Blue Cross: AI Medical Coding Added USD 942m in Costs

On 24 September 2026 the Blue Cross Blue Shield Association put a price on hospitals' use of AI coding tools. Its analysis attributes about USD 942 million in additional cost across its plans over two years, measured against a 2023 baseline.¹

The claim that makes it interesting is not the money. It is that the patients did not appear to get different care. The association's figures describe claims that classified illnesses as more complex in 2024 and 2025 than in 2023, with "no evidence that the patients had received different treatment from the hospitals".²

What the analysis says

The association represents 31 independent Blue Cross and Blue Shield insurers covering more than 100 million people, and the work draws on inpatient billing data from its member plans.¹

Two figures matter. The headline total is USD 942 million in additional cost over two years against the 2023 baseline. Within that, USD 653 million came from more frequent billing for secondary conditions between 2024 and 2025.¹ The arithmetic leaves about USD 289 million attributable to everything else in the coding change, a split the wire coverage did not separate.

The association's own framing of the test is a clinical one. "If patients are truly sicker, we'd expect to see more treatment," its senior vice president of product and data science, Luke Chalker, said.¹ The reported finding is that the extra diagnoses arrived without the extra treatment.

It is worth naming what kind of document this is. It is a payer-side analysis of provider behaviour, published by an association whose members pay the claims in question, using a baseline year it selected. None of that makes the figures wrong, and all of it belongs in any account of them.

The hospitals' answer has numbers in it

The American Hospital Association published a fact sheet in 2026 arguing the case before the paper appeared. Claims that AI drove coding intensity, it says, "not only ignore the numerous underlying drivers of changes in coding intensity, including an aging patient population, but also overlook the legal, ethical and contractual obligations on hospitals to ensure appropriate coding".³

Its counter-evidence is a trend, not an anecdote. Hospital case-mix index, a standard measure of how sick admitted patients are, "rose by about 5% between 2019 and 2024".³ It names three drivers: an aging population, more chronic disease, and the continued movement of less complex care to outpatient settings.

Then it turns the accusation around. "In 2025, MedPAC reported that upcoding contributed to $40 billion in overpayments to Medicare Advantage plans", and "in March 2026, one payer settled a lawsuit with the DOJ on upcoding allegations".³ Both are the hospital association's characterisations, of the Medicare Payment Advisory Commission's findings and of a Department of Justice settlement. This article could not read the commission's own chapter, which is published in a form it could not extract. So the figure is quoted as the association's account rather than verified at source.

Nobody can say what the right number of codes is

Underneath both arguments sits a problem neither side raises. There is no agreed correct answer to code against.

A paper posted on 21 September 2026 had two teams code the same 110 clinical encounters. They agreed on 73 percent of codes for the same note, and after an independent clinical audit removed erroneous codes, agreement rose only to 77 percent.⁴ The authors model the remainder as "coding style", a site-specific policy about what to code and how much to document.

That is a preprint rather than a peer-reviewed paper, and its subject is model evaluation rather than billing. It still establishes something both sides need: a gap of roughly a quarter of codes between two competent human teams on identical notes. A rise in secondary diagnoses is consistent with upcoding, and equally consistent with a documentation policy changing.

The clinical research never looked at the bill

The evidence base for ambient scribes, the tools that draft notes from the consultation, is real and points somewhere else entirely.

A study of 48 physicians over three months reported reductions in task load and burnout after the tools were introduced.⁵ A systematic review of 11 studies found that nine of ten reported improvements in at least one efficiency metric and seven found positive effects on clinician wellbeing.⁶

What that literature does not contain is any evaluation of note length, coding accuracy, diagnosis capture or billing outcomes.⁶ The tools were studied as a burden-reduction technology. They are now being measured as a billing one, and the two bodies of evidence have never met.

How to read the next figure like this

Ask who produced it and what baseline they chose. A two-year comparison against 2023 is a choice, and a different starting year moves the total.

Ask whether the claim is about codes or about care. "More complex coding" is an observation. "No different treatment" is the inference, and it rests on treatment proxies such as transfusion rates rather than on patient outcomes.

Ask what the other side's number is. Here it is a 5 percent case-mix rise and a USD 40 billion finding against insurers' own plans, which does not refute the USD 942 million but does establish that both parties code to their advantage.

And treat the coding literature as the referee it cannot yet be. Until someone measures what ambient documentation does to code counts in a controlled comparison, the honest answer to who is right is that the study establishing it has not been done.