It is easy to build an impressive AI demo and surprisingly hard to build an AI product that a travel business will trust with its customers. These are the working principles our product team has arrived at — mostly by getting things wrong first.
Start from the call, not the model
Before any feature is scoped someone on the team listens to a hundred real calls. The problems that matter are rarely the ones the technology makes easy; they are the awkward handoffs, the misheard references and the policy questions nobody wrote down.
Make the boring things reliable
Transfers that work every time, confirmations that always arrive, numbers read back correctly — none of this is exciting and all of it decides whether a customer keeps using the product. We ship reliability improvements ahead of new capabilities, every release.
Honesty is a feature
The assistant says what it cannot do, identifies itself as an AI where the law or the customer requires it, and never invents a policy. Trust built this way survives the occasional mistake; trust built on pretending does not.
Ship, listen, iterate
Weekly releases, a QA loop in which customers flag transcripts directly from their dashboard, and a roadmap that changes when the data says it should. We measure success in calls completed well, not in features announced.
The product is not the model. The product is what happens on the call.
