The disclosure deadline arrived. Most voice agents are not ready for it.

Regulation & governance · August 2026 · PolyQ Team

The disclosure deadline arrived. Most voice agents are not ready for it.

Article 50 of the EU AI Act took effect on 2 August 2026. Any system that talks to a person now has to say it is a machine, at the start of the conversation, in a way a reasonable member of that audience would notice. Burying it in the terms of service does not count.

On 2 August 2026 the EU AI Act’s transparency obligations became enforceable. If you operate a chatbot, a voice agent, or anything that autonomously contacts a person, the rule is short: disclose that they are dealing with AI, unless it is already obvious. The interesting part is how narrowly “obvious” has been drawn.

  • This one did not slip: The Digital Omnibus amendments, approved by the European Parliament in June 2026, pushed the high-risk obligations back by roughly 12 to 16 months. Article 50 and the Commission’s general-purpose AI enforcement powers were left alone. Teams that read “AI Act delayed” in a headline and stopped there have a gap.
  • First interaction, not the footer: Disclosure has to be clear and distinguishable, at the latest at the moment of first interaction. Guidance is explicit that a line in the terms and conditions or a legal notice does not satisfy it. For a phone agent, that means the opening seconds of the call.
  • “Obvious” is judged by your actual audience: The Commission’s guidelines assess obviousness against an average member of the system’s intended audience, which pushes toward more disclosure wherever children, older users, or other vulnerable groups are plausibly involved. A synthetic voice being good is an argument for disclosing, not against it.
  • Consent and disclosure are different problems: US two-party-consent recording law and EU AI disclosure are separate obligations with separate triggers. A product that handles one is not thereby handling the other, and both belong in the software rather than in a staff checklist.
  • What we do about it: Intake Manager opens every interview by identifying itself as an AI interviewer and applies jurisdiction-aware consent language automatically. It was built that way before the deadline, because retrofitting disclosure into a conversational flow is considerably harder than designing for it.

Regulatory transparency and product quality point the same direction here. People behave better with a system when they know what it is, and the deadline has simply made that a requirement rather than a preference.

If you are running voice or chat in the EU and are unsure where your disclosure sits, that is worth a conversation — whether or not it ends with you using anything we build.

Sources

More from Thinking