
Two tourists hiked 4,000 meters into the Peruvian Andes in 2025 looking for the Sacred Canyon of Humantay. An AI trip planner had invented it. The canyon doesn't exist.
The unsettling part isn't that the AI lied. It didn't. It blended real places into a plausible, fictional one — convincing enough that two people booked a trek around it.
For a travel company, that's not a quality bug. It's a liability. In Feb 2024 a tribunal ordered Air Canada to pay a customer after its chatbot invented a bereavement fare policy. The airline argued the bot was a separate legal entity. The court rejected that: you own every word on your surfaces, model-generated or not.
Here's the math most vendor demos skip: a real flight booking is 10 sequential steps. At 90% reliability each, end-to-end success is 0.9^10 — about 34%. On the TravelPlanner benchmark, GPT-4 completed multi-day itineraries 0.6% of the time. You can't prompt your way out of compounding stochastic failure. And every speculative search has a price — an agent firing four exploratory GDS queries per turn can burn an OTA's entire 3-5% margin in a quarter.
The fix isn't a better prompt. It's taking the LLM out of the control flow — let it read and summarize, but let deterministic code make every inventory call, policy check, and payment handoff.
If you run booking at a TMC or OTA: where has a confident-but-wrong AI answer cost you the most — inventory, policy, or a refund it never should have promised?
#AgenticAI #TravelTech