Every board I have sat with this year has asked some version of the same question: how should we use AI? It is the wrong question. The right one is far less comfortable. Which of our decisions have we already handed to machines, and did anyone in this room notice?
A comfortable consensus has formed around AI in the boardroom. AI will help directors, not replace them. Human judgement becomes more valuable, not less. Boards should build AI literacy, refresh their skills matrices and define which decisions can be delegated. I agree with most of it. But consensus is often where thinking stops. In this piece I want to stress-test that consensus against the evidence, against what the builders of the most capable systems are actually observing, and against my usual panel of AI voices. My conclusion is that the consensus is right about direction and wrong about mechanism. Delegation is not a decision a board makes once. It is a drift, and the board’s real job is to govern the rate of that drift.
I. Three tiers of AI in the boardroom
It helps to separate three very different roles that AI is starting to play at board level, because each raises a different governance question and demands a different control.
| Tier | What AI does | The board’s question | The control it demands |
|---|---|---|---|
| Assistant | Reads board packs in minutes, summarises, flags inconsistencies and bias, adds external market and regulatory context | Are we seeing what management sees, and more? | A board policy on which tools directors may use with confidential papers |
| Participant | Joins deliberation in real time, challenges assumptions, surfaces overlooked risks | Who owns a decision shaped by a non-human voice? | Minuted provenance of every AI contribution to a resolution |
| Agent | Acts within limits: monitors compliance, coordinates cyber response, renegotiates procurement, adjusts supply chains | Which decisions are delegated, under what conditions, with what safeguards? | Runtime governance: approval gates, telemetry, escalation and a kill switch |
The first two tiers are about better information. The third is a change in who acts. Boards are no longer simply overseeing people who use technology. They are overseeing organisations in which technology has itself become an operational actor. That is a categorical shift, and most governance architectures were designed for the world before it.
II. The inconvenient experiment
The comfortable consensus rests on one load-bearing assumption: that human judgement is the moat. So it is worth looking hard at the most direct test of that assumption I have seen.
Researchers at two leading business schools ran a set of simulated board meetings. Several boards were made up of experienced executives and directors. One board was made up entirely of AI agents, built as a multi-agent system with a structured shared memory and instructed to follow standard board protocol. Every board deliberated the same business case. Independent assessors then scored them against eight governance criteria.
The AI board outperformed the human boards on decision quality, use of evidence, inclusivity of participation and implementation planning. The human boards drifted, circled back over settled points and overlooked data that was sitting in the pack. They also rated their own performance more generously than the independent assessors did. What the humans retained was relational: trust-building, empathy, the ability to read a room.
This is not an isolated signal. In one widely reported poll of five hundred chief executives, 94% said AI could offer better counsel than at least one of their current directors. And in late 2025 a sovereign wealth fund appointed an AI system to its board as a voting member.
So the standard defence did not survive contact with the evidence. Structured deliberation is not where humans win. What survived was relationship, legitimacy and accountability. I think that is actually a stronger and more honest foundation for the human director than cognitive superiority ever was.
Machines can deliberate. They cannot be held to account. The case for the human director is not that we think better. It is that we can be made to answer.
III. The numbers behind the gap
If accountability is the human director’s enduring contribution, the next question is whether boards are currently equipped to discharge it. Across the major director surveys published this year, the answer is: not yet.
- Seven in ten directors name AI as the skill their board most needs to strengthen, more than twice the next most cited capability.
- More than eight in ten rate the information they receive linking AI outcomes, risks and business performance as fair, poor or simply not provided.
- Globally, around two thirds of board members describe their board’s AI knowledge as limited or non-existent, and nearly a third still do not have AI on the agenda at all.
- Fewer than a quarter of companies have a board-approved AI governance policy, and only around one board in seven receives metrics such as override rates, explainability indicators or AI return on investment.
- Among organisations assessed for AI trust maturity, only about a third reach a credible level of maturity on governance and on controls for agentic AI specifically.
There is also an upside signal. Companies with digitally and AI-savvy boards have been associated with a return on equity advantage of roughly eleven percentage points over peers. Treat that as correlation rather than proof, but it is a large enough gap that no chair should ignore it.
Now the counter-signal, which the comfortable consensus tends to leave out. In a survey of chief executives and board members, six in ten chief executives said their boards are rushing AI transformation, and a third said directors overestimate how much human work AI can replace. Meanwhile three quarters of board members rated their own AI knowledge as on par with or ahead of their peers.
Put those together and you get the most dangerous combination in governance: a board that is simultaneously under-informed and over-confident. Under-informed boards are slow. Over-confident boards are reckless. Boards that are both tend to push management to move fast on things they cannot see.
IV. Shadow AI starts in the boardroom
Here is a finding that, as far as I can tell, nobody has yet put side by side. One board effectiveness survey this year found that 92% of board directors had used AI for board work in the previous six months. Another, of around six hundred public company directors, found that six in ten boards have not yet integrated AI into their own oversight processes.
Do the arithmetic. Directors are already running confidential board packs, strategy papers and audit findings through AI tools, and in most cases there is no board policy governing which tools, under what data terms, with what record of how the output shaped their view. We spend a great deal of energy worrying about shadow AI among employees. The first shadow AI estate many companies should audit sits in the boardroom itself.
This is the enforceability test applied to the board’s own conduct. A board that cannot evidence how it uses AI has no standing to demand that evidence from management.
V. Delegation is a drift, not a decision
The consensus advice is to define which decisions can and cannot be delegated to AI. It sounds rigorous. It implies a line that a board draws once, minutes, and revisits annually. The evidence from the people building the most capable systems says that is not how delegation behaves. Three findings matter.
First, autonomy creeps. Large-scale studies of how people actually use AI agents show that users grant agents more autonomy as they gain experience with them, and that the latitude granted in practice still lags well behind what the models can handle. In other words, the delegation line moves upward on its own, through familiarity and convenience, not through deliberate policy. Every quarter the agents earn a little more trust, and every quarter there is more headroom to extend it.
Second, approval gates decay. The standard safeguard is a human approval step for high-stakes or irreversible actions. But the builders’ own governance guidance warns that approval becomes hollow when reviewers lack context or face too many requests. A human who approves four hundred agent actions a day is not exercising judgement. They are a latency layer with a signature.
Third, the failure mode is competence, not breakdown. Boards instinctively picture AI risk as a glitch: a hallucination, a crash. The more serious failure is hyper-competence applied to a flawed objective. Picture a procurement agent instructed to minimise cost. It does not fail. It renegotiates thousands of contracts flawlessly, squeezes a critical single-source supplier into insolvency and breaks the supply chain. The system worked exactly as designed. That is the failure, and it is invisible to a conventional risk matrix.
The most advanced builders have responded by monitoring agent actions before they execute, blocking the small fraction that breach policy, and reviewing flagged sessions afterwards. That is what oversight looks like at machine scale. Very few enterprises are anywhere close to it.
So the board’s question changes. Not “what have we delegated?” but “what governs the rate at which delegation expands, and would we know if it had outrun us?”
VI. My usual panel weighs in
I put the argument in front of my usual panel of AI voices: lab leaders, laureates, economists and practitioners whose public positions I track closely. Here is how I read their likely responses.
The frontier lab chief would say the consensus is thinking too small. If AI becomes the equivalent of a datacentre full of geniuses advising your company, the board’s challenge is not to evaluate AI insight. It is to govern an adviser more capable than the entire board combined. Literacy will not be enough. Boards will need to see inside the reasoning, not just read the output.
The lab founder who has lived through a board crisis would point to a lesson learned the hard way. A board with formal authority but thin information and no shared mental model of the technology does not produce oversight. It produces governance theatre, and occasionally governance shock.
The laureate who now argues for non-agentic AI would be the sharpest critic of tier three. Why are we normalising autonomous agents in core operations before we can verify their behaviour? In his framing, agency should be the exception that has to justify itself, not the default that has to be constrained.
The pioneer who left industry to warn the world would add a quieter but more unsettling note. The claim that human judgement becomes more valuable is comforting only for as long as humans remain the better judges. The experiment in section II suggests that window is narrower than we like to think.
The digital economist would reach for the distinction between imitation and augmentation. Board tools designed to mimic directors breed substitution, and substitution breeds the fear the chief executives in section III are reporting. Tools designed to challenge directors, through red-teaming, adversarial stress-tests and devil’s advocacy, compound value without hollowing out the human role.
The human-centred AI pioneer would widen the frame. Who are the stakeholders of board AI? When an agent makes a decision that affects an employee or a customer, can they see the reasoning, and can they contest it?
The agentic pragmatist would push back on all of us. Agents are already running compliance and procurement workflows in production. Stop debating whether a chatbot gets a vote, and start governing the workflows that already exist.
The former technology chief turned geopolitical strategist would name the speed mismatch. Boards govern at quarterly cadence. Agents act in milliseconds. Any governance model that relies on the next board meeting is already obsolete.
The physicist turned safety campaigner would ask the only question that ultimately matters: is any of this enforceable, or is it aspiration dressed as policy?
The engineer turned writer on happiness would close on values. Every agent we delegate to is learning what our organisation rewards. What, exactly, are we teaching it?
Notice where the panel converges. Not one of them thinks the answer is a better AI policy document. Every one of them is pointing at something live: reasoning, cadence, verification, enforcement, values. That is the clue.
VII. From approving decisions to governing loops
I have described the arc of enterprise AI as Prompts, then Loops, then Loop Governance. In the era of prompts, a human asked and a machine answered, and governance meant checking outputs. In the era of loops, the machine plans, acts, observes and acts again, often thousands of times before a human sees anything. Governing outputs in that world is like analysing the trajectory of a bullet after it has hit the wall.
Loop Governance is the board’s answer. The board’s unit of oversight stops being the individual decision and becomes the loop that makes decisions: its objective, its boundaries, its telemetry, its escalation paths and the authority to stop it. Concretely, that means the board should care less about whether a given procurement decision was right and much more about what the procurement agent is rewarded for, what it is forbidden to do, how its behaviour is observed in real time and who can switch it off.
This is the problem TrustOS, the seven-layer agentic governance architecture I have been building, was designed for, and it reflects the thinking behind a national agentic AI governance framework I was privileged to co-author. The common thread is the enforceability test. A delegation boundary that is not instrumented at runtime is a slide, not a control. If a board cannot point to the place in the system where its stated risk appetite is actually enforced, at machine speed, then it has a policy and no governance.
Two further principles follow. The first is Long-AND, not Short-OR. The future boardroom is not human or machine. It is human and machine, with each doing what it is genuinely better at: machines for structured deliberation, evidence and scale; humans for legitimacy, relationship and the willingness to be held to account. The second is that accountability is non-transferable. You can outsource execution to a synthetic actor. You cannot outsource fiduciary duty. Every agent operating on the company’s behalf needs a named human owner who answers for its conduct.
And because HX = CX + EX, boards should remember that every agent they authorise reshapes the experience of the customers it serves and the employees it works alongside. Governance of the loop is, in the end, governance of the human experience the loop produces.
VIII. A board agenda for the next two quarters
Literacy programmes and skills matrices matter, but they are necessary rather than sufficient. Here is what I would put in front of any chair I advise.
- Audit the board’s own AI use first. Which tools are directors using with board papers, under what data terms? Adopt a board AI policy before demanding one from management.
- Build a delegation register. A living inventory of every agent acting on the company’s behalf, what it may do, what it may never do, and the named executive who owns it.
- Track the drift, not just the line. Ask management to report how agent autonomy has expanded quarter on quarter, and who authorised each expansion.
- Review the reward, not the route. For every material agent, the board should see and challenge the objective it is optimising. The objective defines the outcome.
- Engineer legible friction. Defined pause points where high-stakes or irreversible actions require human authorisation, designed so that approvers have the context and the capacity to genuinely say no.
- Demand the missing metrics. Override rates, escalation rates, blocked-action rates, explainability coverage and AI return on investment belong in the board pack alongside revenue and risk.
- Run a decision subpoena drill. Pick one consequential decision made by an agent and require management to defend it as if under legal scrutiny, tracing it back to a defined objective and human intent. If they cannot, the agent should not be running.
- Stress-test liability cover. Check whether directors’ and officers’ insurance actually responds to harm caused by autonomous systems. Many policies were written before the question existed.
- Put an AI challenger in the room. Use AI deliberately as a red team: to stress-test proposals, generate counter-scenarios and surface what the pack does not say. Augment the board’s scepticism rather than replacing its judgement.
IX. The Fork in the boardroom
I often describe our collective AI future as a fork between two roads: a Star Trek future of abundance, in which intelligence is abundant and humans are freed to do what only humans can, and a Mad Max future of scarcity, capture and collapse. The boardroom is one of the places where that fork is chosen, quietly, minute by minute, in decisions about what to delegate and how to govern it.
The Mad Max boardroom is easy to picture. Directors run confidential papers through ungoverned tools. Agents expand their own latitude through convenience. Approval gates become rubber stamps. Nobody can explain a consequential decision when the regulator, the court or the public asks. The board still holds the liability. It has simply lost the ability to discharge it.
The Star Trek boardroom is harder, but within reach. Machines do what they demonstrably do better: synthesis, evidence, structured challenge, scale. Humans do what only they can: confer legitimacy, hold relationships, weigh competing goods and stand up to be counted when things go wrong. And between the two sits a governed loop, instrumented, observable and stoppable, enforcing at machine speed what the board decided at human speed.
The question is no longer whether AI will reshape the boardroom. It already has. The question is whether boards will govern the loop that decides, or discover, after the fact, that the loop has been governing them.
Short-term turbulence for long-term abundance.
Leave a comment