AI for Risk Adjustment: What Health Plans Need to Know Before CMS Audits Catch Up

In May 2025, the Centers for Medicare & Medicaid Services (CMS) announced it will audit every eligible Medicare Advantage contract for Risk Adjustment Data Validation (RADV), replacing the old approach of selecting roughly 60 out of 555 plans each year. With CMS also working through an audit backlog that goes back to plan year 2018, RADV is no longer a low-probability event for any single plan.

The RADV audit process is also getting tighter. Sample sizes per plan have risen from 35 enrollees to as many as 200, and the review window has shrunk. A health plan still relying on a fully manual, after-the-fact review process is working against a shorter clock with a larger expected sample.

This piece looks at why AI is becoming a standard part of the risk adjustment toolkit, what it excels at, and what to consider before bringing it into your own workflow.

Main takeaways

1. CMS’ audit scope has expanded to every eligible MA contract, and it’s reviewing more records on a shorter timeline.
2. Manual review alone doesn’t hold up at this volume. Plans already review tens of thousands to millions of records a year, and regulators are watching more closely than they used to.
3. AI’s clearest use in risk adjustment is spotting patterns. It can effectively flag conditions that resemble past overcoding errors or point reviewers to the most relevant part of a chart.
4. Regulators expect a person in the loop. The OIG and DOJ have both raised concerns about relying entirely on AI or NLP to review medical records.

Why a RADV finding can be so costly

Under its 2023 RADV rule, CMS would extrapolate audit results, applying an error rate from a sample of members to the plan’s entire audited population for that payment year. A few unsupported diagnoses in a couple hundred records can become a contract-wide repayment. A federal court struck down that rule in 2025, and CMS has appealed. For now, CMS recovers only on the records it samples, but plans should prepare for extrapolation to return.

Either way, it’s extremely valuable to catch errors internally. An unsupported diagnosis that CMS finds during its review carries a penalty. The same diagnosis, identified and corrected by the plan before submission, carries none. That’s why interest in proactive, AI-supported chart review has picked up alongside the audit changes.

The manual review model is already stretched

Medical record review has always required people who can read a chart and confirm what is documented. Depending on the size of the plan, that means manually reviewing tens of thousands to millions of records a year.

Electronic record retrievals were supposed to make this easier, and in some ways they have. As of 2021, an estimated 88% of hospitals could send and receive patient health information electronically. But access to a record isn’t the same as a usable record. Missing signatures, absent face-to-face encounter dates, and formatting problems with imported text still show up often enough to slow reviewers down. The combination of rising volume, tighter scrutiny, and uneven data quality is pushing plans to look at AI and natural language processing (NLP) as review support rather than a nice-to-have.

What AI excels at in risk adjustment

The best use of AI in risk adjustment workflows is pattern recognition. Used well, AI can surface conditions that look potentially unsupported across an entire managed population. It can also point a reviewer to the specific passage in a chart that supports (or doesn’t support) a diagnosis, instead of requiring a full manual read of every page.

Both CMS and audited plans are operating under the same constraint where a person confirms what the model flags. While AI won’t replace human experts, it is transforming how care is delivered across the healthcare industry. Rachel Stemerman, MPH, Ph.D., Senior Principal Product Manager of Data Engineering and AI Platform at Inovalon, put it: “AI reflects us. It reflects our decisions, good or bad; the data we put in, biased or not.” The OIG and DOJ have made a related point from the compliance side, noting that relying entirely on AI, NLP, or machine learning for medical record review, without human oversight, is itself a risk factor to monitor. The plans getting real value out of AI here are treating it as a way to prioritize and explain reviewer attention, not a replacement for the reviewer.

Beyond audit defense: what’s still unsettled

RADV defense is the most urgent reason plans are looking at AI in risk adjustment right now, but it’s not the only place the same underlying approach applies. A few questions remain unanswered:

• Can AI pull a clearer picture out of fragmented data? A single member’s clinical picture is often split across an EHR with free-text provider notes and a claims system. The same pattern-matching and NLP work described above is, in principle, well suited to pulling a clearer, fuller picture out of that fragmentation. But it’s unknown how many plans may be realizing these benefits versus still building toward them.
• Should risk models account for social drivers of health? Housing, transportation, and access to care all affect a member’s real risk and cost of care, but most current risk-adjustment models run on diagnosis codes alone. Whether and how SDOH data gets folded into risk scoring, rather than treated as a separate quality initiative, is still being worked out.
• Does overcoding review double as fraud detection? An honest but unsupported diagnosis and a deliberately upcoded one aren’t the same problem, but they can look similar in the data. Extending pattern recognition from one to the other is a natural next step, though it raises its own accuracy and false-positive questions.
• How does risk adjustment stay aligned with value-based contracts? As more MA business moves into value-based arrangements, a risk score that tracks coding accuracy, but not real-world outcomes, becomes a liability rather than a compliance tool. Recalibrating models against actual outcomes is a harder problem than audit defense alone, and one the industry hasn’t fully solved.

While none of these questions are fully settled yet, they provide a reasonable guess at where the audit-defense use case will expand over the next few years, once plans have the RADV-facing basics in place.

What Inovalon’s clinical team expects to change

Kathryn Eshelman, M.D., MPH, VP of Medical Informatics at Inovalon, suggests that health plans focus on audit readiness and better address risk-adjustable conditions with AI support, particularly AI built directly into the EHR workflow physicians already use. The idea is that flagging and documentation happen in the moment of care, so the physician knows what to address, the patient gets the care the record should reflect, and fewer separate assessments are needed after the fact.

Questions to ask before adopting AI for risk adjustment

Not every AI tool marketed for risk adjustment is built in the same way. Here are a few questions worth asking before you bring a new solution into your workflow:

1. Does it explain the flag, or just issue one? A prediction without a reason is hard for a reviewer to act on quickly.
2. Does it fit into the review process your team already uses? A separate system to learn adds friction; a tool that surfaces flags inside existing chart review is easier to adopt.
3. Does a person confirm every flag before action is taken? This is the line regulators are watching, and it’s also just good practice.
4. Does it help you act before submission, not just after a denial or audit finding? Catching a likely issue early is a different job than explaining one after the fact.

How Inovalon approaches AI in risk adjustment

Inovalon builds AI into risk adjustment as a way to direct reviewer attention, not to replace clinical judgment. By identifying potentially unsupported conditions before CMS does, and helping reviewers move faster through the documentation once an audit is underway, Inovalon’s risk adjustment approach provides full coverage. For a closer look at each, see RADV Audits: Taking a Proactive Approach in the Face of New Regulations and Mastering Record Retrieval: Smarter MRR Strategies for Risk Adjustment Success.

If you want to talk through what this looks like for your plan, contact our team.

Frequently asked questions

What is AI in risk adjustment?

AI in risk adjustment refers to machine learning and natural language processing tools that help health plans and reviewers spot conditions likely to be over- or under-coded and surface the documentation that supports or contradicts a diagnosis. It’s used to prioritize and inform human review, not to submit or approve codes on its own.

How is CMS using AI in RADV audits?

CMS has said it will deploy AI to help review and flag unsupported diagnoses as it scales up RADV audits, though it has not released details about the tool. This is happening alongside a broader expansion of audit scope, sample size, and coding staff.

Can AI replace manual medical record review?

AI cannot replace manual medical record review. The OIG and DOJ have flagged full reliance on AI, NLP, or machine learning for medical record review as a risk factor. AI is most useful to prioritize which charts need attention first and explain why, with a person still confirming each flag.

What should a health plan look for in an AI risk adjustment tool, from a compliance standpoint?

When looking for an AI risk adjustment tool, look for one that:

• Explains why a claim or condition was flagged rather than issuing an unexplained score
• Fits into the chart review process your team already uses
• Keeps a person in the loop on every flag before any action is taken

---

Inovalon and design® and Inovalon® are trademarks of Inovalon, Inc.