Call Recording Consent for AI Phone Agents: Notice, Wording, Retention
Set recording notice, consent wording and retention for an AI phone agent, with a checklist that stays valid as the voice agent changes and improves.

A phone agent that records calls needs three things in place before its first live conversation: a notice that the call is recorded and that the caller is speaking with an AI agent, a clear way to decline recording or reach a person, and a retention period for each kind of data the call creates. Keep the wording under version control so it stays accurate as the agent's prompts, tools and voice change. The rest of this guide explains how to write, log and maintain that notice.
Key takeaways
- Disclose the recording and the AI nature of the agent early in the call, before the caller shares personal details.
- Give a real alternative: a way to decline recording and a route to a person, and confirm the agent honors both.
- Log which notice version was played and how the caller responded, so you can show what was said on any given call.
- Set retention per data class. Audio, transcripts, extracted fields and tool logs are not the same thing and do not need the same period.
- Treat the notice as a controlled document. Any change to what the agent captures, shares or learns from calls should trigger a review of the wording.
What the recording notice must say
A recording notice does two jobs. It tells the caller what is happening, and it gives the business evidence that the caller was told. The notice should come early, before the caller gives a name, address or payment detail, and it should state what the recording is used for. If the agent also transcribes the call, extracts fields such as a name or appointment time, or sends a summary to staff, those uses belong in the notice as well.
Because the caller is speaking with software, most businesses also need a plain statement that an AI agent is answering. Whether that statement is legally required depends on where your callers are and how the system is deployed, so treat it as a question for counsel. Our guide on disclosing AI agents to customers covers the disclosure question in more depth.
Consent wording that matches what the agent really does
Consent is not automatically required for every recording. Rules differ by jurisdiction, by whether every party must agree, and by purpose. Some businesses rely on a notice that lets the caller continue or leave, while others ask for an explicit yes. Decide that with counsel for each market you serve. What you can control is the accuracy of the wording and whether it separates the core purpose from anything else.
Keep the recording notice apart from any request for secondary use. A caller who books a visit should not be quietly agreeing to have their audio used to tune the agent. If you ask permission to use recordings for quality review or evaluation, ask in its own sentence and record the answer. A short opening line that covers the essentials might read like this:
This call is recorded to handle your request and check our service. An AI assistant is answering for [company]. Say 'a person' at any time to be transferred.
The bracketed name and the transfer phrase should match the real behavior of your agent. If the phrase does not work in testing, the notice is inaccurate, whatever the legal analysis says.
Retention: keep each data type for its own purpose
A single call produces several kinds of data, and each one has a different purpose and risk. Write a retention period for each class rather than one period for everything.
- Raw audio: keep it only as long as the stated purpose requires. It is usually the most sensitive item because it carries the caller's voice.
- Transcripts: these may support a dispute or a service review, so they often need their own period, which can differ from the audio.
- Extracted fields: a booking time or service request usually belongs in the business system of record. Avoid letting the call log become a second customer database.
- Tool and agent logs: these are needed for audits, but they can contain personal data inside tool arguments. Redact before storage where possible.
- Evaluation samples: reviewing calls to improve the agent is a distinct purpose. It needs its own notice language, its own access list and its own period.
Set a written period for each class, delete on schedule, and test that deletion actually runs. A retention rule that nobody executes is a gap that appears only when someone asks for the record. For the log side, our article on how long to keep AI agent logs and transcripts explains how to balance audit needs against privacy.
Step by step: set up the notice before the first live call
Use this sequence before the agent takes real calls. It is short on purpose, so that someone can run it in a single working session and then repeat it after each change.
- Map every call flow the agent will handle, including transfers, voicemail, after-hours messages and calls that reach the wrong department.
- List what each call creates: audio, transcript, extracted fields, tool actions, summaries and any scores.
- For each item, write its purpose in one sentence and name the roles that can access it.
- Draft the opening notice with counsel. Keep the recording statement, the AI disclosure and the route to a person in separate sentences.
- Decide what happens when a caller declines recording. The agent should stop recording and still complete what it can, or transfer the caller to staff.
- Set a retention period for each data class and name the person who approves any exception.
- Test the flow with staff calling in on each path, and confirm the notice plays before any personal data is requested.
- Save the notice text, its version number, the test results and the approval record together.
Keep the notice current as the agent improves
A voice agent changes in small steps: a new prompt, a new tool, a new language, a model update, a new link to a customer system. Each change can alter what data is captured or where it goes, and the notice describes the old behavior. Treat notice review as a trigger tied to change, not as an annual task.
Review the notice when any of these happen:
- A new tool or integration receives call data.
- The agent starts capturing, summarizing or scoring something it did not before.
- Recordings are used for a new purpose, including evaluation or improvement of the agent.
- A new vendor or subprocessor joins the call path.
- The same phone number starts serving a new market or language.
- A recording, transcript or caller complaint leads to an incident review.
Illustrative scenario: a field-service company launches a phone agent that books visits and records calls for scheduling accuracy. After a later update, the team adds a summary that is emailed to the sales inbox. The summary now includes a caller's address and remarks about budget. The original notice did not mention sales use, and the retention rule covered audio only. The review catches the gap before the summary reaches more staff. The notice gains one sentence, summaries get their own retention period, and access is limited to the scheduling team.
Who checks the notice, and where the evidence lives
A compliant notice is only useful if you can show what it said and what happened afterward. For each call, keep the notice version, the time of the call, whether the caller accepted or declined recording, any transfer to a person, and the result of each deletion job. Those records are the evidence an internal reviewer or an external auditor will ask for. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) offers a useful structure for organizing these duties around governance, mapping, measurement and management, and the [OECD AI Principles](https://oecd.ai/en/ai-principles) include transparency as a core expectation for systems that interact with people.
In managed operations, Be Human maps each call flow with the client, drafts the notice language for review by the client's legal adviser, configures the agent to play the notice and to write the notice version into every call record, sets retention for each data class, and reviews the notice whenever tools, prompts or integrations change. Monitoring and recurring evaluations check that the agent still plays the notice, honors declines and runs deletion on schedule. Be Human does not give legal advice, and the client's counsel approves the wording.
If you are unsure which phone calls should reach an agent first, the free Be Human AI diagnostic is a practical way to find the first task worth handing over.
Frequently asked questions
Do I always need the caller's consent before recording a call handled by an AI agent? Not always. Whether consent is required depends on the jurisdiction, on whether every party must agree, and on the purpose of the recording. Many businesses use a clear notice that lets the caller continue or leave, and some ask for an explicit yes, especially for secondary uses. Confirm the approach for each market with counsel before launch.
Must the AI disclosure be separate from the recording notice? Keeping them separate is the safer design. The recording notice explains what is captured and why, while the AI disclosure tells the caller who or what is answering. Combining them into one vague sentence makes it harder to prove that each message was given and harder to update one without changing the other.
How long should call recordings and transcripts be kept? There is no single answer that fits every business. Set a written period for each data class based on its stated purpose, record who may access it, and make sure deletion runs automatically or on a schedule that someone reviews. Audio often needs a shorter period than transcripts, and transcripts often need a shorter period than the business records they feed.
What happens when a caller says no to recording? The agent should stop recording from that point, tell the caller what it can still do, and either complete the request without recording or transfer the call to staff. Test this path with real phrasings, because callers rarely use the exact words in your notice. Log the decline without storing the audio, so you can show that the choice was respected.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


