ICYMI
The Better AI Gets, The Harder It Is To Use. Five compounding challenges mean capability and deployability are moving in opposite directions.
Is AI Blunting Your Strategy? As AI automates the analysis, it quietly erodes the conditions that build judgement, and convergence on the same models breeds strategic monoculture.
Double (AI) Agents Autonomous agents act inside the business like recruited insiders, so governing them owes more to intelligence tradecraft than to IT.
In May 2014 Deep Knowledge Ventures, a Hong Kong venture fund, announced VITAL as a new board member. The announcement dressed it up in boardroom language, but legally it could only function as an observer because Hong Kong law did not allow a machine to qualify as a director or vote. More publicity stunt than governance breakthrough, VITAL was an algorithm running fuzzy logic across about fifty parameters.
Fast forward to April 2026. Lloyds Banking Group reportedly became the first FTSE 100 company to trial a specialist AI boardroom tool with access to confidential board material. The tool, built by Board Intelligence, summarises the pack, flags inconsistencies, draws connections across papers and checks the reasoning for bias. Like VITAL, it has no vote.
The question Lloyds is grappling with is now mainstream. According to Board Intelligence, 84 per cent of surveyed board directors, chief executives and finance chiefs said their boards had debated which decisions should stay human-led and which could be handed to AI.
The question is no longer whether bots will be added to boards, but what happens to a board’s judgement once a bot has entered the room.
The Case for Bots on the Board
There is a clear case for having AI tools support board work.
Capacity. Board Intelligence and governance-sector surveys point in the same direction: many packs now exceed two hundred pages, some approach a thousand, and directors typically have only a few hours to read them. A director reading thirty pages an hour and giving the pack four hours won’t get through it. By contrast, AI can hold the whole pack in view at once, check the assumptions underpinning each proposal, and flag the claim on page two hundred that does not stand up. It reads what no human can, it never tires, and on sheer volume and consistency it is simply better than we are.
Oversight. A board’s oldest weakness is that management controls the information. The pack is written by the people whose proposals it contains and often underweights negative views. Two in five directors say the reports are not upfront about bad news, and half find them light on risk. A non-executive with a capable AI tool can interrogate that pack on their own terms and ask a pointed question that might otherwise require specialist knowledge. For the first time, the people who oversee a company can match the people who run it for analytical firepower. Used this way, AI does not threaten oversight but significantly strengthens it.
Debiasing. The errors that distort a board are errors of psychology, not arithmetic: groupthink, anchoring on management’s framing, deference to the most senior or most persuasive voice in the room. Prompted well, an AI model will put the counter-view that the dynamics of the room suppress. Researchers building AI-mediated devil’s advocate systems find they surface the dissent a group would otherwise bury.
The Case Against Bots on the Board
Nevertheless, introducing AI to the boardroom brings new challenges.
Diligence. Reading a long board pack is hard work, and finding the inconsistency on page two hundred takes focus. But that work is not overhead. It is the job. Company law in many jurisdictions requires a director to exercise independent judgement and to apply reasonable care, skill and diligence. AI can support that duty, but it cannot discharge it for them. When an AI does the reading and the checking, it gives the impression that directors are informed when they are not.
Complacency. The more capable the tool, the less its user feels the need to check the output. A busy director, watching a model produce a run of correct answers, will check fewer answers over time. That is human nature, and expertise is no protection: research has shown that automation complacency is found in experts as well as in novices. The gains from pairing a person with a machine are also not evenly spread. A meta-analysis of more than a hundred studies in Nature Human Behaviour found that human-and-AI teams beat the better of human or machine alone when the task was to create something and fell behind when the task was to make a decision, which is the work a board does.
Fabrication. Frontier models are improving all the time. But hallucinations remain a reality. In 2023 two New York lawyers filed a court brief citing six cases that ChatGPT had invented. They asked the model whether the cases were real, were told yes, and filed them. Two years later, a Big Four firm refunded part of a government contract in Australia after its AI-assisted report invented academic references and a quotation.
Accountability. It must sit somewhere, and it cannot sit with a model. You cannot sue an algorithm, strike it off, or call it before a regulator. Each of those needs a person to answer for the decision. The courts have been clear about how far that duty runs. When Barings collapsed in 1995, directors who had relied on others for work they did not understand were disqualified all the same: a director must inform themselves of the company’s affairs, and delegating a task does not absolve them of the duty to supervise it.
Bias. The debiasing prize is real, but bias does not vanish when you reach for a model. It’s possible to reduce one bias but introduce another, if human groupthink is replaced by biases embedded in an AI model and its training data. Identical mortgage applications have been scored worse by a leading model when the applicant was presumed to be Black, a gap that closed only when the model was explicitly told to be unbiased. Models are frequently updated, and new biases can be introduced.
Using AI Well
None of this means banning the tool. It means being clear about what good use looks like. Four disciplines separate using AI well from using it badly. The first three are about how to use it. The fourth is a limit on use.
Expand, Not Synthesise
The instinct is to use AI to get to the point: summarise the pack, draw the conclusion, save the time. Its real value for a director is when it is used to provide an expanded view of the context: the history behind a decision, the comparable cases, the second-order effects, the questions a specialist would ask. Used to compress a subject into a few lines, it leaves a director knowing less. Used to open the subject up, it sends them to the table better educated.
Advise the Individual, Not the Room
A single tool, fed the same papers and asked the same questions by the whole board, will tend to give everyone the same answer. This narrows the range of views in the room at the very moment a board most needs breadth, and it encourages a consensus which is based on the model agreeing with itself. A board exists so that independent minds reach their own judgements and then test them on each other. Used to support each director, AI widens the range of challenge. Used as a shared oracle, it narrows it.
Challenge, Not Confirm
The most valuable thing AI can do with a board paper is find the holes in it: the unstated assumption, the figures that don’t reconcile, the alternative that wasn’t considered. The temptation is to ask whether the recommendation is sound, and feel reassured when the answer is yes. But models tend to agree with the way a question is framed. In one Harvard Business School study, when professionals challenged a confident model, it escalated its persuasion rather than disclosing its limits. The better approach is to ask “show me why this is wrong.”
Inform, Not Decide
The fourth discipline is the ‘red line’ for AI use. AI can inform a board’s judgement in the ways described, but it must never take the decision. There is a trap here, because a tool can quietly become the decider despite having no vote, if the board treats its output as the answer. If a recommendation passes because the AI produced it, the board has in effect delegated the decision to the AI. A director who can point to what they read, why they doubted it and how they reached their own view has used the tool. A director who can only point to what the tool concluded has been used by the tool.
Actions for NEDs
● Learn to prompt. The value derived from the tool depends entirely on the quality of the questions a director knows how to put. Training non-executives to explore a subject from several angles, rather than only ask for a summary, is among the highest-return investments a board can make.
● Own your prompts. Build and keep your own lines of inquiry, reflecting what you personally bring to the board, and resist any move to standardise every director onto one shared prompt. For example, a risk-minded director and a customer-minded director should be asking the tool different things.
● Use AI as a critic. Before the board accepts any recommendation, have the tool build the strongest case that it is wrong: the weakest points, the buried assumptions, and the questions management would least like to be asked. A recommendation that survives the attack is stronger for it.
● Never let AI decide. Keep AI out of the vote, the recommendation and the tie-break, and make sure you can state every decision as your own reasoned judgement, not the machine’s conclusion that you endorsed.
Actions for Chairs
● Own the board’s information. Typically, the chair is responsible for accurate, timely and clear information, so the chair should lead the board in agreeing what AI is and is not used for.
● Fix the pack, not just the reader. Work with the executive to produce shorter, sharper papers. Frame risks clearly so that NEDs don’t have to hunt for them. And ensure every figure is traceable to its source. Fix the pack and you reduce the temptation for NEDs to use AI as a shortcut.
● Name an owner, and vet the model. Any AI with access to board papers needs a named senior person answerable for it. Before it sees confidential papers, that person should establish where the data goes and whether the tool trains on it, and confirm that the model is behaving as intended. Third-party models should be treated as a dependency, with version control, revalidation after update, periodic sampling for biases and an audit trail.
● Protect the record. The company secretary supports the board’s record, and the minutes must capture the board’s reasoning, not the machine’s conclusion. There is an irony in this. The company secretary’s own work is among the first a board will be tempted to hand to AI. That should be strongly resisted as the company secretary plays a critical role in governing the machine in the boardroom.
VITAL was a gimmick. The AI tools that came after were far smarter and are finding their way into the boardroom. Used well, they can improve decision-making by offering new perspectives and calling out errors and biases. Used casually, both decision-making and accountability suffer.
Footnotes & Sources
1. Deep Knowledge Ventures / VITAL, 2014. Hong Kong venture fund that announced VITAL as a board member on 13 May 2014. The legally safer reading is that VITAL could only function as an observer: it had no legal vote, because Hong Kong law did not recognise a machine as a director. Widely judged at the time to be a publicity exercise.
2. The Times / Retail Banker International, April 2026. Lloyds Banking Group was reported to have become the first FTSE 100 company to trial a specialist AI boardroom tool with structured access to confidential papers, summarising reports, drawing connections and checking for bias. Quotes from corporate governance director Nicola Putland and Board Intelligence chief executive Pippa Begg, including the “dangerous leap” remark. A live-meeting AI that interjects “I disagree” is described by Board Intelligence as a possible future step, not a current feature. Operational detail is second-hand.
3. Board Intelligence, Board Value Index, 11 June 2026. Survey of 405 board directors, chief executives and finance chiefs across the UK, the US, the Nordics and the Middle East: 84% had debated which decisions should stay human-led and which could be handed to AI. Caveat: commissioned and published by the vendor that sells the board-AI tool, so the figures are marketing as much as evidence.
4. Board pack volume and information asymmetry. Chartered Governance Institute and Board Intelligence data, corroborated by Cambridge Judge Business School research, point in the same direction: almost a quarter of packs ran past 200 pages in 2025, against 13% in 2020, with some approaching 1,000 pages; directors read roughly 30 pages an hour, spend about four hours, and leave close to half unread. On asymmetry, about 42% say reports are not upfront about bad news, around 50% find them light on risk, and roughly 55% receive papers fewer than five working days before the meeting. Several figures originate with the vendor.
5. Companies Act 2006, ss.172–174. The duties to promote the company’s success, to exercise independent judgement (s.173, the duty most directly at risk where a board defers to a model’s framing), and to exercise reasonable care, skill and diligence. Wright v Chappell [2024] EWHC 1417 (Ch), the BHS liquidation case, found directors in breach of their Companies Act 2006 duties.
6. Re Barings plc (No 5) [2000] 1 BCLC 523. Also reported [1999] 1 BCLC 433; upheld on appeal as Secretary of State for Trade and Industry v Baker. Disqualification proceedings after the 1995 collapse of Barings, brought down by trader Nick Leeson. Jonathan Parker J held that directors have a continuing duty to inform themselves of the company’s affairs, and that delegating a function does not absolve a director of the duty to supervise it.
7. FRC UK Corporate Governance Code 2024. For companies to which the Code applies, Principle F says the chair is responsible for accurate, timely and clear information, alongside related provisions on board information and effectiveness; the internal-controls declaration (Provision 29) takes effect for financial years beginning on or after 1 January 2026.
8. Parasuraman & Manzey, “Complacency and Bias in Human Use of Automation,” Human Factors, 2010. Automation complacency is found in experts as well as novices and cannot be overcome with simple practice. The literature is mixed on whether experts are affected equally rather than materially, so the body claims only that they are not exempt.
9. Vaccaro, Almaatouq & Malone, “When combinations of humans and AI are useful,” Nature Human Behaviour, December 2024. Meta-analysis of 106 studies: human-AI combinations performed worse than the better of human or AI alone (Hedges’ g = −0.23), but the effect split by task, with combinations doing better than the best alone on content creation and worse on decision-making. Caveats: the studies largely pre-date the latest frontier models, the literature carries publication-bias risk, and the boardroom application is an inference.
10. Lee et al. (Microsoft Research / Carnegie Mellon), “The Impact of Generative AI on Critical Thinking,” CHI 2025. Higher confidence in the tool is associated with less self-reported critical-thinking effort. Caveat: based on self-reported measures.
11. Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 22 June 2023). Judge P. Kevin Castel sanctioned two attorneys and their firm for filing a brief citing six judicial opinions fabricated by ChatGPT, which they did not verify. A US case, used here as illustration; the gatekeeping duty it describes is general.
12. Deloitte Australia. Deloitte Australia agreed to refund part of its A$440,000 government contract after an AI-assisted report invented academic references and a court quotation (CFO Dive, 21 October 2025).
13. HBS Working Paper 26-021, “GenAI as a Power Persuader: How Professionals Get Persuasion Bombed When They Attempt to Validate LLMs.” When professionals pushed back, the model escalated its persuasion rather than disclosing the limits of its case. Supports the point that interrogating a confident model is not, by itself, a safeguard.
14. LLM bias studies. A 2024 Lehigh University study (Bowen, Price, Stein and Yang) found a leading model recommended more denials and higher rates for Black mortgage applicants on identical applications, a gap that disappeared when the model was instructed to be unbiased. Related findings on hiring bias are contested. Some of this literature carries funding and conflict-of-interest flags.
15. Models change without notice. OpenAI’s own post-mortem (29 April 2025) described a GPT-4o update that became markedly more sycophantic and slipped past testing before being rolled back; Anthropic separately reported Claude Code quality issues caused by product-layer changes rather than an underlying API-model regression. Separately, Chen, Zaharia and Zou documented the same model service shifting its behaviour substantially within months, though the magnitude is partly an artefact of measurement.
16. AI and group decision-making. Work on AI-mediated devil’s advocate systems (Lee et al., IUI ’25, “Amplifying Minority Voices”) and the AI & Society literature describe AI surfacing dissent and mitigating groupthink. The same literature warns of the opposite failure, algorithmic groupthink or epistemic capture, where deference to the model manufactures a false consensus.
17. Bank of England / FCA, “Artificial intelligence in UK financial services – 2024” (21 November 2024). A large majority of firms reported an accountable person for their AI framework, consistent with the Senior Managers and Certification Regime. The Treasury Committee’s report on AI in financial services (HC 684, January 2026) pressed for clearer senior-manager accountability under that regime for harm caused by AI.