AI coding added $942M to claims, with no extra care

Hospitals pointed AI at their medical coding. More patients started showing up in the records as medically complex. The treatment did not change. Blue Cross plans put the extra bill at $942 million over two years, and the number is worth your attention even if you never touch a claim — because it is the clearest case yet of what happens when both parties to a transaction automate their side of the argument.

What actually happened

Per the Blue Cross Blue Shield Association analysis, reported by Fierce Healthcare:

• The complex-case share moved from 37% to 40% of inpatient stays billed to Blue plan members, from the start of 2023 to the end of 2025.
• That drift cost member plans an estimated $942 million over two years. About $653 million of it traced to secondary diagnoses, roughly $11,000 per excess complex case.
• Around 70% of the increase came from more than 55,000 additional cases where a secondary diagnosis pushed the claim into a higher-severity, higher-reimbursement DRG.
• BCBSA's argument is a disconnect, not fraud: a sharp rise in documented complexity with no evidence of a corresponding change in care delivered.
• Hospitals disagree. The American Hospital Association said in August that rising coding intensity also reflects an older, sicker patient population and more accurate documentation of conditions that were always there. That is a real counter-argument, and the data as published does not settle it.

TechCrunch notes the obvious next move: insurers are automating their side too. BCBSA's Luke Chalker described the current state as one-sided. It will not stay that way.

Why automated billing matters for your business

Forget healthcare for a second. The mechanic here is general, and it is coming for your invoices.

An AI that reads a record and picks the highest defensible category is doing exactly what it was built to do. Nobody wrote "inflate the bill" in a prompt. The tool was told to find every documentable condition, and it found them — thoroughly, consistently, at a scale no human coder sustained. The outcome looks like a policy change because the throughput of an existing incentive changed. That is the thing to internalize: automation does not introduce new incentives, it removes the friction that was quietly capping them.

Two practical consequences. First, audit your own automation for drift in the direction of your incentives. If your AI classifies support tickets, prices quotes, picks a shipping tier, or assigns a billing code, plot its distribution over time against a pre-automation baseline. A slow slide toward the outcome that happens to pay you more is the signal, and you will not see it in a spot check of ten records. Second, expect your counterparties to automate their review, and keep the evidence that survives it. When a machine reads your invoice, the defensible line item is the one with a linked artifact behind it — a timestamped log, a signed-off scope, a diff. "Trust us, it was complex" loses to a query.

The endgame is not a war between two models. It is that whoever has the better audit trail wins the dispute, cheaply and repeatedly.

Sources: Fierce Healthcare, TechCrunch.