By Wynda Clayton,
Director of Risk Adjustment Coding & Compliance, RAAPID INC
Two years ago, I would have told you that autonomous AI coding was too risky for risk adjustment. Today, I'm implementing it—but only after finding an approach that met my compliance requirements. Here's what changed, and what it means for your 2026 planning.
The Regulatory Uncertainty Problem
Let's start with where we are right now.
CMS paused RADV extrapolation. That bought everyone breathing room. But it didn't change the underlying challenge: MA Plans still need to defend every HCC they submit.
The pause won't last forever. And when the rules shift again—and they will—the plans that survive will be the ones with defensible documentation.
That's the context for any technology decision you make right now. Speed matters. Scale matters. But defensibility matters most.
Why I Said No to Early Autonomous Coding
My initial skepticism was simple: most AI coding tools couldn't explain their decisions.
You'd input a clinical note. The AI would output HCC codes. When you asked why, you'd get technical explanations about machine learning models and confidence scores.
But there was no clear trail showing how the system arrived at each HCC. No way to see the documented evidence that supported the code.
That doesn't work in an audit.
Auditors want to see the documentation. They want MEAT—Monitoring, Evaluation, Assessment, Treatment—spelled out clearly. And they want to understand how each code was supported by the provider's diagnostic statement in the medical record.
If your AI can't show that evidence trail, you're creating liability faster than you're capturing revenue.
I saw this happen. A plan implemented AI coding without validation guardrails. The codes were submitted. Then came the audit. Half the AI-suggested codes had weak or missing MEAT documentation. When auditors asked, "Why was this code assigned?" there was no clear answer. The financial exposure was significant.
That's when I realized: autonomous coding isn't the problem. Autonomous coding without transparent documentation is the problem.
The Three Things That Had to Change
For me to trust autonomous AI, three things had to be true:
First: Every code needed documented support. Not just "I found this code," but "here's the evidence in the medical record that supports this HCC."
Second: Every code needed a complete evidence trail. Not something a human would need to reconstruct. The full path from clinical note to MEAT documentation to final HCC assignment—all documented and auditor-ready.
Third: Compliance teams had to be able to validate the evidence at scale. If you're manually reviewing every AI decision to verify the documentation, you haven't solved the capacity problem.
Most tools I evaluated couldn't clear all three bars. That's why I stayed skeptical.
What Changed: The Explainability Breakthrough
Here's the critical point: not all AI approaches deliver the same level of transparency.
The breakthrough for me came when I encountered systems built on neuro-symbolic AI—an approach that combines learning with explicit clinical reasoning rules. These systems don't just identify codes; they apply structured logic to read notes the way a trained coder would, identifying the provider's diagnostic statement and looking for specific evidence of monitoring, evaluation, and treatment.
More importantly, they document what they found and where they found it.
Every HCC comes with a complete evidence trail showing exactly where in the note each MEAT component was identified. The system links the documented evidence directly to the provider's diagnostic statement that supports the code assignment.
This level of explainability was what I'd been looking for.
Here's what this looks like in practice: The system suggests HCC 18 (Diabetes with chronic complications). Instead of just flagging the code, it shows you the evidence trail—the provider's statement of diabetic retinopathy, the documented ophthalmology visits, and the treatment adjustments. You can see exactly where each element appears in the note.
That's the level of transparency compliance teams need. It's not about the AI applying clinical criteria to establish a diagnosis—that's the provider's role. It's about the AI showing you where the provider documented the diagnosis and the supporting MEAT elements.
I watched this play out at a mid-sized plan last quarter. They had 40,000 retrospective charts sitting in the queue and 15 coders who were already maxed out. They'd get through maybe 25,000 charts before the deadline hit using their existing process.
They implemented an AI system with complete evidence trail capabilities. The system processed all 40,000 charts. When compliance spot-checked 500 random charts, the accuracy rate was 98%+. But what stood out to me was this: the coders weren't drowning anymore. They'd shifted from hunting for MEAT documentation in every routine chart to focusing on the complex cases that actually needed clinical judgment.
The compliance director told me the biggest relief wasn't the productivity gain. It was opening a chart and immediately seeing the complete evidence trail. No detective work. No guessing. Just clear documentation of what the provider stated and where the supporting MEAT existed.
That's when it clicked for me. This isn't about replacing expertise. It's about making expertise scalable—but only when the technology can prove its work.
What This Means for Your 2026 Planning
If you're planning retrospective programs for 2026, here's what I'd think about:
The capacity problem won't solve itself. Hiring more coders is expensive and slow. Your member population isn't shrinking.
Regulatory uncertainty means you need flexibility. When rules change, you need to adapt fast. Manual processes can't pivot quickly enough.
Audit risk compounds over time. Every undocumented HCC you submit is a future liability. The longer you wait to fix documentation quality, the bigger the exposure.
Autonomous AI with explainability is one solution. But the specific approach matters.
Questions to Ask Before You Implement
Don't take anyone's word for it—including mine. Here's what I'd ask any vendor:
How does your AI explain its coding decisions? Look for systems that provide complete evidence trails, not just confidence scores.
Can you show me exactly where the MEAT documentation is in the note? Every component should be highlighted and linked to the code.
What type of AI architecture do you use? Understanding whether a system uses approaches like neuro-symbolic AI can help you evaluate its explainability capabilities.
How does your compliance team validate the documented evidence? If the answer is "manually review everything," that's not really autonomous.
What happens when the AI is uncertain? Good systems flag edge cases for human review.
Can this documentation survive an audit? Ask for examples. Ask to see rejected charts and how they were handled.
If you don't get clear answers, keep looking.
Where I Landed
I'm not saying every plan needs autonomous coding tomorrow. And I'm not saying all autonomous AI has reached the same level of maturity.
But I am saying that AI systems with true explainability—those that can show their clinical reasoning—have reached a point where they're defensible.
The regulatory environment isn't getting simpler. The documentation requirements aren't getting easier. And your coding teams are already stretched.
The question isn't whether AI in general is ready. It's whether the specific system you're evaluating delivers both the accuracy and the transparency you need to defend results.
For me, the answer changed when I saw an AI that could show its evidence trail clearly and completely. That was the breakthrough.
If you're evaluating options for 2026, make explainability your non-negotiable. Ask vendors to prove their AI can show its work. The difference isn't just technical—it's the difference between adding risk and reducing it.
The editorial staff had no role in this post's creation.