Denied claims are revenue you already earned, walking out the door. For this hospital system, the revenue cycle leaked at every seam. Claims went out with errors, came back denied, and sat in queues while staff reworked them by hand. Denials that could have been appealed often weren't, simply because no one got to them in time.

The work was endless and reactive. Billers chased yesterday's denials instead of preventing tomorrow's, A/R days climbed, and a real slice of earned revenue was quietly written off as the cost of doing business.

The challenge

Could the revenue cycle largely run itself, catching errors before claims go out and working denials the moment they come back? The system needed scale and consistency, but also a way to keep humans firmly in charge of the cases that genuinely need judgment.

The approach

We architected an agentic platform: a set of specialized AI agents, coordinated by an orchestrator, each owning a stage of the cycle. One agent codes the encounter. Another scrubs the claim against payer rules. A prediction model flags the claims most likely to be denied before they are submitted, and an appeals agent drafts and files the response when a denial does land. Humans supervise and take the genuinely hard cases; the agents handle the volume.

01
Agents that own each stage
Coding, scrubbing, denial prediction, appeals, and posting are each handled by a dedicated agent, coordinated rather than crammed into one monolith.
02
Denials predicted, not just processed
A model scores every claim before submission and flags the ones headed for trouble, so they get fixed while it is still cheap to fix them.
03
Appeals that file themselves
When a denial is appealable, an agent assembles the evidence and files the response, instead of letting it expire unworked in a queue.
04
Humans on the exceptions
Specialists supervise the platform and take the edge cases that need real judgment, while the routine volume runs autonomously.

The agents handle the thousand routine claims. The people handle the ten that actually need them.

Agentic loop: claim intake, autonomous coding, claim scrubbing, denial prediction, and automated appeal, supervised by humans
FIG.02A coordinated loop: claim intake, autonomous coding, scrubbing, denial prediction, and automated appeal, learning from every outcome.

The outcome

The clean-claim rate climbed into the high nineties, denials fell by a third, and A/R days dropped sharply as claims went out right the first time and appeals stopped expiring unworked. In its first year the platform recovered millions in revenue that the old, manual cycle had been quietly losing.

The cheapest denial is the one that never happens. Earned revenue stopped slipping away.

The platform gets better as it runs. Every approval and denial sharpens the prediction models, so the system that recovered revenue this year is positioned to leak even less the next.