Patient intake by phone: a voice agent case study

Patient intake sits at an interesting intersection for voice automation — high repetition and clear documented process, but also genuine sensitivity that requires careful boundaries. Here's how we actually built one.

The starting problem

A multi-location healthcare practice was handling new patient intake calls — collecting demographic information, insurance details, reason for visit, and scheduling the initial appointment — entirely by front-desk staff, who reported intake calls as their single most time-consuming and interruption-heavy recurring task, averaging around 12 minutes per call and frequently competing with in-person patients waiting at the desk simultaneously.

What we automated, and what we deliberately didn't

The voice agent handled: collecting and verifying demographic information, capturing insurance details (with a structured confirmation read-back to reduce transcription errors), and initial appointment scheduling based on real-time calendar availability. It explicitly did not handle: any clinical triage or medical judgment about urgency, any question requiring interpretation of insurance coverage specifics beyond capturing the raw information, and any call where the caller indicated distress or an urgent medical concern — those routed immediately to a live staff member, no automated attempt first, given the real stakes of getting that routing wrong.

The technical details that mattered most for accuracy

Insurance information capture required careful design because insurance member IDs and group numbers are alphanumeric strings that are easy to mishear or mistranscribe over phone audio quality. We built explicit confirmation loops — the system reads back captured information ("I have your member ID as A-B-1-2-3-4-5-6, is that correct?") before finalizing, rather than accepting single-pass transcription as final, which meaningfully reduced downstream data-entry errors compared to earlier testing without this confirmation step.

The results, measured against the original baseline

Post-launch, the voice agent handled a majority of new patient intake calls to completion without staff involvement, at an average call duration of about 7 minutes — shorter than the previous human-staffed average, largely because the system didn't experience the interruptions (a walk-in patient, a ringing second line) that had extended human-handled calls. Data entry error rate on captured insurance information, tracked via the practice's billing team flagging claim issues traceable to intake data, showed no meaningful increase compared to the previous human-handled baseline — validating that the confirmation-loop design was catching the transcription errors it was built to catch.

Front-desk staff, freed from the bulk of routine intake calls, reported being able to give more attention to in-person patients and to the calls that did route to them — the genuinely complex or sensitive ones the system was designed to hand off.

How Ndakum approaches it

Healthcare intake automation requires this kind of careful scoping — what to automate and what to deliberately keep human — which is core to how we design every AI Voice Agent for a regulated or sensitive context.