Retrospective Risk Adjustment Coding Is Changing. Here’s What the New Model Looks Like.

The Function That Went From Offense to Defense

For most of its history, retrospective chart review in Medicare Advantage served one purpose: find more diagnoses, submit more codes, increase revenue. Plans measured their programs on codes added and RAF scores gained. Vendors competed on volume and speed. The entire function was structured as an offensive revenue strategy.

That positioning became a liability. The DOJ’s enforcement actions established that programs designed primarily to increase risk scores, without corresponding mechanisms to validate accuracy and remove unsupported codes, create False Claims Act exposure. Kaiser paid $556 million. Aetna paid $117.7 million in March 2026. OIG’s February 2026 guidance formalized add-only chart review as a specifically identified high-risk practice.

Retrospective review didn’t become illegitimate. Its purpose changed. It shifted from an offensive revenue engine to a defensive compliance function. Plans that haven’t made that mental shift are still running programs that look like the ones that generated settlements.

What Defense Looks Like in Practice

A defensive retrospective program starts from a different premise. Instead of asking “what codes can we add?”, it asks “what have we submitted, and can we prove all of it?” The workflow begins with the plan’s existing submissions and works backward to validate each one against MEAT-based clinical evidence.

Codes with strong documentation support stay. Codes without adequate evidence get flagged for removal or provider query. Missed diagnoses that the initial coding process overlooked get identified and added with full documentation trails. The output is a cleaned, validated submission set where every code has an evidence-based justification.

This requires different technology than add-only programs used. AI needs to evaluate documentation in both directions: identifying what’s missing (adds) and identifying what’s unsupported (deletes). The system needs to link every recommendation to specific clinical language in the note, mapping MEAT criteria explicitly. And it needs to produce output that an auditor can follow without additional interpretation.

The cultural shift matters as much as the technology. When coding teams have been measured on adds for years, shifting to a model where a delete counts as a compliance win requires leadership to redefine performance metrics. Revenue teams need to accept that removing an unsupported code protects more revenue than submitting a code that gets clawed back in an audit.

The Population-Level Signal Problem

CMS doesn’t just audit individual charts. It monitors coding patterns at the population level. When a plan’s risk scores rise consistently without corresponding changes in clinical outcomes, hospitalization rates, or cost patterns, that divergence tells a story. Add-only programs produce exactly this signal: steady upward drift in coded complexity that doesn’t match what’s actually happening clinically.

Two-way retrospective programs produce a different population-level signal. Risk scores stabilize at levels that reflect actual patient complexity. Some members’ scores go up (legitimate adds). Some go down (unsupported deletes). The net effect is a coding profile that tracks clinical reality rather than drifting away from it. That’s a pattern CMS finds defensible because it demonstrates the plan is managing its data honestly.

Plans running add-only programs often don’t realize how visible the pattern is until an audit arrives. By then, years of one-directional coding have created a statistical profile that’s difficult to explain and expensive to remediate. You May Also Like Jernsenger.

Building the Defensible Model

The transition from offensive to defensive retrospective isn’t optional. The enforcement precedent is set. The regulatory guidance is published. The model changes are in effect. Plans still running add-only programs are operating on borrowed time, and the interest rate on that borrowed time is measured in settlement figures.

Organizations rebuilding their programs around retrospective risk adjustment coding that validates in both directions are aligning with what CMS, OIG, and DOJ have explicitly demanded. Every submitted code gets an evidence trail. Every unsupported code gets flagged. Every chart review produces both adds and deletes. That’s the model that survives the current enforcement environment, and it’s the only model that should be operating in 2026.