What the Council position actually changes for restaurant operators

It is Monday morning at the newsroom and I have spent the weekend reading the Council document the trades are now calling the streamlining vote. The European Council agreed its position on Friday, March 13, and the inbox here has been a steady drip of operator notes asking the same question in different words: did the AI Act just get pushed back, and if so, do I need to redo my Q3 plan? The short answer is the contrarian one. Most operators do not need to change anything. The deferral is real. It is also narrower than the headlines.

The contrarian read — and this is interpretation, not anything the Council said — is that Friday’s vote moves exactly one part of the compliance calendar in a direction that matters to restaurants, and leaves the parts most operators are actually building for sitting on the original timeline. If your Q3 compliance plan is built around Article 50 disclosures and the August 2 voice-agent deadline, nothing changed on Friday. If your Q3 plan has a high-risk classification workstream attached to it, you just got runway.

What the Council moved, and what it did not

Walk the position document and the structure is cleaner than the press coverage suggests. The Council’s streamlining package holds the prohibitions calendar in place — the Article 5 list of banned uses has been in force since February 2025 and is not touched. The GPAI rules — the general-purpose AI model obligations covering the providers of the foundation layer — remain on their published cadence. And critically for our beat, the Article 50 transparency obligations for systems that interact with humans or generate synthetic content stay on the original August 2, 2026 date.

What moves is the high-risk file. The Annex III obligations — the conformity assessment, technical documentation, post-market monitoring, and registration duties that attach to AI systems classified as high-risk under the Act — get a longer runway. The Council position pushes the effective date for the bulk of the high-risk regime to August 2, 2027, with a transitional regime through that window. The penalty structure on the other side of the door is unchanged: the €15 million or 3 percent of global turnover cap — whichever is higher — applies once the obligations bite.

(Interpretation flag: I am calling Friday’s vote a deferral on the high-risk file because that is the operational read for our category. The Council frames it as streamlining, which is the right political word and the wrong operator word. Same calendar effect.)

Which restaurant deployments actually sit in the high-risk bucket

The high-risk classification is the part of the Act that has been quietly stalking three specific restaurant deployment scenarios since the text finalised. None of them are guest-facing in the way a voice agent is. All of them are operator-side.

HR and biometrics in the back office. Any AI system used for recruitment, selection, performance evaluation, or task allocation of workers — including the algorithmic shift-fitting tools several of our Four Margins interviewees have been piloting — sits in the high-risk bucket the moment it touches employment decisions. The conformity-assessment burden on that workstream is now an August 2027 problem, not an August 2026 one.

Biometric kiosks at the counter. Face-matched loyalty, fingerprint clock-in at the POS, palm-scan payment at the order kiosk — these are biometric identification deployments under the Act, and they are high-risk by default unless the operator can document a narrow exception. The deferral applies. The prohibitions on real-time public-space biometric ID still do not — those are Article 5 and unchanged.

Emotion recognition in drive-thru cameras. The QSR vendor stack has been quietly shipping camera systems that read facial affect to time the order prompt and the upsell. Emotion recognition in workplace and education contexts is prohibited outright. Emotion recognition in a customer-facing drive-thru lane is high-risk, not prohibited — and the high-risk obligations on that surface now ride the new calendar.

The pattern across the three: nothing on Friday’s vote tells an operator they can do something they could not do before. The deferral changes when the documentation and conformity-assessment workstream has to be complete. It does not change the substantive rules.

What Q3 planning should actually look like now

Three threads, in order of operator urgency.

One. The Article 50 voice-agent disclosure work that our coverage of the Second Draft Code walked through two weeks ago is unchanged. August 2, 2026 is still the trigger. The disclosure script, the consent log, the language-localised opening — all of that is the Q3 deliverable it was on Friday morning. Do not let the Council headline slow that work down.

Two. The high-risk workstream — if you have one already running for HR, biometric kiosks, or drive-thru cameras — gets a real twelve-month extension. The right Q3 move is to keep the documentation discipline in place and stop paying the consultancy rate for a sprint. The work shifts from urgent to important. The companion piece our colleagues filed Friday on the broader Digital Omnibus package walks the political logic; I am narrower here on the restaurant-deployment read.

Three. The EU AI Act applied to restaurants framework is the one to pull back out. The article-by-article matrix it builds is the right artefact to update against the Council position. The Commission’s regulatory-framework page is the cleanest single source for the consolidated dates as the trilogue moves through April.

The Council position is not a reprieve. It is a re-calendaring. The bits of the Act most operators are actively building for did not move. The bits a smaller subset of operators were dreading got runway. That is the whole story.

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