Behind every large American health insurer stands an industry most people have never heard of: firms whose entire business is reading medical records. Millions of charts a year flow through their coders and, increasingly, their AI systems, translating the messy prose of clinical notes into the standardised diagnosis codes that determine billions in payments.
It is outsourcing at its most consequential: intimate documents, high stakes, invisible workers. And right now, the entire sector is being remade by the same force reshaping the insurers it serves: audits.
Why the industry exists
The demand is structural. Insurers in the Medicare Advantage programme are paid according to risk scores built from members’ documented diagnoses, which means every year they must review documentation at a scale no internal team comfortably handles: tens of millions of encounters, each potentially containing conditions that were treated but never coded, or coded but never properly supported.
So a service industry grew. Certified medical coders, working for specialist firms across the United States and around the world, read charts against strict documentation criteria. Technology firms layered AI on top, reading first and routing humans to what needs judgment. The result is a global pipeline through which a diabetic’s clinic visit in Ohio may be coded by a specialist thousands of miles away, validated by software, and submitted to the US government, all inside a payment deadline.
At its best, the pipeline is genuinely valuable: real illnesses, documented by rushed clinicians, that would otherwise vanish from the record get counted, and the funding to manage them follows. At its worst, the past two years demonstrated, it becomes something else.
The reckoning
The uncomfortable discovery of the enforcement era is that a reading industry inherits the incentives of whoever pays for the reading. When clients wanted more codable diagnoses found, the industry optimised for finding them, and only them. Federal auditors reviewing three insurance plans reported in March 2026 that 81 to 91 percent of certain sampled high-risk codes lacked adequate supporting documentation. A major insurer settled with the Department of Justice for 117.7 million dollars over chart-review programmes that added diagnoses prolifically while almost never removing unsupported ones, programmes substantially executed through exactly this reading pipeline.
The audits that followed did not distinguish between insurer and contractor. With roughly two thousand federal coders now re-checking submitted diagnoses on a quarterly cycle, and error rates extrapolated across entire contracts, every chart the industry reads is now, potentially, evidence.
The new shape of the sector
The response is visible in how buyers now evaluate the field. Assessments of risk adjustment outsourcing partners have inverted their criteria: where the old scorecards ranked cost per chart and codes found, the new ones lead with validated accuracy, bidirectional review, and audit performance.
Concretely, the firms gaining ground share a profile. They review in both directions, flagging unsupported codes for removal with the same rigour they apply to finding missed ones, because one-way review is now the recognised signature of bad faith. They keep evidence trails: every code traceable to the exact language in the clinical note that supports it, and to the credentialed human who confirmed it. Their AI explains itself, emitting justifications an auditor can follow rather than bare conclusions. And they submit to adversarial testing, letting prospective clients sample and re-code their output the way a federal reviewer would.
The firms losing ground are the ones still selling volume: charts per hour, codes per chart, price per unit. In a market where a single extrapolated audit can claw back a meaningful fraction of a contract, cheap reading that cannot survive re-reading is the most expensive service an insurer can buy.
The wider mirror
The story is worth telling beyond healthcare because outsourced judgment is everywhere: content moderation, loan processing, claims handling, data labelling for AI. In every case, the same quiet law applies. The contractor’s output will drift toward the client’s incentives unless verification is built into the relationship, and when regulators eventually arrive, they audit the output, not the org chart.
The mature version of any such industry looks like what medical chart reading is becoming: work product that carries its own evidence, quality measured in both directions, and a paper trail built for hostile review from the first day, not assembled in panic after the letter arrives.
An invisible industry, forced to make its work inspectable, is quietly writing the standard for every other invisible industry. That is worth a moment of visibility.
