What do voice agents need for call recording consent and AI disclosure?
At the start of every call, say that the caller is talking to an AI, that the call is recorded and that outside providers process it, and follow the strictest state on the call, since several US states require every party’s consent to record. For outbound calls, an AI voice counts as an “artificial voice” under the TCPA, so you need the person’s prior express consent before dialing.
The rules that matter in late 2026
- All-party consent
- California, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Pennsylvania and Washington are among the states that require every party’s consent to record, and the strictest state usually governs an interstate call (Rev (blog, confirm)). You rarely know where the caller is, so disclose and ask on every call.
- California privacy law (CIPA)
- In Ambriz v. Google (N.D. Cal., February 2025), the court applied a “capability” test: a contact-center AI vendor that merely can use call data for its own purposes may be a third-party eavesdropper, and its contract promising not to didn’t get the claim dismissed (ZwillGen, Goodwin). Pair the disclosure with technical limits: per-customer keys and no training on raw call data.
- Outbound calls (TCPA)
- The FCC ruled in February 2024 (FCC 24-17) that AI-generated voices are “artificial” under the TCPA, so outbound AI calls need prior express consent (FCC). Keep a consent record and block any dial without one.
- Patients in California (AB 3030)
- A generative-AI message to a patient about clinical information needs a disclaimer at the start and end of an audio interaction and instructions for reaching a human, unless a licensed provider reviews it (ArentFox Schiff, CDPH). Scheduling-only messages are administrative, but calls drift.
- EU callers
- Article 50(1) of the EU AI Act requires telling people they are interacting with an AI system, applying from August 2, 2026 (Jorijn (news report, confirm)). Confirm against the text on EUR-Lex.
- Card numbers
- If callers read card numbers aloud, recordings and transcripts come into PCI DSS scope, and pausing the recording doesn’t take them out (Paytia (vendor source)). Send a payment link or transfer to a compliant line.
A disclosure that covers the basics
Example only. The customer’s counsel signs off on the final wording
Hi, you’ve reached the virtual assistant for [Company]. I’m an AI. This call is recorded and transcribed, including by our service providers, to help with your request. You can ask for a person at any time. Is it okay to continue?
For clinical calls with California patients, also close with the AI disclaimer and how to reach a person, for example: “You were speaking with an AI assistant. To talk to someone on our team, say ‘person’ or call [number].”
How to make it hold at launch
- Put the disclosure in the agent’s first turn and test it. In replay, every call must open with it, word for word, before anything else.
- Get sign-off before shadow mode. Confirm the all-party consent rules with the customer’s counsel and approve the exact script.
- Decide what happens on “no”. Transfer to a person without recording, or offer a callback. Test that path too.
- Turn off training on raw customer call data and say so in the DPA, with the list of providers that process the audio.
- For outbound, check the consent record before every dial.
- For health, watch for clinical drift. A scheduling call that turns into symptoms brings in the AB 3030 rules.
Short answers
Does an AI voice agent have to say it is an AI?
In a growing number of cases, yes. The EU AI Act requires it from August 2, 2026, and California’s AB 3030 requires it for generative-AI clinical messages to patients. Saying it at the start of every call is the simplest safe default.
Do I need consent to record calls with an AI agent?
Several US states require every party’s consent, including California, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Pennsylvania and Washington. Disclose the recording and get consent at the start of every call, and confirm the rules with counsel.
Is a contract enough to protect an AI voice vendor under California privacy law?
Not on its own. In Ambriz v. Google (2025), a court let a privacy claim proceed against an AI vendor that was able to use call data despite its contract. Combine disclosure with technical limits such as per-customer keys and no training on raw call data.