A director asks a routine question in a board meeting: where did this churn number come from? Normally someone points to the source tab, or names the analyst who pulled it. This time, nobody in the room can answer — the slide was generated the night before, the underlying query is gone, and the person who "built" the deck just describes it as pasting a prompt into a tool. The number turns out to be close enough to plausible that it took a second meeting to catch. That's the moment, multiplied across enough companies, that turned AI-generated slides from a productivity story into a governance problem.

It's not that boards suddenly distrust AI. It's that a board deck carries a kind of weight no other document in a company does, and that weight doesn't tolerate a source nobody can point to.

The problem was never the AI — it was the missing footnote

Every other document that reaches a board goes through a chain someone can reconstruct: a finance analyst pulled the figure, a controller reviewed it, a VP signed off on the framing. That chain is what makes a bad number someone's mistake instead of a mystery. Generative tools don't preserve that chain by default — the output looks the same whether the model pulled from real company data, stitched together something statistically plausible, or quietly filled a gap it never flagged as a gap.

NIST's 2024 Generative AI Risk Profile gave this failure mode a formal name: confabulation — output that's wrong but stated with the same confidence as output that's right. That's the part that makes it dangerous on a board slide specifically. A hallucinated customer testimonial is embarrassing. A hallucinated revenue trend that looks internally consistent, formatted like every other chart in the deck, can influence a real capital-allocation decision before anyone thinks to check it.

A board deck has to survive being read years later, not just presented once

This is where the comparison to other formal company documents actually helps. Anyone who has worked on a shareholder-facing annual report knows the document has to serve two very different readers at once — the person skimming it in two minutes and the person doing real diligence on every footnote. A board deck has a third obligation neither of those documents fully shares: in the UK, board and committee minutes have to be retained for at least ten years under section 248 of the Companies Act 2006, and any AI-generated summary or transcript tied to that meeting risks being treated as part of that record. A slide nobody can trace back to a source doesn't just look sloppy in the moment — it becomes a liability sitting in a file for a decade.

That's a very different bar than "did the chart look professional." It's "can this survive being pulled out and scrutinized in a dispute five years from now" — and an AI-generated number with no audit trail generally can't.

There's a candor problem too, and it's a little counterintuitive

You'd expect the resistance to be purely about accuracy. It isn't entirely. A 2024 poll by GC100 — the body representing general counsel across the FTSE 100 and 250 — found that many directors are uncomfortable with AI tools recording or summarizing board discussions, separate from any concern about factual errors. The worry is psychological: knowing every word might be transcribed and analyzed changes how openly people argue in the room. Boards exist partly to have the blunt disagreement that doesn't happen anywhere else in the company. A tool quietly logging that discussion — even accurately — can flatten it.

So the restriction isn't only "the AI got something wrong." Sometimes it's "the AI's presence changed what people were willing to say," which is a harder problem to fix with a better prompt.

This isn't the first time companies have pulled back on a tool this fast

The pattern has a precedent, just in a different corner of the company. In 2023, Samsung banned generative AI tools on company devices after engineers in its semiconductor division leaked proprietary source code and internal meeting notes to a public chatbot in three separate incidents inside a three-week window. The ban wasn't a rejection of AI as a category — it was a reaction to a specific, traceable failure mode that leadership decided outweighed the convenience. Board-level restrictions on AI-generated slides are following a similar shape: not a blanket "no AI" policy, but a narrower rule aimed at the exact point where an untraceable output could do real damage.

Where this tends to complicate itself is that the fix people reach for first — "just have someone double-check the AI's numbers" — assumes the checker has time and the original source data still exists to check against. In practice, by the time a deck reaches the board, the person doing final review is often optimizing for narrative flow and formatting, not re-deriving every figure from scratch. The verification step people assume is happening is frequently the first thing skipped under deadline pressure.

What's actually getting restricted, and what isn't

Almost nobody is banning AI from presentation work entirely — that would be a strange fight to pick when the tools genuinely save time on formatting, first drafts, and design iteration. What's narrowing is a much more specific thing: AI-generated numbers and claims reaching a board slide without a human-verifiable source attached. Design assistance, layout suggestions, and language cleanup mostly stay untouched. It's the data layer that's under new scrutiny.

That distinction maps almost exactly onto something consulting decks have enforced for decades, long before generative AI existed. The discipline behind McKinsey-style slide construction insists that every title state a checkable claim and every chart carry a source line — not as a formatting rule, but because an unsourced number is treated as an unfinished analysis, not a real one. A generative tool that can't cite where a figure came from fails that test the same way a junior analyst's unsourced chart would have failed it in 1995. The tool changed; the standard a board actually applies didn't.

The templates underneath matter more than people expect

One thing I'd flag for anyone rebuilding a board deck process right now: the chart format itself can either support or undermine traceability. A structured chart layout with a fixed source-line field makes it awkward to skip attribution — the empty field is visible. A chart pasted straight out of an AI tool's output usually has nowhere built in for that line, which is part of why it's easy for the habit to slip. The policy fix people reach for is procedural — "review AI outputs before they reach the board" — but the layout itself is doing quiet enforcement work that a policy document alone doesn't.

None of this means AI is leaving the boardroom — it's being pushed back to where it's actually useful, drafting and formatting, while the numbers that drive real decisions get walked back to a source somebody can name. The harder question for most companies isn't whether to allow AI near a board deck. It's whether their current review process would have caught the slide with the untraceable number before it reached the room, or whether it just happened to get caught this time.