Hospitals across the country have spent the last few years putting AI into their billing departments. The software reads doctors’ notes, lab results and the rest of a patient’s chart, and it finds diagnoses a human coder might have missed. Every one it finds can move a hospital stay into a higher-paying category.
On September 24, the Blue Cross Blue Shield Association put a price on it. In an analysis of inpatient claims, the association estimated that changes in hospital coding added $942 million in costs to Blue plans over two years. About 70 percent of that, more than $650 million, came from secondary diagnoses, the extra conditions listed alongside the main reason a patient was admitted.
Blue Cross says the diagnoses went up and the treatment did not.
More anemia, fewer transfusions
The association’s case rests on a handful of numbers that, it argues, should move together and did not.
| Measure | Before | After |
|---|---|---|
| Inpatient stays billed as medically complex | 37% (early 2023) | 40% (late 2025) |
| Major bowel surgery claims at the highest complexity level | 10.2% | 22.7% |
| Anemia coded: top-quartile hospitals vs. the rest | 9.9% | 13.7% |
| Transfusions for those anemia patients: top-quartile vs. the rest | 19.3% | 16.9% |
Read the last two rows together. The hospitals that diagnosed anemia most often were less likely to give those patients a blood transfusion, not more. To Blue Cross, that is the signature of software finding conditions that pay, rather than doctors finding patients who are sicker. The association says it saw the same gap in ICU use, repeat surgeries and length of stay.
A separate Blue Cross study on childbirth found the same shape. At the hospitals that adopted AI coding fastest, the share of maternity patients coded with acute posthemorrhagic anemia, anemia from blood loss, rose from about 4 percent to more than 12 percent between 2022 and early 2025, while transfusion rates stayed flat. The association put the added cost at about $22 million in one year at the hospitals it studied.
The hospitals’ answer: show us the charts
The American Hospital Association rejects the word “upcoding.” Its argument is that patients really are older and sicker, and that better documentation is not the same thing as fraud. An AHA and Vizient analysis found the average hospital case mix, a measure of how complex patients are, rose about 5 percent from 2019 to 2024.
That criticism lands on a real limit. Blue Cross worked from insurance claims, not full patient charts. It cannot say whether any single added diagnosis was legitimate, overlooked for years, or invented. It does not name hospitals or AI vendors, and it stops short of proving AI caused each increase. In the AHA’s words, “the key question remains whether diagnoses are clinically supported, not whether they increase measured complexity.”
The hospitals also point straight back at the insurers. The AHA accuses commercial plans of using automated “downcoding”, software that cuts payment on claims for care that was medically necessary. Insurers say they use AI to process claims, approve some requests automatically and flag anomalies; Blue Cross and Blue Shield of Texas says it does not use AI to replace physician findings or to deny care. Doctors are not convinced. In a 2025 American Medical Association survey, 61 percent of physicians said they worried insurers would use AI to deny more prior-authorization requests.
An arms race nobody is winning
What both reports describe is two sets of algorithms aimed at the same document. One side’s software reads the chart looking for conditions that raise the payment. The other side’s software reads the claim looking for reasons to lower it. Each improvement on one side gives the other side something new to train against.
Sanghavi is not neutral either: Machinify sells claims-review software to health insurers. But the money he describes is not in dispute. He estimated the revenue cycle management market, the industry that sells billing and coding software to providers, at about $65 billion in 2025, and put health systems’ total spending on billing operations at more than $140 billion a year.
None of that spending treats a patient. Patients see it later, in premiums and in denials. Total premiums for employer-sponsored family coverage averaged $26,993 in 2025. On HealthCare.gov plans in 2024, insurers denied 19 percent of in-network claims and 37 percent of out-of-network claims, though not every denial was improper.
Michigan wants to ban both sides
The first politician to take on both sides at once may be Michigan House Speaker Matt Hall, a Republican from Richland. On October 8, Hall said he wants to ban hospitals from using AI to set prices, including to calculate the maximum price for a service, and to stop health insurers from using AI to deny patients’ claims. He said small businesses keep telling his office their costs are rising.
Hall has not released bill text, and he did not say when a vote might come. But he has put the idea in plain terms: if each side’s AI is mostly being used against the other, maybe neither should have it. “Everyone is going to have to give a little bit,” he said.
The Takeaway
The Blue Cross Blue Shield Association estimates that hospital coding changes added $942 million to Blue plans over two years, more than $650 million of it from secondary diagnoses, while its claims data showed no matching rise in treatment. Its sharpest example: hospitals that coded anemia most often transfused those patients less, not more. The association blames AI coding tools, which it says more than 60% of hospital systems now use, but it worked from claims, not charts, and does not prove AI caused each increase.
The American Hospital Association says patients are genuinely sicker and accuses insurers of automated downcoding of their own. Michigan’s House Speaker wants to bar AI from both hospital pricing and insurer claim denials. What to watch: whether Hall’s proposal becomes an actual bill, whether any regulator audits the charts behind the claims, and whether the next Blue Cross report names the hospitals or the vendors.