Hi everyone! I am currently helping a client migrate their digital support stack from Zendesk to Genesys Cloud. In Zendesk, we used to rely on simple keyword matching for our chat bots, which was straightforward but lacked depth. Now, we are moving to Bot Studio to leverage the advanced NLU capabilities in Genesys Cloud. However, we are hitting a specific technical wall regarding intent confidence scores.
The issue is that intents which were clearly identified by Zendesk’s bot are returning low confidence scores (around 0.4) in Genesys Cloud, causing the bot to frequently trigger the ‘No Match’ fallback instead of the intended action. We have verified that the utterances are identical in both systems. We are using the latest version of the Genesys Cloud Bot Studio and the standard Zendesk Web Widget connector for the migration.
When we test via the Bot Studio simulator, the intent matches correctly with a score of 0.9. But when tested through the actual Zendesk widget integration, the score drops significantly. We suspect there might be a difference in how the text payload is pre-processed or normalized before hitting the NLU engine in the live environment versus the simulator.
Has anyone else encountered this discrepancy during a similar migration? Are there specific API parameters or preprocessing steps in the Architect flow that we need to adjust to ensure the text input matches the simulator’s expectations? Any practical advice on aligning these confidence thresholds would be greatly appreciated!