Most big stories are hard to hold in your head. This one fits into three numbers, and once you have them, an enormous amount of American healthcare news from the past six months suddenly makes sense.
The numbers are 91, 117.7, and 2,000. Here is what each one means.
91 percent
That is the top of the error range federal auditors found this spring when they examined how well insurers could support the diagnoses they had billed for. Across three separate audits of Medicare Advantage plans, between 81 and 91 percent of the high-risk diagnosis codes sampled were not adequately supported by the medical records behind them.
Read that slowly, because the instinct is to assume it means fraud on a massive scale. Mostly it means something more mundane and more structural. The commonest failure was a condition a patient genuinely had at some point, still recorded as though it were active today. A stroke from years earlier, coded as current. A resolved illness that nobody ever removed from the file. Records that grew but never shrank.
Why does that matter financially? Because in Medicare Advantage, private insurers cover older Americans and the government pays them monthly amounts that rise with how ill each member’s documented conditions show them to be. Sicker files, bigger payments. When files drift upward and nobody checks, payments drift with them.
117.7 million dollars
That is what one major Medicare Advantage insurer agreed to pay in March 2026 to settle federal claims about how its diagnosis records were assembled. The detail worth noticing is not the size of the cheque. It is what prosecutors focused on.
The insurer ran a programme that reviewed members’ old charts looking for conditions to add. That, by itself, is legal and can be legitimate, because busy clinicians genuinely do miss documenting conditions they are actively treating. What drew the government’s attention was that the programme almost never worked in the other direction. Years of reviewing, and the reviewing consistently found errors that increased payment while rarely finding errors that decreased it.
An accuracy programme that only ever produces convenient answers is, statistically, not an accuracy programme. That asymmetry was the case.
2,000 coders
That is roughly the size of the federal audit workforce now checking this industry, up from around forty. Audits run on a rolling quarterly schedule rather than occasionally. Payment year 2020 reviews began early in 2026, payment year 2021 followed in May, and further years are scheduled into 2027.
The mechanics carry real weight: when auditors find an error rate in a sample of a plan’s members, that rate can be applied across the plan’s entire contract. A modest sample becomes a substantial repayment. Congressional advisers estimate the overpayments accumulated during the loosely checked years in the tens of billions of dollars annually.
What the three numbers add up to
An industry that spent fifteen years optimising for documentation volume is being forcibly reoriented toward documentation accuracy, and the technology it uses is being rebuilt in the process.
The software that reads medical records used to be sold on how much it was found. Buyers now evaluate ai tools for medicare risk adjustment coding accuracy on almost the opposite basis: can any individual output be traced to the specific sentence in the clinical note that supports it, does the system flag unsupported codes for removal as readily as it surfaces missed ones, and could a hostile reviewer reconstruct the decision three years later. Systems that answer yes are winning contracts. Systems that produce conclusions without reasoning are being replaced.
There is a general principle buried in these three numbers, and it travels well beyond healthcare. Any system that pays out based on records an interested party keeps about itself will drift, not because the people involved are dishonest, but because nobody is incentivised to look for errors that cost money. The correction is never clever. It is somebody checking, regularly, in both directions, with the authority to make the answer matter.
Three numbers, one lesson, and a very large bill for learning it late.
