Those of you who have ever written a chatbot have probably encountered several error messages, such as "Sorry, I don't understand your questions", that the chatbot answers. You may have even experienced your homegrown chatbot receiving extremely random questions from users that are clearly just testing the chatbot boundaries. These irrelevant requests are also called conversational divergences , which cannot be prevented.

Usually, a chatbot is created to solve a specific use case or provide a specific value. Therefore, it is obviously not able to answer all possible questions. So if the chatbot sends out a so-called error message (also known as a fallback response or default response ), this basically indicates that it was not able to process the request and link it to a suitable intention. Considering how new the technology is and how fast AI will still evolve, this is definitely not a criterion for not giving a bot a chance. In fact, if you know how to deal with the user in such situations, there is a high chance of ensuring a positive user experience in communication. On the other hand, there is also the danger of damaging the user experience by sending ill-conceived error messages. One of the main tasks of the chatbot is to keep the user engaged in the communication without leading them into dead ends.

In this blog post, you will learn how to write user-friendly error messages by demonstrating bad and good examples and presenting a standard framework on how to design error messages to avoid users getting lost in the conversation again.

How does error handling actually work?

How not to do it:

Example of bad chatbot error messages

The above examples show very well how error messages should not be handled.

  • The conversation on the left with "The Weather Channel" did not provide a user-friendly answer because it only refers to the error without giving details about the cause. Also, asking the user to "try again" is not really helpful, as they do not know what exactly went wrong and what they should do better next time.
  • The chat with "Dallas Mavericks" in the middle actually shows many improvements. There are definitely more ways to communicate than asking the user to use cryptic input. What stands out, however, is the negative undertone of "Hmm, sorry, I don't quite understand this". This error message simply gives the user the impression that the chatbot is not trying very hard to please or help them.
  • The last example on the right fails to find a helpful wording for the default message. It is obvious that the user will leave the chat because the chatbot does not even work, as it always sends the same answer. Speaking of which, if you have already written user-friendly error messages, it is still important to implement variations of them! The likelihood of misunderstandings occurring more than once in a message is very high. So you don't want to annoy your users by sending the same messages.

Let's take a look at experts: Poncho's Case

The developers of the most popular chatbot Poncho, which can send weather forecasts, have also realized how important it is to write well thought-out fallback answers. For example, they implemented the following error message at the beginning:

"Sorry, I was trying to charge my phone. What were you trying to say? "

So what exactly went wrong? In this message, the chatbot is obviously overplaying the fact that it did not understand the request. This unfortunately only leads to an infinite circle, as the user will repeat his question, which will be followed by an error message. The person has no idea how to deal with the chatbot further and will soon leave the conversation. This is how the developers of Poncho came up with a better answer:

"Well, I'm good at talking about the weather. Other things, I'm not so good at. If you need help, just type in "help".

In this version, the chatbot is straightforward. He explains that its sole purpose is to talk about the weather and that the user should not expect it to know more. He also mentions the help section to provide even more assistance. The people behind Poncho realised that this was a better strategy to address users in these situations. To perfect error messages, we will now outline a framework.

A framework for error handling

1. clarification of the misunderstanding

The first step is to clear up the misunderstanding. It is important to be very transparent, honest and modest about the abilities of the chatbot. Errors should not be overridden. If the chatbot is not able to understand the request, this should also be communicated.

2. remind the user of the capabilities of the chatbot

In the second step, it makes sense to explain to the user once again what information the chatbot can provide. So if you remind the user of the knowledge, function and skills of the bot, he will be ready to ask the right questions.

3rd Call to Action!

Calling actions is the right way to guide the user. You can either suggest him to use buttons, ask his question in a different way, or enter a keyword like "help" to re-read the documentation. It might also be useful to offer the user a handover to a person to ensure satisfaction or sometimes even restart the whole conversation. There are several options to keep the user in conversation to keep the termination rate lower.

Good examples

conversational UX

Above you will find better versions of the standard answers. In the Porsche chat on the left, the user received a clear call to actions, either to a real employee to ask questions or to use the buttons below. Same case with ShoeDazzle. The chatbot is aware of his abilities and tries to lead the user out of the impasse.

To summarise this post, it is extremely important to focus on the critical parts of a conversation to prevent the user from getting lost. Error messages are crucial events. Keeping the aforementioned framework in mind during the conversational design of a chatbot will ensure a great user experience. Finally, keep the following in mind:

Rule no. 1: The user is never wrong and it's never his fault! 😉

Want to know more about training voice and chatbots? Then get our checklist with 20 tips for optimising virtual assistants: