The recorder in the room

I joined a steering committee call this spring and counted the participant list before anyone spoke. Fourteen humans and three notetaking bots — one from the client's tenant, one riding in with the integrator, and one nobody could account for that had presumably followed a director in from a different meeting series. Three separate organizations were about to hold three separate verbatim records of the same hour, under three different retention policies, and not one person in that room had made a decision about it.

That's the part worth sitting with. This wasn't a policy failure in the dramatic sense. Everyone on the call was well-intentioned, and the tools were doing exactly what they were bought to do. But a program that had a formal change-control process for a database field had no process at all for the fact that its governance forum was now producing a court-quality transcript every two weeks.

A transcript is not minutes

Minutes are a deliberate artifact. Somebody sat in the meeting, decided what mattered, wrote it down, circulated it, and took corrections. That editorial act is not laziness or spin — it is the mechanism by which an organization decides what its official position is. Minutes are short because a record of decisions is more useful than a record of noise.

A machine transcript is the opposite object. It captures the half-formed thought, the sarcastic aside, the sentence someone abandoned mid-clause after realizing they had the wrong number in front of them. It attributes with reasonable but imperfect accuracy. And the AI summary generated on top of it makes a third artifact, which paraphrases, compresses, and occasionally asserts a conclusion the meeting never actually reached.

So one meeting now yields four candidate versions of what happened: the recording, the transcript, the AI summary, and the approved minutes. White & Case's litigation group put the problem plainly in a client alert on AI meeting tools — organizations are no longer dealing with a single minutes document but with "a complex ecosystem of recordings, machine transcripts, AI summaries, decks and conventional minutes," and when those versions diverge, an adverse party or a regulator will reach for the most detailed one, not the most official one. On a program, the same dynamic plays out without anyone filing suit. Six months on, two workstreams cite two different versions of the same steering committee, and now you're litigating your own history instead of delivering.

What it costs in the room

The second effect is harder to measure and matters more. People talk differently when they know the tape is rolling.

The most valuable sentence in a program review is usually a hedge. I'm not comfortable with the integration date. We're calling it green but I want to flag something. That sentence is how bad news enters a governance forum early enough to be cheap, and it gets said by someone who is not yet certain, who is testing whether the room agrees, and who would say it very differently in a written statement. Verbatim capture converts every one of those into an on-the-record position. The lawyers noticed this too — the same alert lists the "chilling of board and strategic discussions" as a governance risk in its own right, because moving from selective summaries to verbatim records changes how people engage.

I've watched it happen at the working level. The concern doesn't disappear; it relocates. It moves to the side channel, the hallway, the direct message after the call — which means it is now outside your governance record entirely, arriving weeks later and considerably more expensive. You optimized for a complete record of the meeting and got a less complete picture of the program. That's the trade nobody priced.

The obligations you inherited

Then there's the boring, expensive part. If AI-generated outputs form part of the business process, they may be deemed records for retention and legal-hold purposes — which means transcripts and summaries have to be inside your retention schedule and your hold process, not sitting in whatever folder a third-party vendor defaults to. Mayer Brown's global survey of the issue makes the same point about notice and consent across jurisdictions: requirements "range from one-party to all-party consent standards," and voice capture can implicate biometric rules where the tool builds speaker-identification profiles. A bot visible in the participant list is not, on its own, consent.

None of that is a reason to ban the tools, and I want to be direct about why. I've argued before that prohibition just converts a visible practice into shadow AI, and the notetaker is the purest example on the market — it takes one click from a personal account and produces something genuinely useful. Ban it and you don't get fewer transcripts. You get transcripts you can't see, on infrastructure you don't control.

Govern the capture, not the tool

In the Army we didn't record everything, but we were rigorous about what constituted the record. An operations order was the record. Hallway conversation was not, and everyone knew which was which, because ambiguity about the authoritative version is how units end up executing two different plans. The discipline wasn't about secrecy. It was about designating one version and making it unmistakable.

That's the whole fix, translated. Classify your meeting types and set the rule per type: working sessions capture freely, steering committees capture with an approved tool inside your tenant, and anything touching legal exposure, personnel, or vendor negotiation runs on human minutes only. Name one authoritative artifact for every governed forum and label the rest as working material — an unlabeled AI summary circulating in email will become the de facto record purely by being the easiest thing to forward. Put transcripts and summaries on the retention schedule you already have. And announce capture verbally at the top, every time, including for the people who join at minute eleven. It takes four seconds and it's the difference between a practice and an incident.

The bottom line

Perfect recall is not the same as good governance, and the gap between them is judgment — someone deciding what mattered, writing it down, and putting their name on it. The notetaker is a real productivity gain and I use one. But it produced a new class of program record without a single governance decision behind it, and unowned records have a way of becoming someone's problem at exactly the wrong moment. Decide what your record is. Otherwise the tooling decides for you, and it will pick the longest version.