces a wrong answer at scale, nobody calls the head of data, and the call goes to whoever owns the process.
The Financial Reporting Council's Provision 29 sharpens this for listed companies, and for financial years beginning on or after 1 January 2026 boards must declare the effectiveness of their material controls, covering financial, operational, reporting and compliance controls. The first declarations appear in reports published in 2027, and AI is not named, so it is in scope only where the board judges the controls material. Auditors do not test the declaration and give no assurance on it, so the board's own judgement carries the declaration.
If AI now runs part of a material process in your business, someone has to be able to say the controls around it worked, and in practice that someone is you.
The evidence that ownership is the binding constraint
The productivity research points in one direction with unusual consistency.
BCG's July 2026 CEO survey found that nearly nine in ten CEOs saw cost or revenue benefits in targeted areas. Only 26 per cent had embedded AI into a broader business transformation, and only 14 per cent had clearly defined P&L impact for all AI initiatives. High performers were roughly seven times more likely to have redesigned workflows and 2.4 times more likely to put their best people on AI work.
BCG's survey of 11,749 workers found that workflow redesign raised the likelihood of a measurable improvement by 24 percentage points. It also found that 66 per cent of AI users receive limited or no guidance on how to redirect the time they save. Over half do not redirect it into strategic work, and 47 per cent now spend more time managing and directing AI than doing the work themselves.
A working paper from Bank of England, Stanford and Bundesbank researchers surveying around 6,000 executives found that nine in ten reported no impact on employment or productivity at their own firm over three years, despite 69 per cent of firms using AI.
Read those together and the conclusion is uncomfortable and simple, because in all three the technology is not the constraint. Time saved is not money saved until somebody redesigns the process to collect it, and that redesign is an operations job sitting in a function that does not own it.
The finding that should end the headcount conversation
Gartner surveyed 350 executives at companies with over a billion dollars of revenue in May 2026, and eighty per cent of those who had piloted AI reported workforce reductions. There was no correlation between those reductions and higher AI returns, with reduction rates near identical among firms reporting high returns and firms reporting low or negative returns.
Firms are cutting headcount, and the survey found no correlation between those reductions and reported AI returns.
The UK official data says something similar from the other direction, and the ONS finds around 5 per cent of AI-using businesses reporting a headcount reduction because of AI. That figure rises to 7 per cent among those with 10 or more employees, and about half report no workforce impact at all. The Bank of England's Agents report that AI investment is usually aimed at improving productivity rather than reducing headcount, and that reductions happen mostly through natural attrition.
If your AI business case rests on a headcount line, it is resting on the one mechanism the evidence does not support.
Where the returns actually are, and where they reverse
Two peer-reviewed studies are worth more than the survey literature combined.
In the Quarterly Journal of Economics, Brynjolfsson, Li and Raymond studied 5,179 customer support agents, where AI assistance raised issues resolved per hour by 14 per cent on average. The gains concentrated in novices, who improved by 34 per cent, while the most experienced and highest-skilled workers saw minimal impact.
In Organization Science, Dell'Acqua and colleagues studied 758 consultants, finding that inside the tasks AI handles well participants completed 12.2 per cent more tasks. They completed it 25.1 per cent faster and at significantly higher quality, and outside that boundary AI users were 19 per cent less likely to produce a correct answer.
A UK government randomised trial run by the AI Security Institute with 500 participants found 25 per cent time saved and a 19 per cent quality improvement, alongside no significant improvement on planning and prioritising, and the authors noted high uncertainty around all estimates.
For a COO these are operating instructions, not curiosities, and AI lifts your least experienced people most, changing your training and supervision model before it changes your headcount. It has a boundary, and past that boundary it makes work worse, so somebody has to define where the boundary sits for each process. It also does not help with judgement and prioritisation, and that is exactly the work that your senior people do.
A COO's agenda for the next two quarters
Claim the accountability explicitly. If you are in a regulated firm it is already yours by default, sitting with you under the overall responsibility rule. It is better to hold it deliberately, with a named owner and a documented scope, than to discover it during a supervisory conversation.
Build an inventory before a strategy. You cannot govern what you cannot list, so the inventory needs to cover every AI system touching a customer, a regulated process or a material control. Each entry carries its owner, its decision boundary and its escalation path. Most of the businesses we work with cannot produce this in under a fortnight, and that is itself the finding.
Define the boundary for each process. You need to know where AI helps, where it stops and who checks, and the Organization Science result makes this a control, not a nicety.
Redesign one process properly. The target is one end-to-end process where you own the outcome (not five pilots), redesigned so the time saved lands somewhere you can measure. This is the single measured differentiator in the evidence.
Name where the saved time goes before you start. Two thirds of users get no guidance on this, and if the answer is not written down before the tool arrives, the time evaporates with your business case.
Separate the AI cost line from the software line. The Bank's Agents find savings being offset by rising software, cloud and licence spend, and if those sit in different budgets your programme will look like it is working while your P&L does not move. We wrote about how to build that picture in the real cost of AI for a COO, and for PE-backed businesses the sponsor's angle is covered in AI strategy for PE-backed businesses.
The uncomfortable summary
AI arrived through technology functions and they own the strategy as a result, and the consequences land in operations, so the rules put the accountability there. Most businesses have not closed that gap, and the productivity data suggests the gap is the reason the returns have not appeared.
The businesses getting results are not the ones with the best models, and they are the ones where somebody who owns the process also owns the change.
That is the work we do in a Grow engagement, embedding senior AI leadership alongside a leadership team to hold strategy, governance and delivery together. If AI has become your problem without ever having become your remit, that is a conversation worth having.
Sources
- FCA, AI Update
- Bank of England, Agents' summary of business conditions (July 2026)
- Brynjolfsson, Li and Raymond, Generative AI at Work (NBER working paper)
- Dell'Acqua et al., Organization Science
- AI Security Institute, measuring AI-driven productivity gains
- ONS, Artificial intelligence in UK businesses
- NBER, Firm-level data on AI (working paper 34836)