Hospitality

Why AI personalization fails for luxury hotel guests

The privacy paradox, in Sofiane Ghorbel's words: the better the personalization, the more data AI needs from the clients least willing to give it.

AI personalization fails for luxury hotel guests because personalization needs data and luxury clients are the population least willing to give it. Sofiane Ghorbel, founder of AL Hospitality Group and a hospitality executive with more than twenty-five years across luxury hotels, resorts, travel and destination management, calls it the privacy paradox. The better the personalization, the more data the system requires.

Everything else that goes wrong follows from that.

What is the privacy paradox in luxury hospitality?

It is the collision between what personalization needs and what the client bought. Ghorbel names it directly:

"We have the privacy paradox. The better the personalization, the more data AI requires."

"Luxury clients are often the least willing ... to share data. They value privacy, discretion, security and control. And the challenge becomes: how do we deliver personalization without compromising confidentiality?"

This is worth separating from ordinary enterprise privacy concern. It is not that the data is sensitive and must be handled carefully. It is that the client's expectation of discretion is part of what they purchased. A system that makes a guest feel surveilled has damaged the product even if it never leaks anything.

Why is a wrong AI recommendation worse in this business?

Because the error is relational, not operational. In most software a mistake costs a retry. Here it costs the relationship, and the relationship was the asset.

"Trust is the product in our field ... In luxury hospitality we don't sell rooms. We don't sell cruises, we don't sell flights. We sell trust."

"One incorrect recommendation can damage a relationship [that] took years to build. So an AI mistake is not simply operational, it's relational."

That inverts the usual accuracy trade-off. An assistant that is right eighty percent of the time is useful in most contexts. In luxury hospitality the other twenty percent is not annoying, it is disqualifying, because the failures are visible to the client at exactly the moment they were trusting your judgment.

Why does AI sound so certain when it is wrong?

Because thin data produces confident output rather than an admission of ignorance. Ghorbel connects the two directly:

"Many AI systems appear very confident, but confidence is not accuracy. And confidently wrong is still wrong."

"When we lack data about a specific point, people are thinking that the information is correct, it's good, it's confident, people are confident about that — but it's not accurate."

This is the same failure Artificial Reality, the AI buyer-evidence podcast hosted by Galina Fendikevich, has documented in other physical-world industries. In construction, a model answered a question it lacked the site conditions to answer. In yacht charter, a model told a client he could anchor in twenty-five knots of wind. In luxury hospitality the missing input is the client themselves, and because these clients withhold data by design, the gap is permanent rather than a collection problem somebody can fix.

Did a hotel concierge chatbot actually fail?

One reported case, and the provenance matters. On the show, Fendikevich cited a case study she had read about Marriott: the group deployed a chatbot intended to substitute for concierge recommendations, and a reporter testing it asked the most basic possible question, where to eat nearby, late, on arrival. The system could not produce useful local recommendations because nobody had given it the local restaurant knowledge. The group redirected the technology to revenue management, demand forecasting and back-office operations, where it worked. That account is reported as told on the show and has not been independently verified by Artificial Reality.

Taken at face value, it illustrates the point. The chatbot did not fail because the language model was weak. It failed because the specific local knowledge that makes a concierge valuable had never been captured, and confident output over missing data is worse than no output at all.

What should a luxury operator do about it?

Separate the two data problems, because only one of them is yours to solve.

Consent. Client preference data exists and is deliberately withheld. Ghorbel's own ask is privacy-safe personalization in which clients control their own preference profiles, granting an advisor access rather than handing over a copy. Nothing you do internally changes a client's willingness to share.
Capture. Operator knowledge exists and was never structured. It lives in relationships, private networks, human expertise and local market knowledge, and sometimes on paper. Which hotel suits which kind of family, which concierge in Rome actually delivers, which terrace is worth it and when.

The first is a technology bet you are waiting on someone else to win. The second is available now, is entirely inside your control, and is the harder asset for a competitor to replicate. It is also the exact thing the failed chatbot was missing, and the thing that walks out of the building when a senior advisor retires.

So the question to put to your own team this month: if your longest-serving advisor left in ninety days, what would you have written down by then?