Should directors put board papers into AI, and rely on the summary?

These are two questions and they have different answers. Whether a director may put a confidential board paper into an AI system is a question about the terms of the tool, and it has a clean answer. Whether a director may rely on an AI summary of that paper is a question about what a summary does to the person reading it, and the evidence answers it differently from the way most directors would like. Both need settling before the next board pack goes out, because directors are already doing both, on their own accounts, and the board has not said anything.

The confidentiality question#

A board paper put into a consumer AI account has been disclosed to a third party under that party's terms, which for consumer products commonly permit retention and may permit use of inputs to improve the service. Whether that breaches the director's duties depends on the jurisdiction and the paper; that it breaches the company's expectation of confidence does not. The answer is not to forbid the practice, which drives it onto personal phones, but for the company to provide a tool on enterprise terms it has accepted, meaning no training on inputs, retention the company controls, and access it can audit, and to name that tool in the board's own rules. A director then has a lawful route and no excuse. The same principle, applied to the whole workforce, is at what to do when people work around the AI policy.

The reliance question#

A summary is useful for three things: orienting in a long pack, finding the section that matters, and drafting the questions to ask. It is not the paper. Steyvers and colleagues measured how well a model's own confidence separated its correct answers from its incorrect ones, and how well its readers' confidence did the same; the model discriminated at an AUC of 0.751 to 0.781, its readers at 0.589. The person reading the output is worse than the machine that wrote it at knowing when to doubt it. Kim and colleagues found that people given a system that was right half the time agreed with it 81 per cent of the time and scored lower than people with no system. Xiong and colleagues found models' stated confidence poorly matched to their accuracy and clustering at round numbers. Put together: a director reading a fluent summary will find it more convincing than its accuracy warrants, will not know which parts to doubt, and will vote on it.

A director's duty attaches to the decision. On any matter the board will vote on, the paper is the thing the director is responsible for having read, and a summary that omitted the qualification in the appendix has not discharged that. The individual version of the argument is at should I let AI summarise everything I read; the board version is stricter, because the director cannot delegate the vote.

A policy in three lines#

The company provides an approved AI tool on enterprise terms and names it in the board's rules; board papers go into that tool and no other. Directors may use it to orient, to find, and to draft questions, and are encouraged to. On any matter the board will vote on, the director reads the paper, and the minutes may record that the board did. The third line is the one that will be resisted and the one that matters, because a board whose members voted on summaries has delegated the vote to a system nobody appointed, which is the board-level version of the disagreement rate falling to zero.

What nobody has measured#

No study observes directors using AI on board papers. The reliance evidence here is on individuals reading model output in experimental settings, on tasks that are not board decisions, with models that have since moved. The confidentiality answer is a reading of standard consumer and enterprise terms, which vary and change; the company's own counsel should check the terms of the tool it names. The three-line policy is a proposal.

Key sources

• Steyvers, M. et al. (2025). What large language models know and what people think they know. Nature Machine Intelligence. Graded entry.
• Kim, S. et al. (2024). I'm Not Sure, But...: Examining the Impact of Large Language Models' Uncertainty Expression. Graded entry.
• Xiong, M. et al. (2024). Can LLMs Express Their Uncertainty? Graded entry.

Related SuperSkills research#

On why the output sounds so sure, why does AI sound so confident and what is calibration. On the general failure, automation bias. On the record a board decision should leave, decision provenance. On how directors should use AI at all, what board oversight of AI looks like.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. He has run, grown, bought and advised businesses with AI in them. Findings are attributed to the studies and statements that produced them and kept separate from the interpretation. This is a living reference, reviewed and updated as significant new evidence appears.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.