Deterministic to probabilistic
Travelport CTO Andrew Jordan opened the discussion with the structural framing that keeps coming up. Traditional search is deterministic — you tell the system an origin, a destination, dates, and cabin. It gives you back a set of exact matches. Agentic travel is probabilistic — the input is intent, not parameters. "I want a cool holiday in Thailand." "The least risky way to recover a stranded business traveller inside policy." The system has to interpret, decompose, and act, and the risk of getting it wrong is not academic.
Almost right isn't good enough
Jordan's line: an agent that gets a trip nearly correct is not delivering hospitality — it is delivering a customer service incident. The tolerance for hallucination in a search box is not the tolerance for hallucination in a booking. That reframes what "AI in travel" has to be built to before it goes near a paying customer.
Clean, structured, deterministic APIs
Even in a probabilistic front end, the back end has to be strict. Travelport's position, aligned with the broader panel consensus: agents need clean, structured, normalised data and deterministic APIs to operate reliably. The probabilistic reasoning layer sits on top of a deterministic execution layer — not the other way around. This is where MCP, agentic protocols and standardised inventory schemas become foundational rather than optional.
The real barrier isn't what people expect
The panel — including Jason Cincotta, CEO of Kismet — pushed back on the assumption that trust or technology is the primary blocker. Both are further along than the discourse suggests. The stubborn barrier is consumer behaviour. "You could probably book travel right now agentically on some hotel websites, but that's not how people plan trips." Habits, not infrastructure, are the throttle.
Practical takeaways for hoteliers
- Assume the front end is going probabilistic; make sure the back end is deterministic enough to survive it
- Invest in data cleanliness and structured content — agents cannot rescue bad inputs
- Watch consumer behaviour, not just capability announcements — adoption will lag ability
- Design the recovery path for "almost right" — that failure mode is the one that will define trust