1. The problem Bahn delivers hundreds of cars with drivers across Europe. Most orders go smoothly, but when something breaks on the road, an operator has to sort it out by hand. A ferry slot from Germany to Finland sells out, or a car stops on the Autobahn between Munich and Hamburg. The operator then has to call the driver, check ferry sites and towing companies, work out the new delivery date and decide what to tell the customer. These cases are rarer than pickup problems, but each one is urgent and expensive. Bahn told us they have no agent for this yet, and generic agents stop at "I don't know" instead of offering concrete alternatives.
2. What you built On the Road Control has two parts. First, a Control Room with an attention queue: out of all active orders, it shows only the ones that need a human, ranked P1–P3 by urgency. Second, a decision card for each case. It shows the problem, a dated delivery timeline (picked up, problem, promised and now-expected delivery), the facts with their sources, the chat with the driver, and two or three options with price and time side by side. The operator clicks Book to open the real booking page, then Approve to lock the plan, and the case leaves the queue.
3. How it works — and what's AI about it The agent works step by step with tools: it looks up the order, gathers facts, asks the driver, estimates recovery and searches for trips. It also applies real-world rules. Example: a BMW going from Leipzig to Helsinki misses its sold-out Finnlines ferry. A long land detour through Tallinn isn't possible because the driver needs rest, so the agent proposes the next Finnlines sailing plus a hotel in Travemünde. Another example: a Bentley breaks down inland in Germany. The agent proposes towing to a dealer or a workshop and doesn't suggest ferries, because the route doesn't need one. The agent talks to the driver but never contacts the paying customer, which stays with the operator. Nothing is booked without the operator's approval.
4. What's next Pilot it with Bahn operators on real on-the-road cases. Connect live ferry and towing APIs instead of demo data. Add a check for the driver's next job (will someone be there to hand over the keys if this one runs late?). Once the success rate is proven, let low-risk rebookings run automatically and keep approval for the risky ones.
