RADV audits have moved from a periodic compliance exercise to a standing operational risk. CMS now audits every eligible Medicare Advantage contract annually and extrapolates findings, so plans face a validation workload that scales faster than the coder capacity available to absorb it. The charts under review were often submitted years earlier, by prior vendors, on platforms no longer in use, and the burden of proving support falls on teams that did not create the documentation.

This session examines how AI changes the economics of that work, covering where automation helps and where it does not: classifying document and encounter types, checking signature validity and provider credentials against CMS lists, confirming dates of service, and linking a submitted HCC to the specific sentence in the record that supports it. Sean Creighton and Hasham Ul Haq walk through what a tiered review model looks like in production at Martlet AI, with automated scoring across the full population, human coders reserved for the exceptions, QA sampling and compliance sign-off. They show the evidence packet a plan should be able to produce for any code it defends.

They also cover the constraints that decide whether any of this is deployable: keeping PHI inside the plan’s network, working from scanned PDFs rather than structured data, and building audit trails that hold up under scrutiny. We also cover mock audits, proactive sampling, the five-month post-audit submission window, and the limits of current model performance.