Who better to report on the Swiss Federal Railways SBB chatbot project than Laura Gander. She is Project Manager Innovation & Digitalization at the SBB Contact Center and has been involved in the current chatbot project from the very beginning.
How was the use case found?
Almost every project begins with the definition of the chatbot use case. This was also the case with SBB. The Contact Center team first analyzed the topics on which customers and potential customers have the most frequent questions. A large number of inquiries focused on questions relating to the Half-Fare travelcard or the GA travelcard. It was then decided that the chatbot should initially only answer questions about the GA travelcard and Half-Fare travelcard.
When is the use case suitable?
The use case was clear, but the team wasn’t sure from the outset whether it would work. At the time, it was still unclear what questions customers would ask the chatbot. Bots are particularly suitable for simple, predictable requests. If user questions are too individual, chatbots are quickly overwhelmed and do not deliver the desired performance.
But instead of thinking long and hard, the SBB team simply started with a chatbot for simple questions such as “How much does the GA Travelcard cost?”. After just a short time, it became clear that the majority of chatbot users really do use the bot for rather simple and easily automated questions. The team itself was a little surprised, but of course positively impressed.
Only at the beginning of Corona were there “new” questions, which could then be answered automatically after a short time.
What do I need to consider for the technical implementation?
When it comes to technical implementation, the first question is usually “With AI or without?”. What is meant here is whether the chatbot should also understand free text input from the user by means of an NLP component or whether it should guide the user through a rule-based dialog using buttons.
SBB has opted for an AI-based bot that can also understand free text input from users. Depending on the user’s request, the bot then switches to a response-based dialog in which it provides the user with various options using buttons.
The combination of free text and buttons can prevent misunderstandings with the chatbot. When a user clicks on a button, their statement is clear and there can be no misunderstandings.
. 
What was important in the conceptual design?
As soon as the technology decisions had been made, the SBB team set about writing the dialog. It was clear that the bot should sound fresh and friendly. But how much humor can an SBB chatbot have? Customer surveys have shown that SBB is perceived as serious and down-to-earth in the Swiss market. Too much humor is not welcome here. The chatbot therefore does not use emojis and addresses its users with the polite “you”.
How is the chatbot supported and further developed?
Of course, the chatbot will also be used for other topics at SBB in the future. Furthermore, it will no longer only be available in German, but will also be able to cover English, French and Italian.
A new 50% position has been created to deal exclusively with the dialog optimizations of the current chatbot. If the bot has not understood a question, it is saved in the backend and updated by the responsible employee in the bot. Provided the question matches the chatbot’s subject area.
What lessons has the team learned?
The project is far from over. The chatbot will soon also be able to speak French and in future it will also be able to answer questions on other topics.
Of course, the team has already learned a few things. The most important thing here:
- Do it! Instead of hesitating for a long time and being afraid of making mistakes, you should simply bring the bot out into the open.
- Optimization is inevitable. No matter how hard you try at the beginning, AI-based bots don’t understand all the internals at first and need to be constantly optimized.
- Customer surveys are worthwhile. Thanks to customer surveys, SBB has discovered that users want a more formal and serious chatbot.
- The chatbot is mainly used for simple questions. Although information on prices and validity is very prominent on the landing page, users ask precisely these questions again in the chatbot.
It’s best to listen to the podcast episode with Sophie Hundertmark and Laura Gander from SBB for yourself. Have fun!
And if you don’t want to miss any more podcast episodes, then sign up for the mailing list at the bottom of this page.
