How I taught a robot to talk

Here we go

Everything runs smoothly, and even when I ask Kai some fairly difficult questions, he offers quick and precise answers. It’s time for a little challenge: “Do you know Ray Sono, Kai?” I’m very curious what he’ll say. Until this point, I have put a lot of time and passion into “my” digital service agent. 

When I joined the “chatbot” project in March, many things were still quite vague. What was certain was that a highly motivated team of UX, technology, design, content, and project management talent had one goal: to create a chatbot that answered the most important questions about Deutsche Bahn’s products, services, and offers. By the way, I’m Julia from the Ray Sono content team. I advise my customers on topics related to editorial content in various digitisation projects. I’m also passionate about creative writing, so I am immediately excited to help design a chatbot that must answer a wide range of many, many users.

“Conversational” is currently one of the buzzwords in the content sector. It’s no wonder. This touchpoint is more concrete, personal, emotional, and closer to the customer than anything else. But alongside opportunities, it also carries risks. If a chatbot does not meet the needs of customers, it can annoy them. The challenge is to create an experience that creates value for users and businesses alike. A challenge we are happy to accept. 

In this project, I took over all “content” aspects. This is a big part of a chatbot. After all, the bot stands and falls with its answers and must be measured by what it answers the user. 

Chatting – the basics 

I started as always – with a thorough analysis to deepen my knowledge of chatbots. In addition to scanning relevant technical literature and the latest blog posts on service chatbots, this meant deep immersion in the world of automated dialogue – and of course chatting. The best thing was chatting with many bots about all kinds of topics. It’s good there are now so many of them. One of my favourites is the Woebot, developed by a Stanford University psychologist, which helps users who have depression. It’s a super-exciting and useful case that is set up interactively and has a very pleasant bot. 

In general, I quickly realised that all the chatbots I liked to interact with have recognition value, can tell me something about themselves, and can chat with me. So that Kai will be the same, my UX colleague Carolina and I defined his personality. He had to harmonise with the Deutsche Bahn brand and fit his role as a service employee that offers customers advice and support. This is exciting, because it will obviously influence the bot’s appearance as well as all his dialogue texts. 

Off to bot school

Which brings us to our keyword. Now we must give Kai a “voice”. But how do you write texts for a chatbot? Texts must be easy to understand, friendly, adapted to the bot’s personality, as well as short and precise – i.e. clear from the start. However, the more extensive and complex the content is, the more difficult this becomes. The challenge is to transfer the entire Deutsche Bahn FAQ section to the chatbot. To prepare Kai to talk with real users, he must be able to make small talk with callers, in addition to his service job. He must also respond to other requests such as emergencies, or to praise or verbal abuse. 

The biggest difference from “normal” texts is that most texts are written to answer specific questions. But when customers talk to a chatbot, they don’t always formulate their questions the same way, even if these questions mean the same. Therefore, as a copywriter, you must consider what questions users could ask, and then find answers that cover these questions – but not too specifically. By the way, we are talking about 1,000 question-answer pairs and about 50 small talk modules. Thanks to a pilot project, real users are already providing us with many questions, and these help us train our bot. Nevertheless, in these weeks we are greatly deepening our knowledge about Deutsche Bahn as we write answers and questions and build dialogue trees to make Kai as smart as possible.

Kai speaks!

After everything has been implemented in the system, we go back to the chat. We talk to Kai day after day and refine the content and expressions. We realise how important it is also to have answers to questions that a chatbot cannot or should not answer. Nothing is more annoying than a bot who says he cannot understand you. 

Today I can say that the effort was worth it. Kai is now talking to all of us and freeing up the train company’s customer-service lines. Our customer and project manager Frank Merkel from DB Sales is also visibly happy: