The Risk Coalition's blog post of 22 July 2026 on its Risk Matters roundtable puts a plain question to every board paper: "What is fact, forecast or judgement?". It also warns that where the underlying conversation is weak, technology may simply ‘industrialise the weakness’. When a board pack is drafted with AI assistance, that first question seems to me to have a companion: of the statements in this paper, which did a person actually check?
This is no argument against the tools. The same post notes that AI may reduce the time spent compiling reports and increase the time available for analysis. My concern is narrower. A polished summary can make it hard for a reader to see what has been compared with a source and what has not.
Here is an invented example; no real firm, committee or figure is described. A quarterly paper to a risk committee is marked ‘for noting’. An AI assistant drafted it from the operational risk log and a second-line report. It states: "Third-party outages stayed within appetite in the quarter, with four incidents against a tolerance of six". The risk manager reads it for tone and signs the cover sheet. The chair asks one question. The minutes record that the committee noted the report and challenged management on third-party resilience.
In the log there are indeed four incidents above the reporting threshold. There are also three more, reclassified during the quarter and still pending review. If they were counted, the total would be seven against a tolerance of six. The draft mentions neither the threshold nor the three. The example does not require anyone to have meant to mislead, and the committee might have reached the same view with the extra information. But a later reader, whether an auditor, a new director or a supervisor, now finds a paper, a signature and a minute, and each appears to vouch for the others. None of them shows that anyone compared the sentence with the log.
In this example, one line could have said so: which statement was checked, against which source, by whom, and what was left open. "Four incidents above the threshold matched to the log, 3 March, risk manager; three reclassified incidents and the appetite statement not reviewed." Such a line could make the open point visible to a director before the meeting, provided it reflects a real check.
The roundtable's five gateway questions already ask what decision is required and who owns execution. I do not think this needs a sixth. It sits close to the distinction the post began with. In the example, the count is a claim about the past, while "within appetite" is a judgement that depends on how the open items are treated. On the page, the two read as one settled sentence, and nothing tells the reader which part anyone looked at.
Two cautions belong here. A check line can become its own ritual. If it reads the same way every quarter, it moves the problem to a new place without solving it. And it says nothing about how reliably people notice errors: a reviewer can trace a figure to its source and still miss what the source leaves out. I am not aware of good evidence on how often that happens in board papers, and I would be wary of a precise rate.
Whether such a line belongs on a cover sheet, in the minutes or nowhere at all will depend on the organisation, its sector and its company secretary. I offer it as an observation built on an invented case, not as guidance, and it does not replace a board's own governance arrangements or the Risk Coalition's published practice guidance.
The question I would be curious to hear at a future roundtable is a modest one. When a director says "I saw that in the paper", what does the director know about what was looked at, and by whom?
Andreas Ehstand is an independent AI researcher based in Starnberg, Germany. One of his research projects, on structured human-AI interaction data, has been certified as research and development in Germany by the BSFZ, the body that assesses projects for the research allowance. He studies how reliably people notice errors in AI output and how the reasons behind AI-assisted decisions can be recorded.