Always tell; whether to ask depends on where you operate and why you record. Every customer should know at the start that the conversation is recorded or transcribed and why. In some places, and for some purposes, you also need their consent. An AI support agent should deliver the notice, record that it did, and handle a refusal without pressure.
Why is an AI conversation always "recorded"?
A human phone call can happen without a recording. An AI conversation almost never does. The agent works from a transcript, its replies are generated text, and both are usually stored for quality, audit and improving the service. So the question is rarely whether to record; it is whether customers know, and on what basis you keep it.
That changes the default. Treat every AI conversation, in any channel, as recorded, and tell customers so.
What decides whether consent is needed?
The law where the customer is. Rules on recording calls vary. Some jurisdictions require only that one party to the call knows it is recorded; others require that every party consents. Data protection laws add their own requirements for any personal data the recording contains. Under the GDPR and UK GDPR, consent is one of several lawful bases under Article 6; legitimate interests or legal obligation may apply instead for some purposes, but the customer must still be told. Other regimes put more weight on consent, as covered in PIPEDA consent and AI agents and DPDP Act consent and AI agents.
The purpose. Keeping a record so you can resolve a dispute is different from using conversations to test or improve the agent, and different again from analysing customers' voices. Each purpose needs its own basis and its own notice, and purposes beyond the obvious ones often need consent.
The content. Conversations that include health information or other special category data carry stricter rules.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What should the agent do?
| Moment | The agent |
|---|---|
| Start of a call or chat | States that the conversation is recorded and why, and where to find the privacy notice |
| Where consent is required | Asks plainly and waits for a clear yes |
| Customer refuses | Explains what it can still do, such as limited help without recording, or offers another route |
| Customer asks how the recording is used | Gives the approved explanation and points to the privacy notice |
| Customer asks for the recording | Records the request and routes it as a data access request |
| Handover to a person | Passes on what the customer was told and agreed to |
A refusal should not be met with persuasion. If the business cannot serve the customer without recording in that channel, the agent says so and offers another way, such as a written channel with its own notice, or a callback from a person under different arrangements.
An illustrative example
Take an insurer whose voice agent takes calls in several countries. The opening line in every market says the call is handled by an AI assistant and is recorded for quality and to keep a record of the conversation.
In one market, the insurer's counsel has advised that recordings used to test and improve the agent need the caller's consent. There, the agent adds: "We'd also like to use this recording to improve our service. Is that all right?" A caller says no. The agent confirms the call will still be recorded as a record but excluded from improvement work, flags the conversation accordingly, and carries on. If the caller later asks to speak to a person, the handover note states what they agreed to, so the person does not ask again; the handover itself is covered in how a support agent should hand a conversation to a person.
What stays with people?
Deciding the lawful basis for each purpose, writing the notice, setting retention, and handling requests for recordings or deletion. The agent delivers the notice and records responses. It does not interpret law or decide exceptions.
What controls make it reliable?
- Notice wording per market and channel, approved and versioned.
- Consent captured as a field, not inferred from the transcript.
- Purpose flags on each recording, so excluded recordings never reach testing or improvement sets.
- Retention by purpose, deleted on schedule.
- Notice and consent in the log. What the customer was told and what they said is filed with the conversation, as set out in what a support agent should log about each conversation.
Where Pryme fits
In Pryme, the customer's notice and consent are recorded with the conversation, and they travel with it into the Workspace on handover. Conversations flagged as excluded from improvement are kept out of the cases Attest replays. More is in the AI support agent that knows when to say no.
Questions people ask
Does a chat need a recording notice as well as a call?
Yes. A chat is stored just as a call recording is. Tell customers it is kept and why.
Can we use old recordings to train or test a new agent?
Only if the purpose was covered when customers were told, or another lawful basis applies. Check before reusing.
Should the agent pause recording when card details are given?
Card data should not reach the agent at all. Use a secure capture method that bypasses the recording.