The Director Who Reads the Summary

The Director Who Reads the Summary

0:00

/658.512

1×

In 2007 the directors of Centro approved accounts that classified some $1.5 billion of short-term liabilities as non-current, and that failed to disclose guarantees of about US$1.75 billion given after the balance date (Australian Securities and Investments Commission v Healey 2011, para 9). They had a chief financial officer, a finance team, and PricewaterhouseCoopers as auditors. The Federal Court still found that they had failed in their duty of care, because the one task the law would not let them hand to anyone else was the reading. Nineteen years later, a survey of 531 governance professionals by OnBoard, a vendor of board software, reports that 92 per cent of directors now use AI for board work, up from 69 per cent a year earlier, while 63 per cent of boards have no formal policy for that use and only 6 per cent enforce one (Board Effectiveness Survey 2026, 3). An obvious use of AI is the one the court ruled out—trusting something other than a director’s own reading and judgement.(Burke [1790] 1993)

Simply put, an AI summary is the newest proxy for a director’s judgement, Australian law has already explained why such an approach fails, and that the remedy is an order of governance rather than abstaining from AI use.

Reading Is the One Task a Director Cannot Delegate

The court’s reasoning leaves little room for a digest. Financial statements, Justice John Middleton held, must be “read, understood and focussed upon” by the directors before they adopt them (Australian Securities and Investments Commission v Healey 2011, para 15), and the Act places responsibility for approving them on directors “in a way that they cannot delegate” (Australian Securities and Investments Commission v Healey 2011, para 125). Directors may hand the preparation of the accounts to others. What the Act does not contemplate is that they “simply put the discharge of those functions in the hands of apparently competent and reliable persons, for directors are a part of the process themselves by undertaking the task of approving and adopting” (Australian Securities and Investments Commission v Healey 2011, para 142). Borrowing an American judge’s line, Middleton added that a director “is not an ornament, but an essential component of corporate governance” (Australian Securities and Investments Commission v Healey 2011, para 19).

The statutory protection for reliance covers people, not models. The Corporations Act presumes a director’s reliance reasonable when it is placed in good faith, after their own independent assessment, on an employee they believe to be reliable and competent, a professional adviser or expert acting within their competence, a fellow director or officer within their authority, or a committee of the board on which they do not sit (Corporations Act 2001 (Cth) 2001, s 189). A language model is not—and given its inability to assume responsibility can never be—on that list. How a court would treat a director’s reliance on a digest they commissioned privately is untested, and it is a question for lawyers rather than essayists to settle. However, the list of approved human agents on which reliance can reasonably be placed nonetheless shows the kind of trust the statute had in mind: on someone who can be questioned, whose errors can be attributed, and whose competence the director has had occasion to judge. A chatbot meets none of those conditions, and I would argue the presumption that protects reliance on people does not extend to AI agents.

Directors are trustees, which is why the distinction between trustee and delegate still describes the boardroom better than any governance code. A director who outsources the reading keeps the industry—in so far as they arrive very well prepared at the proxy level—but they surrender their judgement.

Summaries Fail Where the Bad News Hides

Machine summaries fail in two ways that a colleague’s summary need not. Executive summaries are as old as the role of an executive, and a good one is good whoever writes it; but a summary written by a model differs from one written by an experienced professional in where its errors fall and in what it does to the person who reads it.

The errors fall where board packs keep their bad news. Trouble is rarely on page one—if it’s written down at all. It sits in a note to the debt schedule, a changed accounting estimate, a covenant mentioned once in an appendix—precisely the material a summariser treats as needless detail. Liu and colleagues (2024) found that language models make markedly better use of information at the beginning and end of a long document than of information in its middle, and Maynez and colleagues (2020) found that fluent abstractive summaries routinely contain content the source does not support. The models tested in both studies have since been superseded, and the effects have narrowed, yet the principle survives the improvement in model function: a summariser’s attention is unevenly spread, its errors are likeliest where a careful reader’s attention needs to be the most acute, and responses arrive in prose that gives no sign of doubt or uncertainty.

The summary also weakens the reader. Human-factors research has documented automation complacency for decades: when an automated aid is usually right, people monitor it less and miss its failures more (Parasuraman and Manzey 2010). Lisanne Bainbridge’s (1983) irony applies with uncomfortable precision: automating the routine part of a task leaves the human responsible for the exceptions while eroding the practice needed to notice them. Reading the pack is not an inefficient way of extracting its contents. It is how a director builds the sense of the company that tells them whether an update passes the sniff test or if the balance sheet doesn’t balance and is just sheet. An agent cannot answer for what it has done, and a director who has let one read for them has foregone the very context needed to answer for themselves.

A digest is therefore a proxy for preparation, not preparation itself. It produces a director who can speak fluently to whatever the model judged important, and a minute that records a well-informed board. Fourteen months ago I described senior people using AI to surface the salient points for their next board meeting as a kind of executive attention deficit. I treated it then as a failure of reading culture. For a director it is also a failure of fiduciary duty.

Using AI well means questioning it after the reading, not before it—much less in place of it. Ken Sandy, in our conversation on the podcast, made the case that directors should be able to show a credible use of AI in their own work—interrogating the papers they are sent and surfacing the questions they ought to be pressing. He is right, but there is an important distinction here that matters. A director who has read the pack and then asks a model what they have missed is using the tool to stress test a judgement already formed. A director who asks the model to read the pack for them has not saved time in their judgement; they have subjugated their judgement to a machine that carries no responsibility.

Volume is a reason to demand shorter and better prepared materials, not private digests. Board packs have grown, and machine-written management papers are making them longer without making them more informative. Middleton anticipated the complaint fifteen years ago: “A board can control the information it receives. If there was an information overload, it could have been prevented” (Australian Securities and Investments Commission v Healey 2011, para 229). A digest of too much treats the symptom in a way that hides the disease from the one body charged with curing it.

Directors Should Read First and Ask the Machine Second

The remedy is an order of operations, and it can be in place before the next meeting:

1. Read the pack yourself and form a view of what matters before any tool offers one.
2. Use the model afterwards to ask what the pack omits, contradicts, or buries—not to explain what it says.
3. Check every “fact” the model gives you against the page from which it came. No source, no fact.
4. Ask whether the board has a policy on directors’ own use of AI; if it has none, that is the agenda item.
5. If the pack is too long to read properly, tell the chair and the company secretary, and ask management to take the time to prepare more effectively.

The fourth step is the one boards most often skip, because the question is usually asked of everyone except the board. Nick Hassett, a guest on the show, puts it first among the four questions he and his colleagues at Capability argue a board should ask about AI: “What AI is already in use — sanctioned and unsanctioned?”. Asked honestly, the answer includes the pack on a director’s laptop on a Sunday evening. His test for the answer, that it be “evidence, not reassurance”, is the test a digest fails.

Uploading the pack also raises a second delegation problem, which is confidentiality. The terms of a consumer chatbot are not the board’s confidentiality undertaking, and a director who has not read them has delegated more than their reading.

The risk extends to privilege. Board packs routinely carry legal advice, and in Australia that advice stays privileged only while the client does nothing inconsistent with keeping it confidential (Mann v Carnell 1999, paras 28–29). No Australian court has yet decided whether uploading advice to a chatbot is such an act, but the Federal Court now warns everyone appearing before it that information entered into a generally accessible AI tool “may become available to other people”, and lists privileged material among the information whose disclosure the law restricts (Use of Generative Artificial Intelligence (GPN-AI) 2026, paras 4.13–4.14). The Court allows that the risk may be lower for closed tools; most directors’ chatbots are not closed. The first court to rule on a related question offered little comfort. In February 2026 Jed Rakoff, sitting in the Southern District of New York, held that documents a defendant had prepared with a consumer chatbot were protected by neither privilege nor work product, in part because the provider’s terms allowed it to retain what he typed and disclose it to others (United States v Heppner 2026). A director who uploads the pack may be putting at risk not only the company’s confidences but its privilege, which belongs to the company and not to them.

The test for a board is whether each director could still tell, unaided, that something in the pack was wrong. Boards are already expected to test whether the organisation has the capability it claims; this asks them to test their own. The Centro directors did not lack advice, and no one doubted their honesty; Middleton called them “intelligent, experienced and conscientious people” (Australian Securities and Investments Commission v Healey 2011, para 8). They had a finance team and auditors, and they relied on both, as the law allows. What failed was the part of the office that no one else could perform. Nineteen years on, the advice is cheaper, faster, and far more fluent, yet the ability of directors remains largely where it was.

Good night, and good luck.

References

The Card Sharp with the Ace of Diamonds by Georges de La Tour (1593–1652) is licensed under Public Domain.