Hospitality

Why a Paris terrace is a bad recommendation in the rain

Three AI platforms personalise recommendations for hotel guests. None reads the weather or the street, which a luxury operator says a suggestion needs.

No product Artificial Reality identified fuses a guest's known taste with live external conditions — today's weather, tonight's local disruption. Sofiane Ghorbel, founder and CEO of AL Hospitality Group and more than twenty-five years into luxury hospitality, says that fusion is what a recommendation actually depends on. Three platforms personalise for hotel guests in production. None reads the world outside the building.

What does a luxury recommendation actually depend on?

Three inputs, in Ghorbel's account, and the recommendation is wrong if any one of them is missing. The first is taste, held at a granularity most profile systems never reach.

"They will provide us with some example of suggestions of restaurants, for example, museums, because he used to love arts — only arts, arts, arts, but not history."

Art but not history. That is the resolution at which a suggestion starts to feel personal rather than merely competent.

The second input is the weather. The third is what is happening in the city that evening.

"Depends if they have riots in this particular area ... riots sometimes happen suddenly. We don't expect [it] — after a soccer game for example, there are riots like last month, it happened in Paris. So they don't need to be in a specific area where the riots are."

Ghorbel's summary of what a system should do with all of it: "This is the strength of the AI: to get all those information and all those data from everywhere. Makes it like a cocktail and give us a great cocktail."

Why does weather invert the value of a venue?

Because for some venues the weather is the product. Ghorbel's example is a Paris terrace.

"Terraces in Paris — you cannot go to a terrace in Saint-Germain-des-Prés, Les Deux Magots for example, when it's raining. You cannot have fun on the terrace, because the fun is the terrace, it's not inside the cafe ... people want to be seen. And to be seen is on the terrace."

The same restaurant is an excellent recommendation on one evening and a poor one on the next, and nothing about the restaurant has changed. Its quality is not degraded by rain. It is inverted. A recommender that scores venues on stable attributes — rating, popularity, profile similarity — gets that wrong, and gets it wrong confidently.

What did the operator ask a system to do?

Five requirements, stated by Sofiane Ghorbel rather than inferred by Artificial Reality:

Hold granular taste — not "likes culture" but art rather than history.
Factor live weather into whether a venue is right today.
Account for real-time local conditions, including unrest and sudden disruption.
Combine every input into a single recommendation rather than presenting separate signals.
Work at the front desk too, drawing on guest history and preferences.

On the last point, recalling hotel concierge chatbots that failed to produce useful local recommendations, his position is that the missing ingredient was never the model: "If we have an AI tool for that, and we have all the history, the guest history, and the preferences, it'd be so easy to provide them with choices, the right choices, accurate choices."

Galina Fendikevich, who hosts Artificial Reality, the AI buyer-evidence podcast, extended the request to a technical shape: a real-time, multimodal view drawing on social platforms and live reviews rather than static text, deliberately narrow — political and social conditions in a city, restaurants, weather, vetted providers — combined with what the advisor already knows about the client.

What is a contextual recommendation engine?

A contextual recommendation engine combines a known individual's preferences with live external conditions — weather, availability, local events, safety — to produce a suggestion appropriate to that person, in that place, at that moment, rather than a generally highly-rated option.

The live layer is the whole difference. Conventional recommenders rank by rating, popularity or profile similarity, and those rankings are stable across the week. A contextual engine can rate a venue highly on Tuesday and exclude it on Wednesday because it is raining and the point of the venue is the terrace.

Which products come closest?

Three, identified by Artificial Reality as apparently relevant rather than named by the operator. Runnr.ai runs a proactive, WhatsApp-first AI virtual concierge, seed stage in the Netherlands. HiJiffy runs conversational AI across more than 2,600 hotels in 60-plus countries. Duve personalises recommendations per guest profile across 64-plus countries from Tel Aviv.

All three are property-side: a hotel personalising for guests inside its own walls. None appears to fuse live weather and local safety conditions into the recommendation, and none serves the advisor recommending across several cities, dozens of providers and an entire itinerary.

What should a builder do about this?

Stop treating quality as a property of the item. A large share of what makes a recommendation right is conditional on the world at that moment, and almost all recommendation infrastructure is built on the opposite assumption. Fixing that is not a hospitality feature. It is a different architecture, and it generalises well past travel.

Then build the harder half, which is restraint. Sofiane Ghorbel names knowing "when not to recommend something" among the most valuable hospitality skills, and no recommender optimises for silence — the objective function always rewards producing an answer. In a business where confidently wrong is still wrong and one bad suggestion ends a relationship built over years, a system that reliably knows when to say nothing is worth more than one that always has three options. Ship the abstention path before the ranking model.