This article covers common AI Assistant issues and offers quick tips on how to tackle them, helping you get more accurate and helpful responses.
We have divided the tips and instructions in the following three categories:
Conversation Flow: how the chatbot carries the conversation and handles escalation rules.
Content Relevance and Accuracy: how well the chatbot understands questions and how it answers with the right content.
Response format: how the chatbot presents information, in terms of for example language and formatting.
Conversation Flow
The chat is closed when only the AI Assistant is available and humans are offline
AI Assistants have access to chats only in rooms that have been assigned to them in the AI Persona's settings in the AI Chatbot app. For the AI to chat in these rooms by itself, it also needs to be added to the router for the room. To check or edit the router, please visit our Routers-guide.
If the room has been assigned to the AI Assistant in the Chatbot app and it has been added to the router, the final step is to check that the "Hide chat if only bots are online" in room settings (Settings → Rooms → Choose room → Chat settings) is not selected.
My AI Assistant escalates conversations to human agents too easily
We recommend to preferably build the transfer logic in your Q&As - you can even have a separate Q&A just for questions that should always be escalated to humans. The confidence level threshold for the Q&As can be fairly low, i.e. 30-40%. If you want to use Fallback-answers to escalate chats to humans, use very specific instructions on when a transfer should happen. Try to tell the AI what to do, instead of what not to do.
My AI Assistant is not escalating conversations to agents when it should
First, look up the conversation in the "Chat log" tab and see where (which Q&A for example) the question matched. If there is a "Transfer to human" Q&A, consider adding more questions to it and/or lowering it's confidence level threshold. If transferring happens through a Fallback-answer and the question matched that, improve your Fallback-instructions. Try to tell the AI what to do, instead of what not to do.
If the question matched a completely wrong Q&A and the transfer to human happens in the Fallback-answer, we advice you to create a separate "Transfer to human" Q&A for this topic to have it match this new Q&A in the future. If there's other topics you want to always escalate to a human, you can list these in your new "Transfer to human" Q&A.
My AI Assistant answers the visitor's message twice, with a duplicate answer
Look up the conversation and message in question from the "Chat log" tab and check where the answer is coming from; which Q&A did it for example match. After this, check if there are any empty or duplicate answers added in that Q&A and remove those.
Content Relevance and Accuracy
My AI Assistant answers irrelevant topics
Clarify in the general AI instructions and in the fallback answer what the bot can answer and give instructions what to do instead of answering, when irrelevant questions are asked.
My AI assistant is hallucinating links
In the general AI instructions, state clearly that it can only give out links to visitors that it finds from its own databases.
It is also worth checking from "Chat logs" where the answer (with the link) had matched and consider editing the instructions of for example that Q&A too. If there are Q&A answer instructions with many different links, those instructions can cause the bot to hallucinate similar links, that are not real.
My AI Assistant's answer is matching the wrong Q&A
When you've identified the Q&A that was used in the answer, first check that it does not have any half cut sentences. You can then add more questions to the Q&A, to increase it's probability of matching with new similar questions. You can also consider increasing the confidence level of the Q&A, if it was low to start with.
Response format
My AI Assistant's answers are too long
Instruct or modify the desired answer length in the general AI instructions. The AI does not really understand lines, it is better to set the limit to for example 3-5 sentences or 200-400 characters.