AI voice intake that cut patient hold time from 94 to 22 minutes

Patients calling the network’s intake line waited an average of 94 minutes. Staff were triaging calls manually, re-keying the same demographic and symptom data into Epic, and losing callers — many of whom simply drove to an ER instead.

The network needed intake capacity that scaled with call volume, without adding headcount and without moving protected health information outside a controlled environment.

A conversational voice agent built on Amazon Bedrock answers every call immediately. It verifies the caller, runs a structured intake dialogue, and writes clean records into the Epic workflow — so staff review instead of transcribe.

Calls the agent can’t or shouldn’t handle — clinical judgment, distress, ambiguity — hand off to a human with the transcript and structured data attached, so nothing restarts from zero.

Average hold time fell from 94 minutes to 22 — a 77% reduction — while intake data quality improved because records arrive structured instead of re-keyed under pressure.

The system runs inside the network’s own cloud boundary with the safeguards described in the claims note above.

Before a build like this we run the same four checks the readiness assessment does, and error cost is the one that decides the design: a wrong answer in patient intake can never reach a caller unreviewed. So the human-handoff path — distress, clinical ambiguity, anything the agent shouldn’t decide — is built first, not bolted on.

We’d also confirm the intake data has somewhere clean to land (here, the Epic workflow) before promising any hold-time number. The 77% reduction held because the boring integration work came first.