- An AI travel agent confirmed a booking at the Tabacon Springs Eco-Lodge in Costa Rica. Gorgeous photos. Great price. Confirmation number.
The hotel does not exist. The model fused two real resorts into a fictional third.
In 2026, every major booking layer is shipping this. 🧵 - The pressure is real. Sabre + PayPal + Mindtrip ship agentic flight booking in Q2 2026. Google AI Mode books Marriott direct, bypassing OTAs. Amadeus Cytric Easy lives in Microsoft Teams.
- Navan reports 73% touchless and policy violations down 35%→<5%. Your CFO wants that. Now.
- And here's the gap nobody mentions in the board deck. GPT-4 scores 0.6% on TravelPlanner, the standard travel-planning benchmark. Not 60%. Zero point six. We're bolting live bookings onto models that fail the task 99%+ of the time.
- This is not a quality bug to iterate on. It's a liability.
Feb 2024: a tribunal ordered Air Canada to pay $812.02 after its chatbot invented a bereavement fare policy. The airline argued the bot was "a separate legal entity." The tribunal rejected that flat. - And the "blame the LLM vendor" exit is closing. In Jan 2026 a Chinese court narrowed vendor liability — but did NOT shield deployers. The duty of care stays with the travel brand whose surface the answer appeared on. You own every word your agent says.
- Why does the wrapper hallucinate? Because chaining LLM steps is multiplicative. At 90% per step, a 10-step booking flow lands at 0.9^10 ≈ 34% success.
Booking is control flow, not a language task. Prompt chains can't hold a state machine together. - Even if the model were perfect, the plumbing isn't. NDC sells the offer, but exchanges, refunds and IROPS rebooking still run on GDS. Treat GDS vs NDC as binary and your agent strands travelers when a flight cancels — IROPS runs $500–$2,000 per traveler. Both pipes, or neither.
- Then the economics bite. GDS charges ~$3–3.50 per segment plus ~10% commission — on searches, not just bookings. An agent that searches 20 times to book once is lighting money on fire. Lufthansa hiked GDS fees again Jan 1, 2026. No caching = a surprise invoice.
- And the paperwork is real. To issue tickets you need ARC (US, ~25 days) or full IATA accreditation (6–12 months). Then EU AI Act transparency rules for deployers land Aug 2, 2026 — the agent that hallucinates is now a compliance event.
- So we build the opposite of a wrapper. A deterministic state machine: the LLM only extracts language; hard-coded logic makes every GDS/NDC call, policy check and payment handoff. The LLM never touches the error path — if one fails, a Saga rolls it back. Nothing ships unverified.
- Question for TMC and OTA teams: ship the platform's vendor wrapper and accept the lock-in, or build vendor-neutral and own the GDS+NDC complexity yourself?
For most teams there's no middle. Which way are you leaning, and what's the dealbreaker?
#TravelTech #AgenticAI - We wrote up the full deployment posture — NDC servicing gap, look-to-book penalties, EU AI Act, the lot — for TMCs and OTAs who can't bet the company on one platform: https://veriprajna.com/solutions/agentic-travel-booking