The 2% problem: why nobody lets AI press "book"
91% of digital natives plan trips with AI. Almost none of them let it pay. We spent six months figuring out what closes that gap — and it was not a better model.
Every traveler we interviewed had the same story. They asked a chatbot to plan four days in Lisbon, got a decent itinerary back, and then opened four other tabs to book it themselves. The plan was free. The commitment was not.
The gap is authority, not accuracy
When we replayed those sessions, the recommendations were usually fine. What was missing was any way to see why a flight was picked, what it would actually cost after fees, and what would happen if the agent got it wrong. People do not withhold their card because the model is dumb. They withhold it because nothing in the interface is accountable.
What we changed
Three things. Every recommendation carries its reasoning trace inline, in plain language. Every price is quoted end to end before approval, with no post-hoc fees. And spend authority lives in an Identity Wallet the traveler controls, with limits and provider allow-lists they set themselves.
The result
In our closed test group, approval rates on agent-proposed bookings went from a third to just over four in five once the reasoning trace was visible. The model did not change. The contract between agent and traveler did.
Trust is not a model capability. It is an interface promise you keep every single time.