Blue Cross Blue Shield Association estimates that AI-assisted coding at hospitals added $942 million in costs for its health plans between 2023 and 2025. The headline rounds that to "close to $1 billion." The analysis itself uses $942 million.
CNBC is carrying both the estimate and the hospital industry's reply. Luke Chalker, the association's senior vice president of product and data science, said roughly 70 percent of the billing it flagged, or $653 million, was tied to extra diagnoses that arrived with no change in care. Divide 653 by 942 and the share is 69.3 percent. "Roughly 70 percent" is that ratio, rounded. Both dollar amounts are the association's tally over three years. The percentage is a share of their tally. It is not a share of every hospital dollar in the country.
A count of systems is not a count of dollars
The association wrote that this "complex coding" grew during a period when 60 percent of hospital systems started using AI coding tools. Sixty percent is systems. Nine hundred and forty-two million is dollars. Multiplying them would pretend they are the same unit.
Chalker said that in many cases "the diagnoses are legitimate and weren't captured" before. He also said some conditions "clinically just don't really matter and don't influence the patient's care," and still allow another code, and a higher payment.
The American Hospital Association answered CNBC directly. Patients "are older and more clinically complex," it said, and the tools help providers "appropriately capture their patients' conditions to aid in care planning." It said the insurer's write-up "lacks the context needed to meaningfully assess" quality, access, or spending.
8.2 percent is a 2027 forecast
Marsh expects the cost of health coverage per employee to rise 8.2 percent on average in 2027, which would be the steepest increase since 2003. That forecast is a price for a year that has not begun. It is not a piece of the $942 million, and the two should not be divided. One is a three-year coding sum. The other is a next-year premium path.
Marisa Greenwald, a partner at Oliver Wyman who advises both hospitals and insurers, said the tools are already helping clinicians, and that "it's hard to disentangle how much of it is better accuracy." Misuse, user error, and overcoding stay in the picture. If providers code more with AI and insurers deny more with AI of their own, "the arms race is poised to exacerbate." She hopes both sides notice they are "adding cost and burden" to a system that already has plenty. Ten coders, she said, can read one chart and code it ten ways.
Wages are already losing a separate race with prices. Inflation at 3.4 percent has outrun 3.1 percent wage gains. The trades hiring around new data centres is a different bill again. Welder and pipefitter postings on those sites are up 164 percent.
A clinician CNBC quotes, Moldovan, worries that a diagnosis sticks in the record because a model wrote it. "We're in danger of trusting it too much." Christopher Whaley at Brown told CNBC the test is not whether spending falls. If the software only "shuffle[s] the cards a little bit more" so one side comes out ahead "and not do anything to patients," it is probably not worth the money.
The open split is inside the $653 million. Chalker described legitimate codes that used to be missed, and codes that do not change care. The published analysis does not say how many dollars sit in each pile. Greenwald's line that ten coders can disagree on one chart is the human version of the same uncertainty. If the $653 million were entirely missed diagnoses, the insurer's "questionable" label would be the complaint of a payer billed for real illness. If it were entirely codes that do not change care, the hospitals' reply about older, more complex patients would not explain the dollars. BCBSA kept both descriptions. It did not weight them. The 8.2 percent Marsh forecast for 2027 does not weight them either. It is a premium path, dated to a year the $942 million has already finished.