An AI voice agent that took over 40 hours a week of insurance follow-up calls

Industry

Healthcare RCM

Market

USA

Client

US revenue cycle management company

Sponsor

President

What we built

Voice AI agent

The situation

The client is a revenue cycle management company in the US that works claims on behalf of provider groups. Like every RCM team, a large share of its week went to one activity: calling insurance payers to find out what happened to a claim. Is it received, in process, pended for records, denied, paid? The answer usually exists, but getting it means dialling the payer, working through an IVR menu, waiting on hold, reaching a representative, reading out identifiers, and writing the answer back into the billing system.

By the time the President brought us in, that loop was consuming 40 hours a week of staff time: a full-time role spent on hold. Worse, the calls that mattered (a denial that needed an appeal before a deadline) waited in the same queue as the calls that did not (a claim that was simply in process). The collections team's skill was going into dialling rather than resolving.

What we built

A voice agent that makes the follow-up call end to end, the same way an experienced biller does.

It works from the claims queue, not a call list. Each morning the agent picks up the claims that are due for status follow-up, with the identifiers a representative will ask for: payer, member ID, date of service, billed amount, claim number.

It navigates payer IVR trees. Every payer's phone system is different, and they change without notice. The agent listens to the menu, chooses the path for claim status, enters or speaks the identifiers when asked, and handles the "I didn't understand that" loops that trip up simple dialers. We covered the mechanics of this in how AI agents navigate payer IVR systems.

It waits on hold so no one else has to. Hold time is where most of those 40 hours went. The agent waits, detects when a human picks up, and starts the conversation.

It talks to the representative. It identifies the provider, gives the claim details in the order reps expect, asks the status questions, and asks the follow-up questions a good biller would: what is missing, what the denial reason is, what the reference number is for the call, whether reprocessing has been requested.

It writes everything back. Status, reason, next action and call reference go into the client's system against the claim, so the record is complete without anyone retyping notes.

It routes exceptions to people. A denial that needs an appeal, a request for medical records, a payer asking for something the agent is not authorised to give: these land in an exceptions queue with the call summary attached, so a biller picks up exactly where the agent left off.

How it went live

We started from the client's real call patterns rather than a script. The discovery sprint mapped the payers that generated the most follow-up volume and the questions reps asked on each. The agent was then tested against those payers before it touched production claims, and rolled out payer by payer so the team could compare the agent's write-backs with what their own billers would have recorded.

Because the calls carry patient identifiers, the build was scoped for HIPAA from the start: only the identifiers each call needs, defined retention for recordings and transcripts, and access limited to the people who work the claims.

Results

The agent now handles the follow-up calls autonomously. The 40 hours a week the team used to spend on hold and dialling is gone, and the collections team works exceptions only: the denials, records requests and appeals where their judgement actually changes the outcome. The President reported that ROI showed in the first month.

What made the difference

Write-back, not just calling. An agent that makes calls but leaves notes for someone to key in only moves the work. Writing structured results into the claim record is what removed the hours.

Exceptions as a first-class output. The agent is judged on how cleanly it hands off, not just on how many calls it closes. A biller who opens an exception sees the full call summary and never has to call the payer back to find out what was said.

Payer-by-payer rollout. Each payer's IVR is its own small project. Launching them one at a time kept the team confident in the agent and made problems easy to isolate.