A few months ago, in a noisy New York bar, I ended up talking with a consultant who works with payers. She called a common rev cycle practice attestation theater, and I cannot quit thinking about it.
She was describing what happens inside a lot of risk-adjustment programs. A query lands on a physician’s phone, and with a few taps, an auto-generated note populates the chart to support a diagnosis. The software may be accurate and the physician’s answer may be right, but when a RADV auditor or an OIG investigator pulls that claim, they won’t ask whether the diagnosis was clinically correct. They’ll ask whether the documentation reflects the physician’s own independent clinical judgment, or if there’s any evidence of a real chart review.5
Upon reading this, it may seem like the physicians and coders using these tools are doing something wrong. The truth is, they inherited a broken workflow and regulators have finally run out of patience with it.
If a diagnosis can’t stand on a real encounter, don’t expect to be paid for it
CMS has already moved. In its recent rate-setting, it finalized the exclusion of diagnoses drawn from unlinked chart-review records from risk-score calculation.1 The message is not subtle. If a diagnosis cannot stand on a real clinical encounter, do not expect CMS to pay for it. They may have started with Medicare Advantage, but anyone who assumes retrospective queries are not the next obvious target isn’t paying attention.
The enforcement side is moving in the same direction. The recent DOJ settlements over invalid HCC submissions, which run into the hundreds of millions of dollars, share a common theme. Programs have both added diagnoses without the clinical evidence to defend them and also have failed to remove the ones that are no longer pertinent.23 The way risk adjustment was designed in the past no longer makes sense. The information must be real, defensible, and clinically relevant.
Now layer V28 on top of all of this. The model that went fully live in 2026 removed more than 2,000 codes and demands more specificity than most programs were ever built to produce.4 A lot of organizations are watching their average RAF scores drift downward and reaching for the same retrospective machinery to claw it back, right as that machinery becomes the riskiest part of the operation.
The programs that leaned hardest into retrospective machinery are now the most exposed, and I fear that most don’t see it yet.
The programs that leaned hardest into retrospective machinery are now the most exposed.
Recapture was never the real prize
Even setting aside the compliance risk, retrospective review was always solving for the wrong thing.
A retrospective query can only act on what is already written down. It attempts to bring forward and capture previous diagnoses lost in the chart. If the diagnosis is so lost that the clinician isn’t addressing it in their care and needs to be retrospectively surfaced, is it relevant? Is the patient benefiting from that recapture work done later?
The conditions that actually move a patient’s risk picture and their care are sitting in the data fully supported but undiagnosed because the chart is just too massive. This looks like the CKD stage in a trended creatinine or the heart failure in the BNP and the medication list, or the malnutrition in the weight trends. The chart supports the diagnosis, but no one made it. These are not coding gaps, but clinical gaps. And they are invisible to a tool that mines documentation, because the diagnosis for which it is looking was never entered. Recapture cannot find a condition that was never actually diagnosed.
Existing tools reach only what is already documented. The conditions that move a patient’s risk — and their care — sit below the line.
We have spent enormous energy optimizing the back end of this problem with better queries, faster attestations, and bigger appeal engines, but almost none of it touches where the problem actually begins.
The fix runs upstream
Organizations that set themselves up for success are the ones that capture the right diagnosis in the right encounter, while care is actually happening, with the appropriate clinical evidence attached from the start. The results are better for everybody involved. It is better care for patients, because a condition that is newly diagnosed can be better treated and managed. It’s better for the auditor, because every diagnosis traces to true clinical evidence. And it’s the safest way to protect revenue, because it’s the only version CMS and the DOJ have no interest in tearing down.
See the gap in your own data
Most health system leaders suspect there are large gaps between what their data already knows about their patients and what their documentation actually captures; however, very few can tell you how those gaps impact their revenue and quality.
At Regard, this is our passion. At one large academic health system, Regard ran an HCC gap analysis on a 500-patient sample. The diagnostic engine read those charts against the conditions the system had already documented, scored the gap using CMS’s V28 risk model, and extrapolated across the roughly 3,700 patients in the system’s documented-HCC population.
On net-new undiagnosed conditions identified from the labs, vitals, and medication history, rather than recapture or specificity tweaks, Regard added an average of 2.4 HCCs per patient. That worked out to about $4,850 in additional annualized risk-adjusted revenue per patient, and more than $18 million across the documented population. Counting every condition Regard found, not just the net-new, the projection rose to roughly $29 million.
Sometimes that number comes back modest, but more often it does not. Either way, it is a far more honest starting point than a recapture rate.
But the reason I personally do this work is core to my heart as a physician of 25 years. An undiagnosed condition is not just a coding gap. It represents a patient who is not receiving the treatment they need. For most of my career, the limiting factor was that no human had the bandwidth to read the entire chart on every patient. With AI that constraint is finally lifting. The real question now is whether we use these power tools to chase the last dollar of a dying retrospective model, or do we finally document care the way it is actually happening.
I know which one I’m backing.
