Sometimes you test your dialogues and wonder why you do not go down the appropriate branch. To determine the issue, check a few things. This article covers recommendations to avoid errors in dialogue flows.
- Conversation logs
- Conditional blocks
- Customer messages
- Carousels
- Availability blocks
- Escalation blocks
Conversation logs
Review the conversation logs to understand why a conversation fails to progress. Depending on the issue, check the following things.
Confidence score
Confidence score plays a crucial role in topic prediction. To validate a message, click the slider icon next to it. This action opens the expression overview, which shows the presumed topic, confidence level, and whether you trained it as an expression for that topic. The percentage and the threshold set in AI agent settings explain why a topic failed to trigger.
Confidence never exceeds confidence threshold for default messages
If all messages trigger the default reply and the confidence remains consistently low, you face a training issue. For more information, see this article: Optimizing your AI agent performance
Reply not triggered
To check if the reply remains active, navigate to the Topic table. Green indicates active replies.
Ensure you maintain only 1 reply per language.
To manage replies, click > Topic name > Reply. Check the active box or delete the unnecessary reply.
Conversation & session details
View the conversation details to see the performed actions and the set parameters (conversation or action based). Use this method to determine why certain things happen. Check what occurs at the conversation level, and then drill down into how it occurs.
Conditional blocks
Operators
Conditional blocks can have two issues regarding operators.
Check for the incorrect operator (for example, includes instead of in) or if you set the operator as the wrong value (string or array).
At the parent level of a conditional block, you can choose between parameters or segments. Segments allow you to check multiple values at once and simplify logic. Use operators like "match ALL", "match ANY", or "match NONE" to consolidate and streamline conditions.
Parameter collection
Filter based on a parameter. Ensure the AI agent collects and saves it in the conversation branches. Watch for the most common issue: a typo. Check the saved parameters. Confirm they exist, remain set, and match exactly.
Where appropriate, replace multiple parameter checks with a segment to simplify branching and reduce errors.
Customer messages
Order of scenarios
The AI agent prioritizes scenarios from left to right. Thus, place the most likely and least-sensitive scenario in the left-most position.
When you ask for text input, place free-text written in the right-most scenario, as it catches all text input. If you place it earlier, the net captures the entities and topics you seek. Instead, position entities and topics before free-text written.
Entity recognized vs topic predicted
Sometimes, instead of creating topics for affirmative or negation, you can create entity lists. Although this allows you to add a list of all options you can think of, it does not support multiple languages out of the box and ignores typos. The AI agent only looks for exact matches. This means if someone says "yeah, thanks for the suggestion but I do not want to do that", the AI agent catches it as affirmative, even when the overall sentence sentiment means no.
Carousels
Each CRM has its own limitation regarding carousels. Typically, issues arise when you exceed character or button limits or leave fields entirely blank (a lone empty space does not count as a character).
Availability blocks
Under the Team availability option, verify you write the correct transfer department, button ID, or group ID as an exact match.
If escalations fail to occur within the standard operating window when humans remain online, this indicates a problem with the block. The AI agent fails to divert them, or an issue exists with the CRM integration. Review the CRM troubleshooting article.
Escalation blocks
Escalation blocks typically present issues only when you fail to include the correct department or forget to add actions during custom escalations.