By Marcus Bennett

A personalization package that costs roughly €30 to assemble could command €300 from a luxury guest — and €3,000 in the right setting. That spread is the core argument Gilad Berenstein makes for AI in hotels, and it rests on data most operators already own.

Berenstein founded Utrip, an AI trip-planning company that served more than 200 million travelers before its sale. He now invests in hospitality technology through Brook Bay Capital. He laid out his position on the podcast Matt Talks Hospitality, in conversation with Mews CEO Matt Welle.

His claim is narrow. AI has little guest-facing impact today. Few guests check out believing software improved their stay. The measurable gains sit behind the scenes, in operations, maintenance and guest intelligence.

Where the systems already pay

Two properties anchor his case. Shashi Hotel in San Francisco, a new build, placed sensors in its walls to track noise, heat, dampness and water. A leak in one bathroom wall is caught before it spreads through the tower. Berenstein said the payoff has come quickly through property insurance.

At Atlantis The Royal in Dubai, staff carry location beacons for security. The hotel reuses that data to tell guests roughly when a request will arrive. The delivery time is not the point. The knowing is.

Neither example required replacing a person. Both required connecting a system that already existed to a guest-facing use. That is the pattern Berenstein recommends hotels copy.

What does the training data cost?

Nothing, in most cases. Berenstein points to the concierge inbox as the ready-made asset. Every guest email asking what to do in Rome or Venice, and every hand-written reply, forms a training set for a concierge agent. Add an agent that predicts no-shows and another that chases commissions, and the knowledge compounds once the agents talk to each other.

He calls this the ground level of AI swarms: simple agents that grow smarter in combination. The fully connected version — a guest's own agent negotiating dinner with the hotel's concierge agent — is, in his words, "an end state that does not really exist yet." His estimate for arrival: at most a couple of years, probably months. Treat that timeline with caution. Vendor timelines in this sector have slipped before.

What should operators build?

Not the commodity functions. "Whatever your team is doing over and over and over again right now clearly is not accounted for by your system," Berenstein said. Repetitive work marks the starting line. Beyond that, he draws a hard boundary:

• Accounting works the same at nearly every hotel. Not a differentiator.
• Wi-Fi login experience wins no bookings. "No customer's ever going to choose your hotel because you have a better Wi-Fi login experience than someone else."
• A wellness hotel or a sustainability-led property should build its own tech, because that sharpens the reason guests book.

The chatbot itself is no longer the moat. Anyone can spin one up. The moat is proprietary data and a distinctive product.

The chocolate-versus-wine math

Berenstein does not drink much. He likes chocolate. Swap the standard welcome bottle of wine — $30 to $40 even at wholesale — for an $8 bar of local chocolate, and the property spends under 20 percent of the original cost while the guest feels more seen, not less.

Scaled to a family trip, his Paris example runs about €30 in hard costs: a princess-themed itinerary video, toys by the bathtub, a guide armed with fairy tales, gluten-free cookies for his wife. What he would pay: "I think it's almost certainly €300. In a luxury setting, it might be €3,000." The margin, not the gesture, is the story.

Guest intelligence tools such as Abra extend the same logic. A bartender logs that a guest loves mangoes. Housekeeping gets prompted the next morning to swap the strawberries. The software holds the memory. The person makes the moment.

What stays human?

Berenstein's labor position is explicit. "I don't want fewer humans working in your hotel. I want fewer humans locked in the back hallway, in the back office, in the laundry room," he said. And on the concierge question: "I'm not suggesting that you replace your human concierge. I'm suggesting that you superpower the human concierge."

He described a ten-minute check-in at a high-end Asian hotel that was purely transactional. Late at night, automation should have handled it. At any other hour, a team member could have shown him the full spa floor. That is the trade he wants operators to make: automate the transaction, redeploy the labor.

His closing frame sets the boundary. "AI is coming for everything except for those things that we choose to keep in the human domain." Operators who define that domain deliberately will keep the margin gains and the staff. Those who do not will automate by default and discover what they gave up only in the reviews.