Databricks presents a blueprint for voice agents using business context
Databricks’ October 28 article connects spoken customer service with governed business data and points developers to its DIVA blueprint.
The Databricks article separates speech handling from decisions based on customer records and business rules. Its prototype uses approved, read-only lookups into business systems. Engineers recommend starting with a 90-day pilot for one channel, one team and a frequent routine call, measuring current call cost, handling time and satisfaction before the pilot. This is a recommended evaluation plan.
The DIVA repository describes fictional company data. This is an architecture sketch, not setup instructions: consult its complete prerequisites. A browser joins LiveKit; Lakebase supplies the caller’s tier; Deepgram transcribes speech, and Unity Gateway serves the model. Record lookups use only that caller’s records. Optional tracing needs separate configuration. The optional voice-mode decision requires an administrator to enable the ai_decide preview; it can be disabled. Switching voice leaves the model and tier unchanged. ElevenLabs is optional, with a Deepgram voice as fallback.
The separate AI Decide announcement explains the decision component: it evaluates questions against text and returns a probability, a named choice or an ordered-scale score. The function is in Beta and can be called through SQL or REST. These outputs are decision data, rather than evidence that a business transaction has been executed. One documented evaluation example supplies customer questions, generated answers and a refund policy, then asks whether each answer follows that policy and how completely it addresses the request. This is an example of evaluating responses, not a reported customer outcome.
