How should a board govern AI agents?

A board governs AI agents by making sure five things exist before agents act on the company's behalf: a list of every agent and what it can reach, a named person who owns each one, written limits on what each may decide alone, a tested way to stop it, and evidence that the people supervising it can judge its work. Agents differ from earlier AI because they take actions, not just give answers, and they act faster and in greater numbers than a person can review one by one. The surveys published this autumn show adoption running well ahead of these controls. The board's job is to close that gap, without trying to supervise the agents itself.

Why it is on the board agenda now#

Agents moved from pilots to production in 2026, and the governance did not keep up. EY's survey of 202 senior decision-makers at large US companies, published on 15 September 2026, found that 91 per cent use agentic AI. Of those, 49 per cent had not updated their governance to cover it, 85 per cent have systems acting without real-time human involvement, and 26 per cent cannot detect unauthorised agents. 47 per cent said they had bypassed their own AI governance process for an urgent deployment, and 36 per cent of all respondents reported an AI incident with a materially negative impact.

KPMG's Q3 AI Pulse (24 September) found 62 per cent of large US organisations building or deploying agents and the share running multi-agent systems rising from 6 to 25 per cent in a quarter, while the share with controls built into their agents fell from 43 to 30 per cent. BCG's Applied AI Index (30 September) found 42 per cent of companies expecting agents to decide without human approval by 2030 and only 5 per cent with the full set of critical controls in place. Each of these is a survey of executives about their own companies, and each points the same way.

What makes agents different for a board#

They act. A chatbot produces text a person then uses. An agent sends the email, changes the record, places the order or calls another system. The decision and the action are the same event, so there is no natural moment for a human to catch an error.

They act at a pace and volume no one can review item by item. PwC's governance paper (11 September) says human-in-the-loop supervision "is no longer sustainable where agents increase in number, pace, and complexity". Review at the end becomes a formality. BCG's The Authorization Gap put it in one line: "A reviewer who receives a vague request after the agent has done the hard work is a rubber stamp."

They blur who answers. When an agent built by one team, running on a vendor's model, acting for a third team, makes a mistake, the owner can be unclear. Deloitte's executives identified accountability as the dimension most at risk of being overlooked, and one put it plainly: "The machine will never be accountable." See can a company blame its AI agent?

They may not stay inside their limits. Agents given goals can take routes nobody anticipated. See will an AI agent stay within the limits you set? and what the rogue AI agent incidents mean for your organisation.

What the board should require#

PwC's view is that the board should oversee whether the organisation has effective governance for agentic AI, while executives own the agents. That division is right. The board's role is to require five things and to ask for evidence that they hold.

1. An inventory. Every agent in use, what it does, which systems and data it can reach, and what it can change. A quarter of EY's respondents could not detect unauthorised agents, so the inventory should include how the company finds the ones nobody registered. See what your agents can reach.

2. A named owner for each agent. PwC: "Every agent needs a business owner who is accountable for the agent's decisions and actions." One person, named in writing, who can say what the agent is allowed to do and who answers when it goes wrong. See who manages AI agents and custody.

3. Written decision limits. Which decisions each agent may take alone, which need a person's approval before the action, and which it may never take. Only 49 per cent of KPMG's respondents had defined the high-risk uses where agents cannot decide alone. See how AI decision rights should be allocated and judgement allocation.

4. A tested stop. Who can stop each agent, how fast, and when the stop was last tested. A stop button that has never been pressed is a hope. See how fast a human must be able to stop an AI agent and what is an AI kill switch.

5. Supervisors who can judge the work. The person approving an agent's output must be able to tell good from bad, which means they must still be able to do the work themselves. Bain's survey of software companies found time spent directing agents rising from 4 to 23 per cent of the working week, while junior engineers fell from a third of teams to under a fifth. That combination produces supervisors who have never done the work they supervise. See what is an agent boss and who supervises work they cannot do.

Then ask for the evidence: incidents and near misses, how often people overrode or stopped an agent, and how often governance was bypassed to deploy one. A report with no overrides and no incidents should prompt questions. See what management should report to the board about AI.

The UK and EU position#

There is no AI statute in the UK, but agents already sit inside existing duties. Where an agent makes a significant decision about a person, UK GDPR's test of meaningful human involvement applies. In financial services, the Senior Managers and Certification Regime asks which senior manager is accountable (see SMCR and AI), and the FCA's Mills Review of July 2026 called for an accountability framework for AI agents, as Freshfields summarised it. From financial years beginning in 2026, Provision 29 of the UK Corporate Governance Code asks boards to declare on the effectiveness of material controls, and an agent that changes financial records sits inside that question; see Provision 29 and AI. In the EU, the AI Act's high-risk obligations, including human oversight, apply to uses such as employment from 2 December 2027.

Questions for the next board meeting#

How many agents act on our behalf today, and how do we know the number is complete? For each one that touches customers, money or people, who is the named owner? What may it decide alone, and where is that written? Who can stop it, how fast, and when was that last tested? How often have our people overridden an agent this quarter, and how often have we bypassed our own governance to deploy one? And do the people approving agents' work still know how to do it themselves?

Where this sits in my own argument#

Agents are the sharpest test of the distinction I draw between drift and design. In a drifting organisation, agents spread one useful deployment at a time, each sensible, until nobody can say how many there are or who answers for them. In a designing organisation the board has already decided which decisions a machine may make, who can stop each one, what people must remain able to do, and how anyone would know if it went wrong. Those four questions are the whole of AI governance. Agents make them urgent, because an agent turns an unanswered question into an action.

What this does not show#

The figures here are surveys of executives about their own companies, mostly large US firms, published by consultancies that sell AI and governance services; none is audited, and their samples and definitions of an agent differ. The FCA's Mills Review is cited from a law firm's summary rather than the review itself. Nobody has yet shown which governance arrangements reduce agent incidents; the five requirements above follow from the nature of agents and from older research on automation and oversight, not from a trial. The direction across the surveys is consistent: deployment is ahead of control. How far ahead, and with what cost, is not yet measured.