How to troubleshoot chatbot issues

How to troubleshoot chatbot issues

Solve common chatbot failures to enhance your self-service customer experience
Last Updated June 16, 2022
Table of Contents

Chatbots have the potential to take your customer experience to higher levels. However, sometimes your chatbot may not perform as intended or fail to connect with users, which can lead to losing customers and valuable business. This misbehaviour can be fixed with easy, actionable solutions. In this article you’ll learn how to build a well-trained chatbot and solve common failures.

Common chatbot issues

Before you begin:

  • Make sure AI is enabled: Configuration > Artificial Intelligence > Natural Language Understanding
  • Make sure your AI model status is not Failed.
  • There must be at least 3 good performing intents in your model for the bot to behave correctly.

Bot is not starting conversations

By default, your bot won’t begin with greeting customers. To let your bot do so:

Bot does not reply to questions

  • Make sure AI is enabled.
  • Make sure NLU (Natural Language Understanding) is enabled.
  • If the bot still does not reply, try refreshing the web page.

Bot is unable to provide an answer

Sometimes your bot replies with a sentence like “”Sorry, I did not understand your question…”, this happens due to poor training, and is simply solved by checking the scores in your model:

1. Check intents’ scores:

  • Your model must have a minimum of 2 intents of the same language.
  • Add more examples. For better results use 25 distinct examples or more.
  • Validate examples accuracy; 70% or more is considered a good accuracy to know if you should add more examples for this intent.
Intents

2. Check examples’ confidence:

  • Increase confidence. This is achievable by providing a large number of examples that share the same meaning, which makes the bot confident about what customers really want when they write their inquiries. Confidence is updated each time your AI model is trained.
  • Avoid misclassification of samples. A miscalssified example is the one confused with samples in another intent. If this happens, add more similar examples to increase its confidence score, or edit the example to change a word.
Intent examples

If you want customers having the option to request to talk with a human agent when the bot fails to find the right answer, enable handover on unknown.

Bot does not understand customer’s intent

If your bot cannot identify what the customer really wants, it may be confused between two or more intents, items or categories.

  • Use Intents Confusion Report to solve confusion across intents:
    • Try merging intents.
    • Add more examples (recommended number of samples 25 samples per intent).
    • Refine data and make sure examples are unique per intent.
Intents Confusion Report
  • Enrich item description, the more text you enter the better the bot can search in it. Chatbot searches the item name, short description, and full description using common keywords.
  • Use entity annotation to teach the bot to recognize context, named entities, and key phrases in the text to narrow the categories in which the bot searches for results.

Replies are not aligning with your brand’s tone

Use the language editor to customize how your bot replies in each situation.

Never stop training

Keep optimizing; monitor your interactions for trends and good learning experiences!

Want to learn more?
Was this article helpful?