Two-thirds of professionals now have an AI notetaker in their meetings, and 84% say they speak differently because of it. Almost none of it was ever approved.
Ask an Irish leadership team which AI system has the widest deployment in their organisation and you will usually hear Copilot, or a customer chatbot, or something in the finance stack. The answer is the meeting notetaker. It is in more meetings, touching more sensitive conversation, held by more people, than anything that went through a procurement process. And in most of the organisations we look at, no one can tell you who approved it, where the transcripts live, or how long they are kept.
The adoption numbers from 2026 survey work are not marginal. Roughly two-thirds of professionals now use an AI notetaker at work, up from about a quarter in 2023. Two-thirds of Fortune 500 companies have one running somewhere. The curve is inverted by size: small firms sit near 80% adoption, organisations above 5,000 people sit closer to 43%, because the big ones eventually blocked it. The Irish mid-market is squarely in the high-adoption band and has done none of the blocking.
It arrived the same way Copilot did
Nobody made a decision here. Teams and Zoom shipped recap and transcription as a default-on feature. Individual staff signed up for Fireflies, Otter or Read on a free tier and pointed it at their work calendar. External participants started sending bots into meetings they were invited to, so a supplier's notetaker is now sitting in your commercial negotiation, recording your side of it, on terms your legal team has never seen.
That last one is worth sitting with. An organisation can have a sound internal policy and still have every meeting transcribed by someone else's tool, because a bot joined from the other party's invite and nobody felt able to ask it to leave.
A transcript is a record, and minutes were never meant to be verbatim
Minutes are a curated instrument. Somebody decides what is material, what was decided, and what was an aside. That curation is a governance control, and it has been quietly removed.
A transcript captures the half-formed opinion, the frustrated aside about a customer, the bit where someone says the number is probably wrong. Courts in 2026 have been unambiguous that AI-generated transcripts and summaries are electronically stored information, discoverable on the same footing as email, and subject to litigation hold from the moment proceedings are reasonably anticipated. An HR grievance, a supplier dispute, a regulatory inspection — each of them now reaches a body of unreviewed verbatim material that used to be a two-page minute.
The privilege exposure is sharper again. When counsel is in the meeting and a third-party transcription service is processing the audio, the confidentiality that privilege depends on is being handed to a processor whose terms may not recognise it at all. That is a waiver risk that cannot be undone after the fact.
Then there is accuracy. These tools misattribute speakers, and their summaries state things as decided that were floated and dropped. An AI summary that says "Finance agreed to fund the pilot" carries the weight of a record, gets circulated, and nobody goes back to the audio to check.
Article 50 raised the floor in August
From 2 August 2026 the EU AI Act's transparency obligations apply. People have to be told when they are interacting with an AI system, and deployers carry duties around AI-generated content and around emotion recognition where it is in play — and several of the meeting tools now market sentiment and engagement scoring as a feature. Layered on top of that is the ordinary GDPR position: a transcript of a meeting is personal data about everyone in it, needing a lawful basis, a retention period, and an answer when someone asks for a copy.
Most organisations we assess cannot answer the retention question. The default on the consumer tiers is indefinite.
The productivity cost is the part leaders miss
84% of people who work with an AI notetaker present say they change how they speak. That is the finding that should reach a leadership team, not the legal exposure.
The whole value of a working meeting is that people will say the uncertain thing out loud. Put a permanent record in the room and the conversation gets more careful, more performative, and slower to reach the disagreement that the meeting existed to resolve. Organisations adopt this tooling to cut the cost of meetings and end up needing a second, smaller, unrecorded meeting afterwards to have the actual conversation. The diary gets longer.
Four decisions, and they take an afternoon
Decide which meetings can be transcribed and which cannot — board and sub-committee sessions, HR matters, and anything with counsel present should be off by default. Decide the retention period and configure it, rather than leaving the vendor default in place. Decide who is permitted to admit an external notetaker and give people explicit permission to refuse one. Decide who reviews an AI summary before it circulates as the record of what was agreed.
None of that requires a project. It requires someone senior to own it, which is the part that has not happened.
We run this as part of the Cognitive Mirror diagnostic, because meeting data and meeting tooling are the same problem viewed from two sides. If your organisation has notetakers running and no one can say where the transcripts are, contact Acuity AI Advisory for a diagnostic conversation.