AI Meeting Notes and Transcription at Work: Consent, Accuracy and Records

AI notetakers arrived in most organisations without anyone deciding to adopt them. Someone enabled a feature, someone else saw the summary and liked it, and within a few months a bot is joining meetings that include client matters, personnel issues and commercially sensitive discussion.

The tool is genuinely useful. The problems are not technical, and they are rarely considered before the first recording exists.

Consent: who has to be told

Recording a conversation and transcribing it are both processing of personal data. In an internal meeting that is usually manageable under your existing arrangements, but it still has to be a decision rather than a default.

The practical minimum:

• Tell people at the start that the meeting is being recorded and transcribed, who can see the output, and how long it is kept. A bot name appearing in the participant list is not notification.
• Ask before external meetings. Clients, candidates, suppliers and patients have not agreed to anything. Ask in the invitation, not thirty seconds in.
• Allow it to be declined. If someone objects, the meeting proceeds without it. This matters most in the conversations where a transcript would be most sensitive.
• Be careful with recruitment, grievance, disciplinary and performance meetings. These have their own procedural rules and an automated transcript can cut across them.

What the transcript becomes once it exists

This is the part most organisations miss. A recording is a record, and records are discoverable. A verbatim transcript of a meeting where someone speculated loosely about a client, a colleague or a liability is a document that can later be requested — in litigation, in a regulatory inquiry, or by the individual themselves under a subject access request.

Three consequences follow:

• Set a retention period and apply it. Indefinite retention by default is the common state and the worst one.
• Decide who can access the transcript, not just the summary. They are different documents with different risk.
• Be clear whether the transcript or the minutes are the record. If both exist and they differ, you have a problem.

People also speak differently when they know they are on the record. That is sometimes the point, and sometimes it removes exactly the candour the meeting needed.

The accuracy problem

Transcription is strong and summarisation is weaker, and the second is what people actually read.

Summaries routinely attribute an action to the wrong person, state as a decision something that was raised and dropped, miss the qualification attached to a commitment, and mishandle names, figures and technical terms. The output is fluent and confident, which is precisely what makes a wrong action item survive to the next meeting.

Two habits fix most of this: whoever chaired reads the summary before it circulates, and actions are confirmed in the meeting rather than extracted from it afterwards. Our method for checking AI output applies here as elsewhere — scale the checking to what the summary will be used for.

Confidentiality and where the data goes

Before any of the above: does this tool send audio outside your environment, and to whom? A notetaker in a meeting covering client matters raises the same question as pasting a client document into a public chatbot, with the difference that nobody consciously pasted anything.

Worth establishing: where processing happens, whether recordings train the provider models, who at the provider can access them, and what happens on deletion. Our guide to AI vendor due diligence covers the questions to ask.

A workable position

1. Decide where notetakers are allowed — internal only, or external with consent, and a named list of meeting types where they are not used at all.
2. Standardise the announcement so it happens the same way every time.
3. Set retention and make deletion automatic rather than a task.
4. Name the record. Either the summary is the minutes or it is not. Say which.
5. Review the summary before it circulates. One person, every time.

None of this requires new software. It requires one decision and a sentence at the top of meetings.

Frequently asked questions

Do I need consent to record an internal meeting?

You need a lawful basis and you need to be transparent. Telling people clearly, every time, is the practical floor. Specific rules vary by jurisdiction and by meeting type.

Can a client refuse?

Yes, and they should be asked rather than informed. Proceed without it if they decline.

Is the transcript disclosable?

Treat it as if it is. Retention and access controls are the mitigation, not hoping it is never requested.

Should the summary replace minutes?

Only if someone reviews and approves it. An unreviewed summary is a draft, not a record.

Our CPD courses cover AI use in professional settings including the records and confidentiality questions, and our guide to AI fluency levels sets out what each role needs to know.