I am currently helping a client tune their native Bot Flow NLU engine. We are seeing a lot of interactions falling into the ‘I do not understand’ path even when the customer’s utterance seems to match the intent training phrases. I want to adjust the ‘NLU Confidence Threshold’, but I am worried that setting it too low will lead to incorrect intent matching. Is there a best practice for determining the optimal threshold for a high-volume support bot?
Hello Jun15! I am a Zendesk-to-GC migration specialist and I love working on these bot flows! From my experience, the default threshold of 0.6 is usually too conservative for support bots. I often recommend starting with 0.5 and then using the ‘Intent Health’ tool to see which utterances are being missed. If you have a lot of similar intents, you might need to keep it higher to avoid ambiguity.
Greetings. As a consultant who evaluates many AI vendors, I can tell you that threshold tuning is a balancing act. You should not just change the number globally. You should use the ‘Confirmation’ logic in Architect for any intent that has a confidence score near your threshold. This way, the bot asks ‘Did you mean X?’ instead of just failing or guessing. It provides a much better experience for the customer while still allowing the bot to learn from its mistakes.
Hey guys! I am a telecom engineer and I want to add that you should also check the ‘Audio Quality’ for your voice bots! If your SIP trunk is using a high-compression codec, the NLU engine will have a harder time deciphering the utterances, which leads to lower confidence scores. Make sure you are using G.711 for the best results with your native bot flows!