Article 50 of the EU AI Act kicks in August 2, 2026 — voice agents have to disclose they’re an AI, audibly, right at the start of the call. Here’s our plan on LiveKit, curious how others are doing it.
Our approach:
Play a fixed disclosure line via session.say() the moment the session starts, before the LLM loop — “You’re speaking with an AI assistant.”
Hardcoded string (not LLM-generated) so wording stays compliant.
Barge-in disabled so it actually gets heard.
TTS pre-cached (same line every call) to save latency/cost.
Configured per-language for multi-region calls.
Questions:
For inbound SIP, where do you trigger it — participant_connected or the agent entrypoint? Any issues with audio not being ready yet?
Always-disclose, or are you relying on the “obviousness” exception? (We’re going always.)
How are you logging proof of disclosure for audits?
(Not a lawyer — just our engineering take.)
Thanks @s , I highlighted this post internally as we are also considering what we can do to make the compliance journey easier for our developers.
This is a very important topic, I’m also interested in how others in the community are preparing for this, and any questions they might have.