We A/B tested Agent Assist with our predictive routing model enabled.
Group A: Predictive routing + Agent Assist → 12% AHT reduction
Group B: Predictive routing only → 6% AHT reduction
Group C: Agent Assist only → 8% AHT reduction
The combination produces a multiplicative effect because the ML model routes the call to an agent who is MOST LIKELY to benefit from the suggested article.
Under GDPR, the real-time transcription data processed by Agent Assist constitutes automated processing of personal data.
If the customer mentions their name, address, or account number during the call, that PII is processed by the NLU engine. You must include Agent Assist in your Record of Processing Activities (ROPA) under Article 30.
The latency you are seeing is consistent with the Speech-to-Text pipeline architecture.
The audio stream is chunked into 3-second segments, transcribed, tokenized, and then matched against the knowledge base index. Each step adds latency. The total pipeline is approximately 8-12 seconds under normal load. The 15-20 second delay you report suggests either high tenant contention or a large knowledge base with poor indexing.