Expedia Group Shares AI Chatbot Lessons Learned

Hotel News Resource reports Expedia Group began testing AI chatbots two to three years ago and publicly launched Romie in 2024, which was later discontinued. Per Hotel News Resource, Shilpa Ranganathan, Expedia Group's chief product officer, spoke at a recent industry summit about users losing confidence when chatbot answers were not grounded in accurate, real-time inventory and reservation data. Hotel News Resource reports the company subsequently shifted to building a more robust AI framework and platform, and developed better evaluation methods and scalable architecture.

Companies developing conversational travel assistants commonly confront two technical gaps: grounding large language model outputs in authoritative, time-sensitive systems of record, and integrating low-latency inventory/reservation APIs so responses match live availability. Industry-pattern observations: firms often invest in retrieval-augmented generation, stronger provenance signals, and strict output filters to reduce hallucinations and protect customer trust.

Industry context: Travel commerce amplifies the cost of incorrect answers because availability and price are time-sensitive and tied to transactions. For practitioners, that raises engineering priorities different from generic consumer chatbots: tight coupling to booking systems, robust fallback to human agents, and deterministic verification of any booking-related assertion.

Observers should track indicators that measure real-world reliability rather than novelty, for example customer adoption rates for conversational flows, mismatch rates between chatbot claims and authoritative inventory, average time-to-confirmation for bookings initiated via agents, and whether teams publish evaluation metrics for grounding and hallucination rates. Hotel News Resource also reports Expedia describes experimentation as low-friction internally but emphasizes that not all experiments succeed.

Industry-pattern observations: organizations testing conversational AI at scale commonly iterate on evaluation tooling, invest in production-grade observability for model outputs, and embed automated checks against transactional systems. Those practices reduce user churn driven by incorrect or irrelevant answers and create a clearer path from prototype to production.