A booking agent that's 90% reliable at each step books a correct trip just 34% of the time. Ten sequential steps, 0.9^10. The math of chaining LLM calls is unforgiving. That's not a tuning problem, it's structural — and it's why the OSU NLP Group's TravelPlanner benchmark clocked GPT-4 completing real multi-day itineraries at 0.6%. For a TMC or OTA, that gap is a line item. A family asks for an eco-lodge under $200; the model blends two real Costa Rica resorts into a fictional "Tabacon Springs," generates a gorgeous confirmation, and the family flies into a property that doesn't exist. Fluent is not the same as correct. And correctness isn't the only cost. GDS providers bill per search, not just per booking — so an agent firing four exploratory searches a turn burns through an OTA's 3–5% merchant margin in a quarter. It's the single most overlooked number in agentic travel pitches. The liability already has a precedent. In Feb 2024 the BC Civil Resolution Tribunal ordered Air Canada to pay $812.02 after its chatbot invented a bereavement fare policy — and rejected the argument that the bot was a separate legal entity. You own every statement on your surface, page or model. The fix isn't a better prompt. It's removing the LLM from the control flow: let the model read intent and summarize, and let deterministic code run every GDS/NDC call, every policy check, every payment handoff. That same code closes the NDC servicing gap — the exchanges, refunds and IROPS rebooking that still fall back to GDS rails after an NDC booking. State machine over prompt chain. Verification loop before anything reads "confirmed." We work with travel management companies and OTAs that can't bet the company on one vendor's agentic stack — and need the booking to be right, not just fluent. Save this for the next "what's our AI travel strategy" meeting. #AgenticAI #TravelTech #TMC #AITravelBooking #AIGovernance