AI notetakers in meetings: what to switch on, and what to switch off
The AI notetaker arrived in most small businesses without anyone deciding to adopt it. It is genuinely useful and it quietly creates records you may not want. Both things are true.
Nobody in your business decided to start using an AI notetaker. Microsoft, Google and Zoom added them, put a button in the corner of the meeting window, and somebody pressed it. Within a month it was normal.
This is the most widely adopted AI tool in British small business and almost none of the adoption was deliberate. Which means the useful parts and the awkward parts both arrived unexamined.
What they are genuinely good at
Actions and owners. The single best use. A summary that ends with who agreed to do what, by when, is worth more than a full transcript nobody reads. Most tools do this well now.
Meetings you could not attend. A twenty-minute read of a summary instead of an hour in a meeting is a real saving for an owner pulled in several directions. Used properly this reduces the number of people who need to sit in.
Site visits and client calls. For surveyors, consultants and anyone who takes notes while also doing something else, having the record captured means you can pay attention to the person instead of the notepad.
Handover. When someone leaves or is off sick, meeting summaries turn out to be the only written record of half the decisions a small business makes.
Where it goes wrong
People stop saying things. The most underrated cost. If a difficult conversation about a client, a supplier or a member of staff is being transcribed, it either does not happen or it happens afterwards in the car park. That is worse than a meeting with no notes.
You have created a disclosable record. A transcript is a business record. In a dispute, a grievance or a tribunal, it can be requested. Most owners have not thought about the fact that the AI is now writing down the bit where they were annoyed.
The summary is confidently wrong sometimes. Names get mixed up, and the tool occasionally attributes a decision to the wrong person or records an action nobody agreed to. If the summary is the only record, that error becomes the truth. When AI hallucinates covers why this happens.
Welsh and strong local accents. Transcription quality drops on Welsh-language content and on some North Wales accents. It is better than it was and still not good enough to trust unread. The Welsh language post goes into this properly.
Consent. Recording a meeting where an external party is present without telling them is at best rude and at worst a problem. The tools announce it in the meeting window, which is not the same as anyone noticing.
The settings worth checking
Whichever platform you are on, these are the five to look at.
- Who can start recording or summarising. Default is often anyone in the meeting.
- Whether transcripts are retained, and for how long. Indefinite retention is a common default and rarely what you want.
- Whether the summary is emailed automatically to all attendees, including external ones.
- Whether meeting content can be used to train the provider's models. On business tiers this is usually off; confirm rather than assume.
- Whether one-to-one meetings are covered by the same rules as team meetings. They usually should not be.
A workable house rule
The businesses that handle this well tend to land in roughly the same place, and it is simpler than a policy document.
Notetaker on for project meetings, client calls where everyone knows, supplier discussions, training and anything where the actions matter more than the atmosphere.
Notetaker off for one-to-ones, anything involving a person's performance or conduct, early commercial conversations, and any meeting where somebody needs to be able to think out loud and be wrong.
Always say at the start that it is running, and always have one person read the summary before it is treated as the record. That second habit takes two minutes and catches the confidently wrong bits.
Where this fits everything else
Meeting AI is worth putting in your AI policy, because it is the tool most likely to be in use already. The one-page policy template has a place for it, and the risk register post covers the retention question in more depth.
It is a good tool. It got adopted without a decision, which is the only real problem with it. Spending twenty minutes on the settings and the house rule turns it from a quiet liability into one of the more useful things AI does for a small business.