Whoever is paid for the outcome should not set the limits. That sounds like a constraint on the most capable people in a business, and it is the opposite: it is what lets them run.
The most capable people now build their own agents. They set targets, wire in tools and measure everything. This is how growth gets made. It is also how governance gets bypassed, because the person who wrote the agent's permissions is the person rewarded for what the agent produces.
What are the surveys describing?
On 15 September EY reported that 47 per cent of 202 senior AI decision-makers at US listed companies with revenue of at least one billion dollars said their organisation had previously bypassed its AI governance process for an urgent deployment. On 15 October IBM reported that 98 per cent of 3,221 executives at organisations of 1,000 or more employees across Europe and the Middle East say board-level responsibility for AI risk has been formally assigned, and that when asked which role is responsible for harm caused by an AI agent, no single answer drew more than 17 per cent. On 19 October Gartner reported that 4 per cent of 2,002 CIOs and technology executives had plans in place to revamp roles for a human-agent workforce. All three firms sell governance or advisory services, so weigh them as vendor surveys. They still point the same way.
Boards have assigned the risk. The gap is below the board, between the person who owns the outcome and the person who should own the limits. In most firms today those are the same person, and the agent runs on a self-issued licence.
The board assigned the risk. Nobody assigned the stop button.
Why does the same hand fail?
A person rewarded for growth will optimise for growth. That is the job, and the firm hired them for it. Asking the same person to also set the ceiling on what their agents may do asks them to work against their own targets, and in a quarter when the numbers are short the ceiling moves.
Biology separates the two signals. The genes that tell a cell to grow and the genes that tell it to stop are different genes, read by different machinery. A tumour is what happens when the stop signal fails while the growth signal keeps firing. The body did not solve this by making growth signals wiser. It gave the stop its own hand.
The usual fix is a policy. The EY figure shows what a policy does under pressure: it is bypassed for the urgent case, and the urgent case is the one that matters. The deeper fix is structural. Clinical trials have known this for decades: the team running the trial does not hold the stopping rule. An independent data monitoring committee holds it, fixed in advance, because the people who most want the trial to succeed are the people least able to call a halt.
A limit set by the person it limits is a target, not a limit.
Does board assignment fix it?
The IBM pairing says no. Ninety-eight per cent assigned at board level, and no agreement on who answers when an agent causes harm. Assignment at the top does not reach the agent. Between the board and the team sits the question nobody has been asked: who, by name, decided what this agent may do, and who, by name, can stop it today?
The ICO is now asking a version of that question in law. Its call for evidence on agentic AI, open until 20 November, asks how organisations allocate controller and processor roles for agents and, as reported, says that decisions about protocols, permissions and access controls may weigh more heavily than training data in determining who controls the means of processing. If that reading holds, the person who set the agent's permissions is the person the regulator will look for. In most firms that is a product owner or an engineer, acting under a licence they issued to themselves.
Last week's argument was that a senior manager must be able to show their challenge. Borrowed assurance, held on someone else's word, fails at the moment someone asks. A self-issued licence is the same failure one layer down: the assurance that an agent's limits are adequate comes from the person who set them and benefits from them.
Permissions are a governance decision wearing an engineer's clothes.
What should a board ask for?
Four names and facts for each material agent, none technical.
First, who owns the outcome: the person whose targets the agent serves.
Second, who sets and holds the limits: what the agent may access, decide and spend, and the authority to stop it. This name must differ from the first.
Third, what the agent may decide without a person, in a sentence a director can read.
Fourth, when it must stop: the thresholds, and who is told when one is reached.
None of this slows the builders. It frees them, because a builder who did not set their own limits cannot be blamed for them, and a board that can name the limit-holder has something to show a regulator, an insurer or a court.
Whoever is paid for the outcome should not hold the stop button.
This is a practice, not a project: a four-line card per agent, reviewed when the agent's permissions change, and a standing agenda line that asks whether the two names are still different. If the ICO asked your firm tomorrow who set the permissions on its most consequential agent, would the answer be the same person who is paid for its results?
Roche-Review. Executive Governance, with Human Observability.
Sources: EY, US AI Risk and Governance Survey (released 15 September 2026, 202 respondents, fieldwork 28 May to 15 June); IBM, EMEA enterprise AI governance study (Censuswide, 3,221 executives, fieldwork 14 to 24 September 2026, released 14 to 15 October); Gartner, 2027 CIO and Technology Executive Agenda (2,002 respondents, released 19 October 2026, as reported by CIO Dive); ICO, agentic AI call for evidence (published 8 October 2026, closes 20 November 2026, as reported).
The Roche-Review is the weekly publication of Dr Ivan Roche FRSS FRSA MInstP, Founder of Otopoetic Limited. Subscribe at roche-review.com.
/
