Expedia learned where AI breaks. It wasn't the model.
A retired booking chatbot points hotels to the part of AI that's actually hard — the grounded data underneath it
Driving the news. Speaking at Skift's Data + AI Summit in early June, Expedia Group's chief product officer, Shilpa Ranganathan, explained why the company's first booking chatbots stalled — including Romie, the assistant it launched in 2024 and has since retired. PhocusWire's account carries the detail worth a hotel group's attention: travelers left when the bot's answers floated free of Expedia's real inventory, rates, and reservations. Not because it sounded robotic. Because it was guessing.
Where it broke. Ranganathan described a clean failure mode. Generic, ungrounded responses; customers gone within a few exchanges; trust spent and slow to win back. The repair took roughly a year, and almost none of it was model work. It was rebuilding the evaluation framework and the data plumbing underneath, so the assistant answered from live inventory instead of plausible-sounding fiction. You can't put a tower on "a foundation of sand," as she put it.
Why it lands harder for hotels. Spinning up an AI chatbot now costs almost nothing. Making one travelers trust costs whatever it takes to ground it in a coherent picture of what a guest is actually asking about — the property, the policies, what's bookable, the guest's own history with the brand. An OTA assembles that picture in one place, because it was built to. A hotel group's version lives in pieces: a PMS holding the stay, a CRM holding the guest, a booking engine holding the rates, none of them designed to answer an open-ended question in a single voice. The model is the cheap part. The grounded picture beneath it is the structural part, set by how the company is built long before anyone writes a prompt.
Read the source's stake. This is an OTA executive recasting a slow pace as patience — a framing that happens to flatter the side already holding unified inventory data. Worth the discount. The customer signal underneath it is harder to wave off: ungrounded answers broke trust quickly, and the broken trust didn't return on its own.
The organizational tax. Ranganathan put more weight on culture than code, calling AI an organizational problem before a technological one — democratizing expensive tools, managing change, building the unglamorous scaffolding that lets people across functions experiment. In multi-brand and franchise structures, where guest and rate data split across owners, operators, and the flag, that organizational barrier is the one that actually binds. The hotel that ships a chatbot first won't be the one that keeps the guest. The one that grounds it in data it controls will.
Enjoying this analysis? Hospitality.today delivers daily insights on hotel distribution, AI trends, and travel commerce — straight to your inbox. Subscribe for free at Hospitality.today →