A truck does not wait. When a driver needs a wash bay, a parking spot or a repair on a route across Europe, TRAVIS is the account that finds it, books it and pays for it. Support sits between the driver on the road and the fleet office paying the bill, and both of them are in a hurry.
The knowledge was in people, not in documents
TRAVIS had years of good support answers and almost none of them written down. The team knew which fuel-card validation emails were routine, which parking questions needed a location look-up, and which requests were really a billing dispute wearing a different hat. None of that lived anywhere an AI agent could read.
The team also works across languages. The chat used to open by asking the customer to pick one, purely so the conversation could be routed to a colleague who spoke it — a question no customer wants to answer before asking their actual question.
And TRAVIS had two hard rules going in: never share data that should not be shared, and never send a confident answer that is wrong. Their head of development was explicit about not handing write access to anything until both had been proven.
What Open does
For the first thirty days the agent watched. It read what customers asked and what the human agents answered, and wrote the pattern down — around 200 unique instructions, generated from the team’s own behaviour rather than from a document nobody had written. The support lead reviews them in Companion and publishes or unpublishes each one, so a human mistake never becomes a permanent lesson.
Those instructions call real actions. A question about where a driver can stop is answered from a live locations look-up, not from a general description of the network, and the DKV fuel-card validation emails — the ones whose subject line is always some version of "validation request, driver added a new main card" — are answered on autopilot as their own topic.
Every reply can be opened up. The reasoning view shows the steps the agent took, the instruction page it followed and the API call it made, which is how TRAVIS satisfied itself that an answer was safe before letting it go out.
The escalation rule is TRAVIS’s, not ours: if a request would change something in the database — cancel a booking, adjust a transaction, unlink a fuel card — it goes to a person. The agent reads; people write. Open mapped which of those cases could be automated without write access and left the rest on the table until TRAVIS is ready.
The Impact
The language question is on its way out of the chat. The agent fills the ticket’s language field itself at handoff or closure, so a conversation reaches the right colleague without the customer being asked to sort themselves first.
TRAVIS is not publishing an automation rate yet. The pilot was scoped to get the knowledge written down and to prove the agent could be trusted with read access before anything else, and this story will carry resolution figures when TRAVIS has them and is ready to share them.