Capital One: a machine learning voice agent on live regulated calls

A bank's inbound call is not a customer-service conversation with some rules attached. It is a regulated interaction in which what the caller is told, what is done on their account and who was allowed to do it can all be challenged later. Any system that answers those calls has to hold to that standard on every call, not on average.

The engagement covered a machine learning voice agent on live regulated calls for Capital One, with the boundary around the agent designed before the agent itself.

The agent recognises what the caller wants and routes the call accordingly. Where it acts, it acts within actions the bank's policy allows, and nothing outside that policy is available to it. Where the call needs a person, the agent escalates with the full context of what has been said and done, so the caller does not start again.

Every decision is on record. That is what makes the agent usable in a bank: a call can be reconstructed afterwards, the reasoning for a routing or an action can be shown, and an examiner gets an answer rather than a recording to listen to.

The structure of the Capital One agent, intent recognition, bounded actions, logged decisions and escalation with context, is the structure Opsian builds every agent on, whether it answers a hotel's phone line or a bank's. The regulated setting is where the pattern was proved hardest.