ICYMI
Should You Add a Bot to Your Board? – AI can improve board decisions, but only if directors use it to open up questions rather than to summarise the pack.
Executive Summary
Boards rely on periodic attention, information provided by management, the judgment of non-executives, and assurance mechanisms.
AI exposes weaknesses in all four of these dependencies.
To close these gaps, boards should make seven changes, from continuous monitoring of controls to directors with more time for the company.
In 1994, Nick Leeson was Barings’ star trader. He also ran his own back office, allowing him to conceal mounting losses. An internal audit highlighted the risk and proposed that these responsibilities be separated. Barings, Britain’s oldest merchant bank, failed to act. The following February, the bank collapsed and was subsequently sold to ING for £1.
Barings had two problems. First, the person taking the risks was marking his own homework. Second, it failed to act on a clear warning. AI can reproduce the first problem and make the second more costly. Some agents keep their own logs and write reports on their work. They can also continue making decisions and commitments at machine speed while a warning goes unheeded. The damage can mount long before the next board meeting.
Why the Board Model Is Breaking
Boards are responsible for setting strategy, allocating capital, appointing leaders, overseeing controls and approving reporting. Four weaknesses in the way boards operate are becoming harder to live with. In each case, AI is at least partly the driver.
Speed. Periodic attention leaves gaps. The largest 150 FTSE companies average 8.2 board meetings a year. Few boards revisit the assumptions behind their strategy at every meeting, but AI can call those assumptions into question in days. In February 2024, Teleperformance’s shares fell almost a fifth during trading after Klarna claimed its chatbot was doing the work of 700 agents. A company’s claim about AI productivity was enough to move a competitor’s share price. Boards need a way to test, between scheduled reviews, whether the assumptions behind their strategy still hold.
Visibility. Management selects most of the information directors receive. AI is increasingly employed to gather and present the data that underpin the board pack. When agents both act and describe what they have done, boards need access to another source of truth. Otherwise, automation can make the company harder to see even as reporting appears to improve. At the Post Office, the findings of a review questioning evidence from the Horizon IT system were not shared with the full board.
Judgment. Directors’ judgment is rarely tested systematically. AI can make a weak argument look convincing, and using it can increase confidence without improving accuracy. In an experiment reported in Harvard Business Review, nearly 300 executives and managers forecast Nvidia’s share price. Those who consulted ChatGPT became more confident but produced worse forecasts than those who consulted colleagues.
Stability. Assurance can become stale. A change to an agent’s model, instructions, permissions or connected systems can undermine a previously effective control. Some agents are built with self-improvement loops that revise their instructions or working methods in response to feedback. The agent in operation can therefore differ markedly from the version that was tested. Boards need to know when those changes invalidate earlier assurance.
Seven Changes for a Board Built for AI
A faster cadence. Boards need to agree triggers that signal a need to reconsider strategy. The chair should require an alert when a major investment assumption is disproven, a competitor changes the economics of the market or risk exceeds an agreed limit. Each alert should identify the decision needed and its urgency. Boards can already call ad hoc meetings. The difference is that these triggers are agreed in advance, tied to specific assumptions and monitored by the board’s AI agent.
An independent source of information. Management should still write board papers, because doing so forces a stock-take. But the audit committee should commission its own AI-assisted analysis through the internal audit function. That requires access to underlying transactions and independently captured records that the operating AI agents cannot overwrite. It is tempting to use one agent to check the work of the first, but two models can be wrong in the same way. The head of internal audit is best placed to take responsibility for validating the analysis and surfacing findings to the board.
A named owner for each material agent. Each material agent, or process that uses agents, needs a business owner accountable for its decisions and commitments. One executive should oversee the company-wide arrangements and ensure that all agents are registered and monitored. The register should keep tabs on what each agent can do, its limits, and who has the authority to stop, override or amend it.
A record of decisions that can be tested. A feedback loop lets boards use AI to learn by analysing their own decisions. Board minutes already explain the reasons for decisions. Boards should add a machine-readable record of expected outcomes, key assumptions, alternatives and any dissent. The board’s AI agent can test these assumptions against new data, compare results with expected outcomes and flag meaningful deviations for the board’s attention. The company secretary should oversee the record, taking legal advice on retention, access and disclosure.
A scorecard. AI makes judgment both more important and more challenging. But judgment can be improved with feedback. Directors should record their forecasts, and how confident they are in their forecasts, before debate. The board should then record its collective reasoning. At agreed milestones, with the help of AI, expectations can be compared with outcomes to examine how the board and individual directors fared. A good decision can have a bad outcome, and vice versa, and one forecast doesn’t represent a pattern. Rather, the purpose is to expose recurring errors, such as overestimating synergies or accepting management’s timelines too readily. Each director’s own results should be private to that director; only the board-level pattern should feed into the board evaluation. These records may have to be disclosed in litigation, so the board should again take legal advice on retention and access.
Directors with higher availability. The increased cadence and requirement to look ‘beyond the pack’ will inevitably require board members to dedicate more time. Deloitte estimated a typical commitment of 35–45 days per director in 2023. A larger commitment implies fewer outside appointments and explicit availability between meetings. That narrows the pool of candidates and raises the cost of each; Deloitte found about 30 per cent of companies were already likely to consider resetting fees. Directors should be able to explain how AI changes the company’s economics and where authority has been delegated to agents. The nominations committee should test that understanding when appointing and evaluating them.
Continuous monitoring, independently checked. Management should monitor agent activity to ensure compliance with agreed permissions and limits. Internal audit should test whether that monitoring is effective and whether escalation paths are clear, with material exceptions escalated to the chair or relevant committee. Provision 29 already expects boards to monitor controls and review their effectiveness at least annually. AI increases the need for evidence that remains current between reviews.
Four Objections
Drift into executive action. A board with better information and more frequent meetings could start to behave like a second management team. But even with these changes, the traditional division of responsibilities can be maintained. Management operates the controls and responds to routine exceptions, whilst the board sets limits, challenges assumptions and intervenes when concerns remain unresolved.
Liability and records. More information creates more work, as it must be assessed and acted on. And recording more detail of the board’s deliberations creates evidence of what directors knew and why they acted. In August 2026, a Delaware judgment cited an AI transcript revealing a motive for a defensive measure that differed from the motive recorded in the official minutes. Boards need to decide explicitly what is recorded, ensure their records accurately reflect their reasoning, and take legal advice on disclosure obligations.
Over-reliance on agents. Directors should use AI to raise questions, but must reach their own conclusions. Different AI models can expose disagreements, but agreement between models should not be treated as independent confirmation, as models can make correlated errors. The board has to remain responsible for its own judgment.
Boards already have the structures, and still fail. This is the strongest objection. Silicon Valley Bank’s risk committee met 18 times in 2022. Carillion had an audit committee, an internal auditor and a Big Four external auditor, yet MPs found its non-executives had failed to scrutinise or challenge its executives. In both cases, the board was warned but failed to act effectively, and the company failed. Part of the solution is to name the decision required and its urgency, as the triggers above do, and to record how the board responds to contrary evidence, as the decision record does. These measures don’t guarantee action, but they do make inaction visible.
Where to Start
These changes can be tested within existing board responsibilities. At the next meeting:
Clarify accountability for agents. Require named agents and their owners, limits and authorities to be registered and one executive to oversee the lot.
Establish a learning loop for major decisions. Start with one decision. Record its assumptions, expected outcomes and conditions for reconsideration. Commission AI-assisted analysis (independently validated) to test material assumptions against source data and bring contradictions back to the board. If possible, review earlier decisions using their original papers and forecasts.
Pilot continuous monitoring. Choose one material automated control. Commission an independent assessment of the pilot.
Review appointments. Ask the nominations committee to assess directors’ understanding and availability to ensure that appointed directors are suitable for the future board model.
Barings failed as the person taking the risks was simultaneously overseeing the controls. AI makes that weakness easier to reproduce at scale. It also gives boards a way to examine the business directly, track their assumptions and test their own judgment. As companies delegate more work to agents that can learn and change quickly, boards will need to enlist AI to support these activities between meetings. A better-written board pack will not suffice.
Questions for the Board
Which assumptions behind our strategy and largest investments could AI invalidate?
Could we reconstruct the reasoning behind our biggest decisions, and learn from that?
Who can stop an agent from making decisions or commitments beyond its authority?
Do our directors have the time and understanding this role requires?
Sources & Notes
Barings. Board of Banking Supervision report, 18 July 1995: trading and settlement unseparated, audit finding ignored, bank sold to ING for £1. The lesson is about evidence and acting on warnings, not that AI would have prevented the collapse. Report
Board meetings. Spencer Stuart, UK Board Index 2025: average 8.2 meetings (7.5 scheduled) across the top 150 FTSE companies. Index
Teleperformance and Klarna. Klarna release, 27 February 2024. Klarna Reuters, 28 February 2024: shares down 19% mid-afternoon, after an intraday low of −29.3%. Reuters.
Self-improving agents. Zhang et al., “Agentic Context Engineering”, arXiv 2510.04618 (ICLR 2026). Shows the technique, not how widely it is used. arXiv
Post Office. Post Office Horizon IT Inquiry, document UKGI00019313: letter from the BEIS permanent secretary to the Post Office chair, 7 October 2020, confirming the 2016 Swift review’s recommendations were not shared with the rest of the board. Inquiry
Judgment. Parra-Moyano, Reinmoeller and Schmedders, HBR, July 2025. A share-price forecast, not a board decision. HBR
Decision records. FRC Corporate Governance Code Guidance expects minutes to record the reasons for decisions. FRC
Time commitment. Deloitte, Non-executive director fees: time for a re-set?, November 2023; based on polling. The same report found about 30 per cent of companies likely to consider a fee re-set. Deloitte
Provision 29. UK Corporate Governance Code 2024; applies to periods beginning on or after 1 January 2026. Code; Mythbuster
Duties and disclosure. Companies Act 2006, s174. Legislation ATG Capital Opportunities Fund LP v Lane, C.A. No. 2026-0447-LWW (Del. Ch. 28 August 2026): an illustration of evidential risk, not UK law. Opinion
Silicon Valley Bank. 18 risk committee meetings in 2022, per SVB’s 2023 proxy. Forbes Supervisory and governance failings: Federal Reserve review, April 2023. Fed
Carillion. Business, Energy and Industrial Strategy and Work and Pensions Committees, Carillion, May 2018: the board held responsible for the failure; non-executives failed to scrutinise or challenge executives; Deloitte as internal auditor and KPMG as external auditor both criticised. Parliament