AI in hospitality — what's worth it, and what's just hype

Every POS pitch in 2026 promises “AI-powered”. Read the homepages — Toast, Square, Lightspeed, every Shopify-for-hospitality entrant — and you will find a fairly consistent vocabulary: AI-powered menus, AI-powered ordering, AI-powered pricing, AI-powered insights, AI-powered concierge. Some of that is real product. A great deal of it is wallpaper.

This post is the case for being narrow and specific about where machine learning actually earns its keep in hospitality software, and where the marketing copy is doing work that the product itself is not.

The signal vs the noise

Gartner’s Hype Cycle for Emerging Technologies is a useful framing here, however over-familiar. The pattern it describes — peak inflated expectations, trough of disillusionment, plateau of productivity — applies almost line-for-line to AI in hospitality between 2023 and now. The peak was around mid-2024 when every restaurant-tech founder was demoing a chatbot at the table. The trough came faster than expected, somewhere in 2025 when the Air Canada chatbot misinformation case and a handful of similar incidents made operators wary.

We are now somewhere on the upslope toward the plateau. Which is the most interesting moment, because it is the point at which the genuine use cases get separated from the genuine hype.

McKinsey’s 2024 review of generative AI in operations lands on a similar conclusion across industries: ROI is high for narrow, bounded, evaluable tasks; ROI is unreliable-to-negative for open-ended, generative, unconstrained tasks. Hospitality is not exempt from that pattern.

Where AI actually helps in hospitality

Machine translation of menus. The single highest-ROI AI use case in our product. A venue in Soho serves customers in eight languages; a venue in Madrid in five. Machine translation, with a careful human-verification gate, gets a usable menu in front of a customer who could not otherwise read it. The translation is not perfect — see our post on where machine translation gets it wrong — but the failure mode is “occasionally awkward phrasing”, not “wrong dish on the bill”.

Google Cloud Translation and DeepL are both production-mature; the marginal quality gap matters less than the platform you can build a verification gate around.

Fraud and anomaly detection on transactions. Stripe Radar, Adyen RevenueProtect and the other major card processors have shipped real, evaluable, narrowly-scoped AI in this space for years now. Stripe Radar’s documentation sets out exactly what features the model trains on and what the false-positive trade-off looks like. This is genuine AI doing genuine work, with a clear evaluation metric and a clear cost of error.

Demand forecasting for stocking. Reservations + weather + day-of-week + past behaviour predicts how many covers a venue will do tonight, which feeds into how much of which ingredient to prep. Toast’s Smart Reports and similar products do this. The technique itself is decades old (it is the same regression problem that retailers have been doing since the 1990s); the modern improvement is mostly the data plumbing.

Receipt and invoice parsing. OCR + structured extraction has become genuinely good. Amazon Textract and Google Document AI both produce production-quality results on hospitality-style documents.

In all four of these cases, the pattern is the same: a narrow, bounded task with a clear right-answer, evaluated against a clear metric, where being wrong has a bounded cost.

Where AI is mostly hype in hospitality

Conversational ordering. “Tell our AI what you want and it will build your order.” Almost every demo of this we have seen at industry conferences degrades within a few exchanges. The failure mode is usually that the model is given a menu of 150 items and is asked to map the customer’s natural-language request onto that menu and handle modifiers, allergens, and pricing. Each of those is a structured problem the customer can already solve with three taps. We have decided not to ship this; we may revisit when retrieval-augmented techniques are dramatically better. Anthropic’s own writing on the limits of conversational agents is honest about what these systems still get wrong.

Generative menu descriptions. Cute as a demo. The marketing copy effectively becomes whatever the model thinks is plausible — which is often plausible-but-wrong, which is the worst failure mode in food labelling. Two restaurant-chain pilots that we are aware of (we will not name them publicly) walked this back after legal review.

“AI-driven dynamic pricing.” Possible in theory. In practice, requires a level of data and a sophistication of forecasting that the typical 80-cover venue does not have, and creates customer-trust problems if the same drink costs different amounts at different times. Surge pricing in food and drink is a reputational hand grenade, as Wendy’s discovered. We do not offer this.

“AI concierge.” Usually a thinly-disguised chatbot pointed at a help centre. Where it works, it is because the help centre underneath is well-written. Where it fails, it is because the chatbot hallucinates a refund policy that does not exist.

Our rule, plainly stated

Use AI for narrow, bounded, evaluable tasks where being wrong has a bounded cost. Avoid AI for open-ended, generative, unevaluable tasks where being wrong is unbounded — particularly anything that touches a customer’s wallet, a customer’s allergies, or a venue’s pricing.

This is roughly the position Cal Newport reaches in his ongoing writing on AI productivity, and roughly what DHH argues in his less-polite manner. It is also a position that is increasingly visible in serious operations research once the demo gloss is stripped away.

What ORDR does about this

ORDR ships machine translation for menus with a human-verification gate, uses Stripe Radar for card-fraud signals, and is honest about the AI features it does not ship. We will add new AI features when they pass the narrow-bounded-evaluable test. We will not add them because the homepages of competitors do.