AI Quoted a $1,500 Seaplane Fare; the Resort Charged $1,910
Lumina's audit of 31 Maldives resorts found 38 AI errors, most from wrong source choices — including a $410 seaplane fare gap traced to a 2015 guest review.

A $410 gap sits at the center of new research into how hotel AI answers go wrong. Lumina, a Zurich company building AI marketing teams for hotels, ran a check in which its AI quoted a seaplane transfer to a Maldives resort at $1,500. The resort's own live page listed $1,910. The AI hadn't invented the number — it had pulled it from a guest review written in 2015.
The finding comes from Lumina's audit of AI-written fact sheets covering 31 Maldives resorts, which surfaced 38 errors in total. The most common failure wasn't fabrication. It was source selection: in 18 of those 38 cases, the AI picked the wrong source when several disagreed. And the wrong source often looked credible. The quoted figure could sit right there on the hotel's own website, ready to pass an editor's simplest open-the-link test.
Recency was the second trap. In another case, a fact sheet quoted a seaplane fare of $488 return. The page did give that fare — for travel from January 2023, with no current price listed. A traveler comparing two resorts on the AI's numbers would have budgeted hundreds of dollars short for the journey to one of them.
What does this change for hotel operators?
Marco Di Cesare, Founder & CEO of Lumina Frontier, frames the operational fix as a reservations question: can the team see what a guest will find about the property today, including in old PDFs? And when two official pages disagree, who has the authority to settle the answer before an agent sends it out under the hotel's name?
Lumina learned the hard way. The company had published 119 pages claiming a hotel reseller didn't tell buyers which room they were purchasing. The claim came from Google Hotels, which showed the price but omitted the room name. One more click — to the reseller's own site — would have shown the room named. Lumina fixed all 119 pages and rewrote its agent instructions: any claim about what a seller tells a guest now requires visiting the page a guest would actually see. "Check your facts" had sounded clear enough, Di Cesare writes. It took 119 pages to realize it left out where to look.
Conflicting pages are a group-level problem
The research points to a concrete case: a Tuscan resort whose house rules listed the dog fee as €35 per stay while its FAQ said €35 per night. If the property confirms the per-stay figure, someone on site can answer the guest and correct the FAQ. Group marketing has the broader job — giving every hotel in the portfolio the same instruction for the next conflict: keep both passages, flag the difference and route it to the person at that property who sets policy.
Di Cesare's suggested test for multi-property groups: take the last correction reservations sent, open the pages a guest could still find, and check whether any repeat the old answer. Then have a marketer at another property draft a reply to the same enquiry using only their own hotel's sources. An unfinished reply counts as success, provided it lands on the right desk with both conflicting passages attached.
As hotels hand guest-facing answers to AI systems, the audit suggests the near-term risk isn't hallucination — it's stale, misplaced sourcing that looks perfectly verified until someone checks the date.
More from Daniel Okafor
Show full bio
Correspondent covering companies, deals and strategy in Restaurant and hotel operations at The Pass Brief.
761 articles


