What is the reason the Genesys Cloud Bot Analytics dashboard is showing a significant divergence between reported intent recognition rates and actual agent handoff volumes for our Paris-based customer service queue? The environment is running the latest stable release, and the bot flow was updated two weeks ago to incorporate new NLP models for order status inquiries.
The specific issue manifests in the Performance view. The bot reports a 94% successful intent resolution rate for the ‘Order_Status’ intent over the last 7-day window. However, the Agent Assist metrics indicate that 18% of conversations flagged with this intent resulted in immediate transfers to human agents due to ‘Confidence Threshold Not Met’ or ‘Fallback Triggered’. This suggests the bot is claiming success where the underlying confidence score was actually insufficient to maintain the session.
We are utilizing the standard Architect flow configuration with a confidence threshold set at 0.75. The bot logs show that these fallbacks occur when the NLP engine returns a score between 0.65 and 0.74, which should theoretically trigger the fallback path, not count as a resolved intent. Yet, the summary dashboard aggregates these as resolved interactions until the transfer occurs. This discrepancy makes it difficult to assess the true ROI of the AI implementation for the finance team.
Is there a known limitation in how the analytics engine calculates ‘Resolved’ versus ‘Transferred’ for intents that are initially matched but subsequently abandoned by the bot? We need to reconcile these numbers to justify the current NLP model costs. Any insight into the calculation logic or a workaround to filter these low-confidence matches from the success rate would be appreciated.