A hemoglobin of 9 after bowel surgery surprises no surgeon. For most of my career it also didn't surprise the billing office. Now an algorithm reads the lab value, proposes acute posthemorrhagic anemia, and the same operation moves up a payment tier.

The Blue Cross Blue Shield Association says that move, repeated tens of thousands of times, cost its plans $942 million over two years. Its evidence is that the diagnoses went up and the transfusions didn't.

Inpatient payment runs on diagnosis related groups. Many surgical and medical DRGs come in three steps. The base case, the case with a complication or comorbidity (a CC), and the case with a major one (an MCC). Medicare sets the weights, and many commercial contracts pay off the same grouper. One secondary diagnosis can move a stay up a step, and the step is worth thousands of dollars.

Anemia, hyponatremia and acidosis are common and easy to miss in a busy note. For decades hospitals hired clinical documentation specialists to find them. Every hospitalist knows the query asking whether a sodium of 131 is clinically significant. AI does the same job without getting tired, on every chart, including the old ones.

BCBSA is the trade group for the independent Blue plans. The analysis comes from Luke Chalker, its senior vice president of product and data science, with clinical commentary from Razia Hashmi, its vice president of clinical affairs. It is at least the second report in this vein this year. An earlier one attributed $22 million in maternity costs to a single anemia code.

The interest is plain. Every dollar of coding intensity is a dollar the plans pay. That does not make the numbers wrong. It explains why the counterfactual is built the way it is, with 2023 as the honest baseline and everything since as drift.

The American Hospital Association is no more neutral. Its July fact sheet says case mix rose about 5% from 2019 to 2024 because patients are older and sicker, calls payer downcoding programs automated and unilateral, and points out that insurers claim sicker members when it raises their Medicare Advantage risk scores. MedPAC has estimated that coding intensity in Medicare Advantage runs to tens of billions a year. Both sides know how this game works because both sides play it.

The Blue Cross report reads as if hospitals brought software to a paper fight. They didn't. In October 2025 Cigna launched a policy to automatically knock high level office visit codes down a level for clinicians whose billing looked like outliers against their peers. It paused after California regulators got involved. In April my own state's Insurance Administration fined Cigna $80,000 and ordered it to stop automatic downcoding and ask for records instead. Aetna's Level of Severity policy for Medicare Advantage, pushed to January 1 after hospital objections, pays many one to four midnight emergency admissions at a reduced rate close to observation.

Notice the asymmetry. The hospital's software reads the whole chart, every note and every lab. Most automated downcoding reads the claim. The AHA complains that payers cut codes without reviewing the documentation at all, and the Maryland order amounts to an instruction to open the chart before swinging.

HCA's chief financial officer says hospitals are behind the payers on AI claims processing. The revenue chief at NYC Health + Hospitals called it a rock 'em, sock 'em robot fight in which nobody wins. Hospitals think they are losing. Insurers think they are losing. That is usually a sign both sides are spending a lot.

An immune system is an arms race with a memory. A pathogen finds a surface the body doesn't recognize. The body builds an antibody. The pathogen mutates. Neither side wins permanently, and the Red Queen keeps them both running to stay in place.

Computer scientists borrowed the idea long before large language models. In 1994 Stephanie Forrest and colleagues at the University of New Mexico built software that learned what normal looked like on a machine (the self) and flagged everything else. In 2016 DARPA put seven autonomous systems on an air gapped network in Las Vegas for nearly twelve hours, finding flaws, patching their own software and probing each other's, with no human at the keyboard. Offense and defense at machine speed is no longer a thought experiment in security. Billing is just a slower host.

Billing now works the same way, with agents on both sides. The hospital's model reads the chart and presents every true diagnosis it can defend. The payer's model reads the claim and flags every pattern that looks inflated. Each learns from the other's last move. The hospital's next model documents the fluid restriction. The payer's next model asks for the sodium trend.

Immunologists know the two ways this ends. Tolerance, where the system learns what belongs to the self and stops attacking it. Or autoimmunity, where the response itself becomes the disease. In billing, the inflammation is administrative cost, and the host who pays for it is the patient, in premiums and in time.

There is a hopeful version. Two agents negotiating every stay, with full access to the chart and the contract, could discover something close to a fair price for that stay faster than any rate committee (I sit on one, so I say this with some humility). Scale it up and a hospital submits a thousand claims at once, the insurer's agent scores the batch against the hospital's own history and against its peers, and a price clears for the whole lot. The pieces already exist. Cigna's policy flagged outliers against peers. CMS mails physicians Comparative Billing Reports that do the same thing on paper. Usage based auto insurance reprices drivers monthly from telematics. The staircase would stop mattering because the price would follow the care instead of the codes. The cost is predictability. A CFO who can't forecast next month's reimbursement will want a very good reason to sign up. Medicare has a blunter tool. When coding intensity rose after the 2007 switch to MS-DRGs, CMS simply clawed the gains back through years of payment adjustments. Tolerance by decree.

The less hopeful version is the one we have now. Each side buys a better weapon, the other side buys a better shield, and the vendors selling to both are the only ones with a clean balance sheet.