Minutes Without a Minute-Taker: Governing AI Notetakers in UK SMEs

Consider a scene now familiar in British offices. A team lead, weary of writing up calls, adds an AI assistant to their calendar. It joins meetings, produces a summary and emails everyone a tidy list of actions. Within a fortnight three colleagues have done the same; by the end of the quarter the firm holds a searchable archive of conversations it never intended to keep — including the ones about a restructure, a customer dispute and a colleague's health.

None of this is necessarily unlawful. But it is a decision, and it has been taken by default rather than deliberately. The questions that ought to precede installation — who owns the transcript, who may read it, how long it is held and on what lawful authority — typically arrive afterwards, once the tool is embedded and removing it would be awkward.

What follows is a practical sequence for managers who suspect the tool is already in use, or who are about to approve it properly.

Start with the data, not the feature list

Before comparing vendors, sketch the flow. Audio is captured in the meeting; it is transmitted to a processing environment; it may be held temporarily; it is converted to a transcript; the transcript is summarised, sometimes by a second model; the summary is stored and distributed; and copies are retained on participants' devices and in email. Every one of those stages is a place where personal data leaves your control.

Two distinctions matter enormously here. The first is whether the vendor processes only what it needs, or retains the audio as well as the text. The second is whether the summaries are generated inside the vendor's own systems or passed to a third-party model provider. Ask for the sub-processor list in writing; a named list is a contractual artefact, whereas a paragraph promising "trusted partners" is marketing.

The vendor questions worth putting in writing

A short list, sent as an email rather than discussed on a call, will separate the serious suppliers from the rest:

• Where is meeting data stored and processed? "UK or EU region" should be a setting you can verify, not an assurance you must trust.
• Is customer data used to train or improve models? The answer should be an unambiguous no, with a contractual commitment.
• What are the retention defaults, and can an administrator change them per workspace or per meeting type?
• On termination, how is data deleted, and can you obtain a deletion certificate?
• Who inside the vendor can access transcripts, under what conditions, and is that access logged?
• Which certifications apply — ISO 27001, Cyber Essentials Plus, SOC 2 — and will they share a recent penetration test summary under NDA?
• Can the tool be barred from specific calendars, meeting types or individuals?

If a supplier cannot answer these in writing within a week, that is itself an answer.

Lawful basis is where most rollouts stumble

Under the UK GDPR you need an Article 6 basis for processing. In an employment setting, consent is a fragile foundation: staff may feel unable to refuse, and consent can be withdrawn at any time, leaving you with a half-populated archive. Most firms land instead on legitimate interests, supported by a documented assessment that weighs the operational benefit against the effect on individuals. That assessment must be genuine, not a template with the company name inserted.

The harder issue is incidental capture. A meeting about a sick leave return, a grievance or a religious holiday observance can bring special category data into the transcript without anyone intending it. That requires a further condition under Article 9, and it is often simpler to exclude such conversations from automated capture altogether than to construct an elaborate legal justification after the fact.

Employee monitoring, described plainly

A notetaker that joins every calendar invitation is, in effect, a monitoring technology. UK regulators have been consistent on the principles: monitoring must be proportionate, targeted and transparent, and covert monitoring is difficult to defend outside narrow cases of suspected wrongdoing. A one-to-one performance conversation recorded without a clear business need is unlikely to pass that test.

A data protection impact assessment is sensible before rollout and arguably required where monitoring is systematic. It should record what is captured, why, who can see it, how long it is retained and what alternatives were considered — for example, restricting capture to external client calls where the record has a demonstrable purpose.

Retention: the setting that does most of the work

The single most effective control is a short default retention period. Many tools ship with indefinite storage because it suits the vendor and flatters the product. Set a workspace default of 30 or 90 days depending on need, permit exceptions only for meetings with a clear record-keeping purpose, and review the exceptions quarterly. Bear in mind that transcripts are personal data: they fall within subject access requests and may be disclosable in litigation, so a sprawling archive is a liability as well as a convenience.

Also decide, deliberately, whether participants may download or forward transcripts, and whether the bot appears in the meeting roster under a name that makes its presence obvious.

A short script for telling staff

You do not need a policy document to begin the conversation. Something close to the following works:

"From next month we are trialling an AI assistant that takes notes in internal project calls and external client calls. It will not join one-to-ones, HR conversations or anything marked private. Recordings are deleted after 30 days and only the project team and IT administrators can access them. You can ask for it to be removed from any meeting, and we will review the trial in three months. If you have concerns, speak to [name] or raise them with [name] in confidence."

Plain, specific and revocable. It does more for trust than a four-page policy circulated after the fact.

The checklist before you switch it on

Confirm the region, the retention default, the training-data position and the sub-processor list. Complete a short impact assessment. Exclude sensitive meeting types by rule rather than by hope. Name an owner for the tool. Set a review date. Then, and only then, add it to the calendar.