What the EU AI Act Prohibitions Mean for Hotels and Restaurants

I am sitting in a glass-walled meeting room in a four-star European city-centre hotel, a printed copy of Article 5 of the EU AI Act on the table, three highlighters lined up against the spine of a binder, and the General Manager’s compliance lead halfway through a coffee that has gone cold while we work out which of the hotel’s existing AI tools are now, as of six weeks ago, contraband. The compliance lead is calm. The GM, who has joined for the last hour, is not. The reason the GM is not calm is that one of the tools we have just put into the prohibited column is the host-stand “guest-mood” camera that the chain’s innovation team rolled out across thirty-one properties last summer with a press release and a LinkedIn announcement post celebrating the partnership. The vendor pitch was that the camera would read arriving-guest emotion and prompt the front-desk agent to adjust their script. Article 5(1)(f), in effect since February 2, prohibits exactly that deployment.

This is the desk review I have been wanting to write since the prohibitions clock started ticking. The contrarian thesis I want to mark up front, before the section breaks: the Article 5 prohibitions are not the regulatory housekeeping the trade press treated them as. They are a substantive product-line ban on a category of hospitality-AI that vendors have been selling into European hotel and restaurant groups for two years, and the Commission’s February 4 guidelines made the ban broader, not narrower. Operators who have not done an Article 5 inventory by the time enforcement powers activate on August 2 are running an unmitigated penalty exposure of up to €35M or 7% of global annual turnover, whichever is higher.

Vibe Check verdict

Verdict: Comply. This is not a vendor-relationship to negotiate; it is a contract category to retire. The four deployments in the callouts below are either prohibited outright or sit close enough to the line that the cost of being wrong is structural.

Ratings, on the standard Vibe Check 1-to-5 scale, with interpretation marked:

• Prohibitions breadth (4.5/5). The eight prohibited practices in Article 5 are broader than the consultation drafts suggested, and the February 4 Commission guidelines extended the practical reach further into hospitality use cases than vendors had been pricing for.
• Literacy scope (3.5/5). The Article 4 AI literacy obligation, also effective February 2, applies to every operator using AI in the workplace — not just deployers of high-risk systems. The scope is real; the enforcement teeth around literacy specifically are softer than around the prohibitions.
• Penalty teeth (5/5). €35M or 7% of global annual turnover is the highest cap in the Act, reserved for Article 5 infringements. By comparison, GDPR’s cap is €20M or 4% of turnover. The drafters wanted operators to pay attention.
• Enforcement timeline clarity (2.5/5). Four deadlines, three of them inside an eighteen-month window, and a trade press that has confused them in print at least four times this month. The deadlines are knowable, but the operator-facing communication has been muddy enough that I have spent more time disambiguating them than explaining what they prohibit.

The case for marking enforcement-timeline clarity low rather than high is in the next section.

The four deadlines, marked clearly

The most common error in the operator-facing trade-press coverage I have read this month is collapsing the AI Act’s four staggered deadlines into a single “AI Act effective date.” There is no single date. There are four, and they govern different obligations, and the penalty exposure varies materially depending on which deadline applies to which deployment. I am going to spend a beat on each because the rest of the piece does not work if the timeline is not clear.

Deadline (a) — February 2, 2025 (past). Article 5 prohibitions and the Article 4 AI literacy obligation took effect. From this date, deploying a prohibited AI system in the EU is unlawful, regardless of whether the formal enforcement powers under the Act are active. Operators using a prohibited system today are exposed to liability through national-law channels and through downstream civil exposure even before the central enforcement framework switches on.

Deadline (b) — August 2, 2025 (forward). The formal enforcement powers under the Act activate. Penalties — up to €35M or 7% of global annual turnover for Article 5 breaches — become directly enforceable through the designated national competent authorities and, for general-purpose AI, through the AI Office at the European Commission.

Deadline (c) — August 2, 2025 (forward, same date as b). Governance provisions and the rules covering general-purpose AI models (GPAI) take effect. The GPAI rules are mostly addressed to model providers rather than deployers, but the deployer-side obligations for operators using GPAI as part of their stack — transparency, instructions-for-use compliance, basic record-keeping — start counting from this date.

Deadline (d) — August 2, 2026 (forward). The high-risk AI system rules apply. This is the deepest set of obligations and covers the bulk of operator-relevant AI categories outside the Article 5 prohibitions — biometric identification, AI used in employment decisions, credit-relevant scoring, and a long list of others. Most operators I have spoken with have been planning their compliance work around this deadline. They have been planning it around the wrong one. Deadlines (a) and (b) are the ones that matter for 2025 budgets. Deadline (d) is the one that matters for 2026 and 2027.

The prohibitions section below is governed by (a) and (b), the literacy section by (a), the GPAI question by (c), and the high-risk question — which I treat briefly because it is not the focus of this review — by (d). Conflating them gets you compliance work scheduled against the wrong fiscal year.

What Article 5 actually prohibits in a hotel or restaurant

The eight prohibited practices in Article 5 are written for AI deployments generally; the operator question is which of them bind specifically on hospitality use cases. Reading the Commission’s February 4 guidelines against the vendor catalog I have seen pitched to European hotel and restaurant groups in the last eighteen months, four prohibitions are doing nearly all the operator-relevant work.

The first is the prohibition on emotion-recognition systems in the workplace under Article 5(1)(f). The plain-text scope is that AI systems used to infer the emotions of natural persons in the workplace are prohibited, with narrow medical and safety carve-outs. The February 4 guidelines made the carve-outs narrower than the vendor lobby had hoped. Workplace includes any place where employment or work is performed, which the guidelines read to include the back-of-house of a restaurant, the housekeeping floor of a hotel, and — crucially for the host-stand vendor I mentioned in the opener — public-facing areas of a hospitality venue where staff are also working. Emotion is defined to include happiness, sadness, anger, surprise, disgust, embarrassment, excitement, shame, contempt, satisfaction, and amusement. Inferring covers both real-time and post-hoc analysis. The carve-outs cover detection of fatigue states in safety-critical roles (pilots, surgeons, long-haul truck drivers) and certain medical applications. Nothing in the carve-outs preserves the host-stand “guest-mood” deployment, the housekeeping-cart “staff-stress” deployment, or the call-centre “agent-emotion” deployment that the major outsourcers have been piloting on European-language queues.

The second is the prohibition on biometric categorization to deduce or infer race, political opinions, trade union membership, religious or philosophical beliefs, sex life, or sexual orientation under Article 5(1)(g). The hospitality use case here is narrower but real — some of the loyalty-personalization vendors have built features that effectively cluster guests by demographic-adjacent categories using camera or wifi-handshake data, and the line between “demographic clustering for marketing” and “biometric categorization on prohibited grounds” is fuzzier in practice than vendor pitch decks suggest.

The third is the prohibition on AI systems that exploit vulnerabilities of a natural person due to age, disability, or specific socio-economic situation, in a way that materially distorts the person’s behavior, under Article 5(1)(b). This is the prohibition the Commission’s guidelines treated most expansively. The relevant operator deployments are AI-driven upsell systems that target elderly guests or guests in evident distress with high-margin add-ons, AI-driven minibar pricing that adjusts on traveler-stress signals, and certain forms of casino-adjacent or in-room-gambling personalization that the guidelines flagged as a worked example.

The fourth is the prohibition on subliminal techniques beyond a person’s consciousness or purposefully manipulative or deceptive techniques under Article 5(1)(a). The operator-relevant version of this is the new class of “dark-pattern” AI optimizers that have started appearing in hospitality booking flows — systems that A/B-test scarcity messaging, false-urgency timers, and emotion-laden personalization at a level the guest cannot consciously detect. The bar for proving manipulation is high, and the guidelines acknowledged that not every persuasive UX optimization is captured by the prohibition. But the deployments that are captured are captured fully — there is no carve-out for hospitality marketing.

The other four Article 5 prohibitions — on social scoring, on predictive policing based solely on profiling, on untargeted scraping of facial images for biometric databases, and on real-time remote biometric identification in publicly accessible spaces for law-enforcement purposes — are written for state-actor and adjacent deployments and bind on operators only at the edges. I am not going to walk through them line by line. The four above are where the hospitality compliance work lives.

Callout: Host-stand emotion cameras

The host-stand “guest-mood” camera deployment is the cleanest prohibited example in hospitality. Vendor pitch: a camera at the front desk reads the arriving guest’s facial expression, classifies the emotion, and prompts the agent’s screen with a script adjustment. The deployment infers emotion of natural persons in a place where employment is performed (the front desk, by the staff member). It falls squarely under Article 5(1)(f). The carve-outs do not apply. Compliance action: uninstall before August 2, 2025. The intermediate-state risk between February 2 and August 2 is non-zero through national-law channels, but the structural penalty exposure starts August 2.

Callout: AI-driven vulnerability targeting

The upsell-optimizer category that targets guest segments by inferred stress, fatigue, or socio-economic-distress signals is the deployment the February 4 guidelines flagged most expansively. The line between “personalization based on observed booking behavior” and “exploitation of a vulnerability due to socio-economic situation” is the line the guidelines drew most sharply. Operators running revenue-management AI that adjusts prices or surfaces add-ons based on inferred guest-state should run their feature inventory against Article 5(1)(b) explicitly. Compliance action: audit any AI-driven upsell or dynamic-pricing system that takes stress, distress, or socio-economic-adjacent inputs. The audit cost is real. The penalty exposure is realer.

Callout: Biometric age-verification at the bar

The casino-adjacent age-verification cameras that some venue operators have deployed are the most interesting edge case. The deployment itself — using a camera to confirm a guest is over 18 or 21 before serving alcohol — is not, by itself, prohibited. The system is permitted under the high-risk AI rules that apply from August 2, 2026, with appropriate conformity-assessment work in the interim. What is prohibited is the secondary inference layer. If the same camera that confirms age also classifies guests by gender, race, or other protected categories for marketing or service-pattern adjustment, that secondary inference is the prohibited deployment. The hardware can be retained. The software pipeline has to be partitioned. Compliance action: map the data flow from the camera through every downstream system, and confirm the only inference produced is age estimation. If the vendor’s API returns additional demographic fields, even unused, the deployment is exposed.

Callout: Predictive scheduling by personality

The HR-tech category that has been pitching “personality-based scheduling” — assigning shifts based on an AI inference of staff temperament — is the most squarely-prohibited workplace deployment I have seen pitched into European hospitality. The pitch is that the AI reads staff communication, predicts who works well together, and builds rosters accordingly. The system infers emotion or emotion-adjacent psychological states of natural persons in the workplace. Article 5(1)(f) prohibits it. The vendor talking points have begun to reframe the deployment as “scheduling-preference optimization” rather than “personality-based scheduling,” which is the kind of rebrand that does not survive a competent regulator’s read of the underlying model. Compliance action: retire. Replace with a labour-scheduling tool that takes only operator-input preferences and historical performance data without inferring psychological states. This is also the right operating decision regardless of the regulatory frame, which the piece on labour-AI integrity made the case for earlier this year.

The AI literacy obligation, briefly

Article 4 took effect on the same February 2 deadline as Article 5 and is the single most under-covered piece of the early-2025 AI Act surface. The plain-text obligation is that providers and deployers of AI systems must ensure a sufficient level of AI literacy among their staff and other persons dealing with the operation and use of those systems on their behalf.

The scope is broader than the high-risk-AI scope. The Article 4 obligation applies to every operator using any AI system in the workplace, including AI tools that are not classified as high-risk and that fall outside the Article 5 prohibitions. The training scope must account for the technical knowledge, experience, and context of the staff and the persons on whom the system is used.

The penalty regime for Article 4 specifically is softer than for Article 5 — it is not the €35M-or-7% cap. But the obligation is real, the enforcement powers around it activate on August 2, 2025, and the practical effect for operators is that every hotel and restaurant group running AI tools needs to have a documented AI literacy programme by end of Q3 at the latest. The Hogan Lovells briefing and the Littler ASAP from February both have useful operator-facing scoping on what an Article 4 programme needs to cover. The short version: training has to be substantive, role-specific, and documented, and “we sent the team a webinar link” does not satisfy the obligation.

The penalty arithmetic and what it means for a mid-size operator

The €35M-or-7%-of-global-annual-turnover penalty cap is the highest in the AI Act and is reserved for Article 5 infringements. For a hospitality group with €500M of global annual turnover, the cap is €35M. For a group with €1.5B of turnover, the cap is €105M. The 7% turnover figure is the higher of the two for any operator above €500M of turnover, and the cap scales with the business rather than with the deployment.

The operator-relevant calibration question is not whether the cap is theoretical. In the GDPR context, the cap was treated as theoretical for the first two years of enforcement and then became, in specific cases, very much not. The AI Act trajectory appears similar — a first wave of enforcement focused on the cleanest Article 5 cases (host-stand emotion cameras, personality-scheduling rollouts, unambiguous biometric categorization) at penalty levels well below the cap but well above the threshold of meaningful.

For budget owners reading this on March 19: the 2025 trade-off is no longer “compliance work now versus later.” It is compliance work between March and July at a cost knowable in advance, versus penalty exposure after August at a cost knowable only after the fact. The operators audited in the first wave will set the precedents every subsequent compliance budget gets measured against — involuntarily.

What a March-19-to-August-2 compliance sprint actually looks like

If you are running a European hospitality portfolio and you have not started, the sprint between today and August 2 has five workstreams, partitioned by what binds when. The interpretation is marked.

Workstream one (weeks 1-3): Article 5 inventory. Map every AI deployment across the property portfolio against the eight prohibited practices. Output is a register that names the deployment, the vendor, the data flow, the matched prohibition (if any), and the disposition (retain, partition, retire). The host-stand camera, the upsell optimizer, the personality scheduler, and the biometric age-verification pipeline are the four to look for first. The output is not a regulatory submission — it is the document you need on file the morning a competent authority asks.

Workstream two (weeks 2-5): Article 4 literacy programme. Stand up a role-specific AI literacy training programme covering executive, operational, customer-facing, and back-of-house staff. The training has to be substantive, documented, and refreshed at least annually. Vendor-provided generic webinars do not satisfy the obligation; operator-specific content is required.

Workstream three (weeks 3-8): Vendor contract audit. Every AI-vendor contract signed in the last 24 months needs an Article 5 representation-and-warranty review. The vendor’s product must not, on a current reading of the Commission guidelines, fall within any of the eight prohibited categories, and the contract must allow termination without penalty if it does. The standard hospitality-tech vendor contract has not historically priced this risk, and the audit will surface contracts that need to be amended. The amendment leverage is highest in the next four months; it falls steeply after August 2.

Workstream four (weeks 6-12): GPAI deployer obligations. Any general-purpose AI model used in the operator stack (the conversational AI in your guest app, the chat-based concierge, the AI-drafted marketing copy generator) generates deployer-side transparency and instructions-for-use obligations from August 2, 2025 under deadline (c). The compliance burden is moderate but it has to be scoped before the deadline, not after.

Workstream five (weeks 10-20): High-risk-AI roadmap. The deadline (d) obligations — full conformity assessments for high-risk AI deployments — apply from August 2, 2026, and the bulk of the compliance work for those deployments needs to be scoped now and executed across the next twelve months. This is the workstream the trade press has been writing about. It is the right workstream to plan for, but it is the wrong workstream to lead with. Workstreams one through three are the ones that bind on 2025; workstream five binds on 2026.

The honest read is that an operator who starts the sprint in March and runs it competently through end-of-Q3 will have a defensible compliance posture by the August 2 enforcement deadline. An operator who starts in June will not. The window for getting this done at a knowable cost is six weeks past its start date and roughly four months from its close.

Vibe Check verdict, restated for the close

The Article 5 prohibitions are a real product-line ban on a category of hospitality AI that vendors have been pitching into European hotel and restaurant groups for two years, and the Commission’s February 4 guidelines made the ban broader. The four hospitality deployments that need to come down — host-stand emotion cameras, AI-driven vulnerability targeting, demographic-inference layers on biometric age-verification, and personality-based scheduling — are not edge cases. They are the deployments the regulators wrote the prohibitions for.

The four deadlines matter independently. February 2 already passed; the prohibitions and the literacy obligation are live. August 2 is the date that converts a quiet legal exposure into an enforceable €35M-or-7% penalty regime. The same date pulls in the GPAI deployer obligations. August 2, 2026, is the high-risk-AI deadline that the trade press has been focused on. Operators who plan their compliance work around 2026 and ignore 2025 are planning for the wrong fiscal year.

The contrarian read I marked at the top of this piece — that the prohibitions are a substantive product-line ban rather than regulatory housekeeping — is what I would tell my GM at the four-star this afternoon when we finish the binder. The work to be done is knowable. The cost of getting it wrong is knowable too, in a way it was not a year ago. The window between knowable cost and unknowable penalty closes on August 2. The companion piece on the operator-stack implications of the Act is the labour-AI integrity review; the broader vendor-relationship framing is in the March 12 piece. Mark the deadlines, run the inventory, retire the prohibited deployments. The compliance posture compounds either way.

— Sofia runs Vibe Check. Tips: tips@tabletransfers.com.