Brussels Just Told Voice Agents to Wear Name Tags

It is Wednesday morning at the newsroom and the filing I have been waiting for since the AI Office’s January workplan finally hit the European Commission’s Shaping Europe’s Digital Future page yesterday afternoon. The Second Draft Code of Practice on the Marking and Labelling of AI-Generated Content is live, the comment window is open through March 30, and the document is doing exactly what the Article 50 watchers have been bracing for. It is telling synthetic-audio systems that they have to identify themselves to the human on the other end of the line.

The contrarian read — and this is interpretation, not anything the Commission said — is that the restaurant voice-agent category is the first hospitality surface that has to ship a compliance artefact, not a feature, before August 2. Wendy’s FreshAI, SoundHound, Vox AI, Slang, PolyAI: every one of them is on the hook in EU markets the moment the Code finalises. UK operators are watching. They are not yet bound.

What Article 50 actually says, and what the Second Draft adds

Article 50 of the EU AI Act is the transparency floor for systems that interact with humans or generate synthetic content. The relevant operative line, for our category, is the one requiring that a natural person interacting with an AI system is informed of that fact unless the interaction is obvious from context. The phone is not obvious context. A diner calling a restaurant at 7:14 p.m. for a Friday booking has no reason to assume the voice asking how many in your party is anything other than a host. Article 50 says: tell them.

The Commission’s implementation timeline is unambiguous on the date. Article 50 obligations enter into force August 2, 2026. The Second Draft Code published yesterday is the operational manual — what counts as a valid disclosure, what counts as a valid marker on synthesised audio, what counts as a documented exception. The comment window closes March 30. The finalisation cadence the AI Office has been running suggests a published final text in May, roughly three months before the obligation bites. That is not a long runway.

The piece of the Second Draft that matters most for voice-agent operators is the carve-out language for real-time interaction contexts. The draft, by my reading, accepts that a synthetic voice does not have to interrupt every utterance with a fresh disclosure — but it does require that the disclosure happen at the start of the interaction, in clear language, and in the language of the caller. That is a script change. It is also a logging change, because the burden of proof that the disclosure was made will sit with the deploying party.

Who has to write the script

The deploying party, in EU vocabulary, is the operator. The vendor provides the system; the operator deploys it. That means a restaurant in Dublin running PolyAI’s reservation surface or a Madrid group running Slang has the legal exposure, not the vendor — though vendor contracts will increasingly carry indemnity language as the August date approaches. The named vendors operating into EU markets that I would put on the immediate watch list:

• Wendy’s FreshAI — limited EU footprint today; the obligation lands on any pilot that crosses an EU member-state border before August 2.
• SoundHound — broad voice-AI deployment across QSR; existing automotive and restaurant deployments in EU markets are already in scope.
• Vox AI — restaurant-specific phone-answering layer with European operator interest growing through Q1.
• Slang — North-American base, but operator groups with cross-border footprints (notably the Founding Farmers and DineAmic-equivalent multi-market chains my colleague Eitan flagged in his February ROI refinement) will face mixed compliance regimes.
• PolyAI — the closest-to-native EU vendor, founded in London, with the most existing European bookings volume; the operational burden of the Code will fall hardest here first.

The compliance artefact for each operator is the same: a documented, language-localised opening disclosure; a logged consent or acknowledgement; and a defensible record of which calls went to the agent and which were handed off. Nothing about that is technically hard. All of it is process-hard, and the gap between we plan to disclose and we can produce the disclosure log on request is the gap operators have to close between now and August.

The UK is watching, not bound

UK operators reading this for second-hand panic should slow down. The UK’s AI regulatory approach remains the principles-based, regulator-led model the ICO and CMA articulated in 2024 and have not materially revised since. There is no UK equivalent of Article 50 with an August 2 trigger. The Online Safety Act’s transparency duties touch some of the same surface for consumer-facing AI but do not bind a restaurant’s phone agent in the same way.

What UK operators should do is exactly what their EU peers are now forced to do: write the disclosure script, log the consent, and keep the audit trail. The market will move there regardless of whether the law does, because cross-border operator groups will standardise on the EU artefact rather than maintain two regimes. If you are an operator in London with a sister property in Paris, your Paris counsel will write the script and your London property will run the same one. Compliance arbitrage at the script level is not worth the operational complexity.

What I would do this week

Three steps, in order. One: ask your voice-AI vendor for their draft EU disclosure language. If they do not have one circulating internally by mid-March, that tells you how serious their EU readiness is. Two: pull the log architecture. Can your vendor produce, on demand, a record showing the disclosure was made on a given call to a given number? If the answer is we would need to build that, build it. Three: file a comment by March 30 if you operate in the EU. The Commission’s window is open for a reason; operator-side voice on a voice-AI Code is the constituency the AI Office has heard from least.

The forthcoming Voice Agent Maturity Curve piece — and the broader EU AI Act applied to restaurants read — sit alongside this. Today’s filing is the moment the abstract framework becomes a deliverable on the August 2 calendar. Name tags. Voice agents wear them now.

— Hana edits the newsroom for TableTransfers. Tips: [email protected].