AI & Technology
AI in the revenue cycle: what's real, what's hype, and what to ask vendors
Sam Whitfield, Chief Technology Officer · April 22, 2026 · 9 min read
AI genuinely works in billing — for specific, well-defined jobs. Here's where it delivers measurable results today, and the questions that expose vendors selling buzzwords.
Every RCM vendor now claims AI. Some of it is transformative; some of it is a rules engine wearing a lanyard. Having built and deployed these systems in production, here's our honest map of the territory.
What works today: denial prediction. Models trained on claim outcomes can score denial risk before submission with enough accuracy to route high-risk claims for human review. In production we see 25–35% denial reductions from pre-submission scoring — this is the most mature AI application in RCM.
What works today: work queue prioritization. Ranking A/R follow-up by expected recovery value (not just age or amount) measurably increases dollars recovered per work hour. It's unglamorous and very effective.
What works today: anomaly detection. Payer behavior changes — a quiet policy shift, a new bundling edit, systematic downcoding — surface in claims data within days when models watch for distribution changes. Humans reviewing monthly reports catch the same shift a quarter later.
What's still maturing: autonomous coding. AI-assisted coding with human review is production-ready for high-volume, low-complexity encounter types. Fully autonomous coding of complex encounters is not — accuracy claims that sound too good usually exclude the hard cases from the denominator.
Questions that expose hype: Does the AI act autonomously or recommend to humans? What specific metric improved for a reference client, by how much, over what period? Is PHI used for cross-client training? Is the model's work auditable claim by claim?
Our position: AI prioritizes, humans decide. The leverage is real — but it comes from pairing models with experts, not replacing them.