In a rail operator's customer service much is repetitive: timetable enquiries, questions about tickets and refunds, connections, luggage. The same answers again and again and in between, the cases that genuinely require experience.
What we delivered
A sovereign chatbot that talks directly to customers and answers around 90 per cent of enquiries on its own. It knows the company's information, replies in its tone of voice and hands over cleanly to a person when a question goes beyond that.
Two properties were non-negotiable:
- Sovereign operation. The language model runs in a controlled environment, not on an arbitrary platform. Availability does not depend on a provider's business model.
- Data-protection compliant. Enquiries stay within the agreed boundaries; nothing flows into third-party model training. For a company handling large volumes of customer data that is a prerequisite, not a bonus.
The real return
The figure sounds like a saving, but it works mainly as a shift: the team no longer answers the same question for the twentieth time, but handles the ten per cent where judgement matters, complaints, special cases, upset customers. That is exactly where a transport company's reputation is decided.
As a side effect, service quality becomes predictable: the bot answers at three in the morning as it does on Monday morning, and waiting times for difficult cases fall because the simple ones are no longer in the same queue.
Our assessment
Ninety per cent is not achieved with a bot that does everything a little, but with one that does a clearly defined field very well and that knows when to hand over. Handing over to a person is not an admission of failure but part of the design.