Somewhere in your business, right now, an AI notetaker is probably sitting in a meeting. Maybe you put it there. Maybe someone on the sales team installed it during a free trial eighteen months ago and nobody has thought about it since. Either way it is listening, transcribing, and storing a conversation that involves at least one person who never signed anything.
This is the most common form of shadow AI in small businesses, and it is the one most likely to produce an uncomfortable email from a client's solicitor. The good news is that getting it right is genuinely straightforward: a short script, a retention rule, and one page of policy. This guide covers what the rules actually require, the three consent models that hold up in practice, and wording you can use on your next call. It is general guidance rather than legal advice — if you handle health records, financial advice, or anything involving children, have a solicitor review your policy.
Why AI notetakers became a compliance problem
Recording a meeting used to be a deliberate act. Someone had to press a button, and everyone in the room watched them do it. Three things changed at once, and small businesses absorbed all three without noticing.
First, the bot joins on its own. Most notetakers connect to your calendar and attend every event that has a video link — the client review, the supplier negotiation, and the one-to-one with a team member who is struggling. Nobody decides to record those meetings. The default decides.
Second, the recording leaves your premises. A 45-minute call produces roughly 7,000 words of verbatim speech, and that transcript is processed and stored by a third-party company that you are, in data protection terms, legally responsible for choosing.
Third, the output is searchable forever. A handwritten note gets filed and forgotten. A transcript sits in a shared workspace where anyone with access can search "redundancy" or "price" or an employee's name across every conversation your firm has had this year.
Owners describe the result in familiar terms: "we've got four tools, and nobody knows what any of them actually do." With notetakers, that gap is not just untidy. It is a record you are accountable for.
What the rules actually require
Two separate obligations apply, and people routinely satisfy one while ignoring the other.
Recording law governs whether you may capture the conversation at all. Rules vary by jurisdiction: some places allow recording with the consent of a single participant, others require everyone's agreement, and several US states fall firmly in the second camp. Checking the rules call by call is not realistic for a small team. The practical answer is to adopt all-party consent as a blanket house rule — it is cheaper than being wrong once.
Data protection law governs what happens next. Under the UK and EU GDPR, a transcript of identifiable people is personal data, which means you need four things in place:
- A lawful basis for processing — usually legitimate interests for ordinary business meetings, with a documented balancing test; consent where the power balance makes legitimate interests shaky, such as with employees.
- Transparency — your privacy notice should say that you use AI meeting assistants, what they capture, and how long you keep it.
- A processor agreement with the vendor (a data processing addendum), which almost always exists but almost never gets signed on a self-serve free plan.
- Knowledge of where the data is stored and whether it leaves your region.
There is also an AI-specific layer. Under the EU AI Act, most notetakers are low-risk tools with transparency duties — people should know an AI system is involved. But one feature set crosses a brighter line: inferring emotions from employees in the workplace is a prohibited practice. If your tool scores staff on "sentiment", "enthusiasm" or "engagement" during internal calls, that is a materially different risk from transcription, and it should be switched off. Our guide to AI and GDPR for small businesses covers the wider compliance picture.
The three consent models, and when each one works
You do not need a bespoke approach per meeting. You need three patterns and the judgement to pick between them.
1. Standing consent in the engagement terms
A clause in your contract or engagement letter stating that calls may be recorded and transcribed using AI assistants, with the purpose and retention period named. Best for retained clients and recurring project work. Mention it once at kickoff so it is not a surprise buried in clause 14. Weakness: it does not cover people who join the call but are not party to the contract — your client's new finance director, for example.
2. Verbal consent at the top of the call
The workhorse for sales calls, discovery calls, and anything with a new counterparty. You say the line, you get an audible yes, and the first thirty seconds of the recording is itself your evidence. Keep that opening segment even if you delete the rest.
3. Advance notice in the calendar invite
A standing line in your meeting template. This is the right baseline for internal meetings, where it should sit alongside a written policy, but it is not sufficient on its own for client calls. Notice is not the same as agreement, and a bot's own small on-screen banner is weaker still — nobody reads it, and you cannot prove they did.
In practice most small firms run all three: terms for retained clients, a verbal line for everyone else, and invite notice for internal calls.
Where small businesses get caught out
The failures are consistent enough to list.
The auto-join setting. The bot attends a performance conversation, a grievance discussion, or a call with your accountant about a difficult quarter. Fix: disable blanket auto-join and turn recording on per meeting, or exclude internal one-to-ones by calendar rule.
Recorded job interviews. The moment you transcribe an interview, you have created an HR record. The candidate can request a copy, and your hiring notes are now verbatim rather than considered. Many firms decide the risk is not worth the convenience.
The open transcript folder. Summaries sync to a shared drive or a team channel that everyone can search. Commercially sensitive negotiations and candid staff comments end up visible to the whole company.
Consumer-tier training terms. Free and personal plans frequently permit the provider to use content to improve their models. Business tiers generally do not. If client conversations are going through a free account, that is the first thing to change.
The tool nobody chose. Shadow installs are the norm. Run the audit: ask every person with a calendar which assistant is connected to it. The answer is usually more tools than anyone expected. If you want a structured approach to that, our piece on how to audit your AI tool stack sets out the process.
A consent script you can use tomorrow
Keep it short, say it confidently, and move on. Hesitancy is what makes people uneasy, not the request.
At the start of a call: "Before we get going — I use an AI assistant to take notes so I can stay in the conversation rather than typing. It produces a transcript and a summary that stays with us and is deleted after 90 days. Are you happy for me to run it?"
In the calendar invite: "Please note: we use an AI notetaker for meeting summaries. Let us know in advance if you would prefer we don't, and we'll take manual notes instead."
In a client onboarding email: "For our calls we use an AI meeting assistant to produce summaries and action points. Transcripts are stored in our workspace, are not used to train any AI model, and are deleted after 90 days. Tell us at any point if you would rather we switched it off."
When someone says no: turn it off visibly, say so out loud, take manual notes, and send a written summary afterwards. Then record the preference somewhere durable — the CRM record, not your memory — so it holds for the next call too. One declined request handled gracefully buys more trust than a year of smooth recordings.
The same openness applies more broadly; we covered the wider question in should you tell clients you use AI.
Retention, access, and the transcript nobody deletes
Consent gets the attention. Retention is where the actual exposure accumulates, because almost every tool keeps transcripts indefinitely unless you tell it otherwise.
Set a default and enforce it in the tool's settings rather than by intention. For most small businesses, 90 days is a sensible figure: long enough to revisit a conversation, short enough that you are not sitting on two years of verbatim client speech. Once the summary and action points are in the CRM, the raw transcript has done its job.
Restrict access deliberately. Client call transcripts should be visible to the people on the account, not to the whole workspace. Internal meeting transcripts should be visible to the attendees.
Finally, be ready for the request. A client or an employee can ask for a copy of what you hold about them, and a transcript is squarely within scope. "We're not sure where it is" is not an answer that ends well. A quarterly check is enough: export the list of recordings, count them, look at the oldest date, and confirm the retention rule is actually running.
Your one-page AI meeting policy
One page, circulated to everyone, reviewed twice a year. It should answer seven questions and nothing more:
- Which tool is approved — named, on a business tier, with a signed processing agreement.
- Which meetings may be recorded, and which never are (one-to-ones, grievances, interviews, anything legally privileged).
- How consent is obtained for each meeting type, using the scripts above.
- How long transcripts are kept, and who is responsible for checking the setting.
- Who can access transcripts, and where they are stored.
- What is switched off — emotion or sentiment scoring on staff, auto-join, public sharing links.
- Who to ask when something falls outside the policy.
If you do not yet have a broader framework for this, our walkthrough on how to write an AI policy for a small business gives you the surrounding structure, and this page slots into it.
None of this is onerous. An afternoon of settings changes and a sentence you learn to say will put you ahead of most firms your size — and it protects the thing AI notetakers are genuinely good at. Used properly, they let you stay present in a client conversation instead of typing through it. That is a real gain. It is just not worth trading a client relationship for, and with twenty minutes of setup you never have to.
Where does your business stand on AI?
Take the free 3-minute AI Readiness Quiz and get a personalised score with your next steps.
Take the Free Quiz →