The first large-scale, measurable return on AI investment in American healthcare is a bigger bill for the patient.
That’s the read I take from the Blue Cross Blue Shield Association’s analysis, which found that hospitals’ use of AI tools during insurance claims submission added $942 million in healthcare spending over two years. Not $942 million in diagnostics. Not $942 million in earlier cancer detection. Paperwork. The claim is that AI has inflated billed complexity without a matching change in the care delivered. Same patient, same treatment, more expensive description of it.
The New York Times framed it as a fight between hospital AI and insurer AI, both sides escalating a feud that predates any of this technology by decades. TechCrunch picked it up too. And as someone who spends most of the week testing AI tools and writing up whether they actually do the thing they promise, I want to be precise about what’s remarkable here, because it isn’t the dollar figure.
The tools are working exactly as designed
Most AI deployments I review fail because the model underperforms the pitch. This is the opposite problem. Revenue cycle AI is, by all appearances, doing its job well. Give a system the objective “maximize reimbursement within the bounds of the coding rules” and point it at a mountain of clinical documentation, and it will find justifications a tired human coder would miss. That’s not a malfunction. That’s product-market fit.
Which is why I get twitchy when vendors in this category talk about efficiency. Efficiency at what? Every AI tool encodes an objective function, and in billing software that objective is money extracted, not accuracy achieved. The insurers, for their part, are running their own models to deny and downcode at speed. Two automated systems optimizing against each other, both technically performing, neither pointed at a patient outcome.
The economics of that are grim in a specific way. When humans argued over claims, friction was a natural brake. Appeals took time. Staffing costs limited how aggressively either side could fight. Software removes the brake on both ends. The volume of the dispute goes up, the cost of the dispute goes up, and the thing being disputed stays exactly the same.
Consider the source, then consider the mechanism
I’d be a bad reviewer if I took the number at face value. This analysis comes from insurers, who have an obvious interest in casting hospital AI as the villain and a long history of using cost narratives to justify their own behavior. Insurers also deploy AI to deny claims, and that side of the ledger doesn’t show up in a figure like $942 million. Treat the number as a directional signal from an interested party, not as a neutral audit.
The mechanism, though, is completely believable, and that’s what should worry anyone evaluating AI tools for a living. You don’t need to trust Blue Cross to believe that a language model trained to find billable complexity in a chart will find billable complexity in a chart. That’s the easiest prediction in this entire field.
What this says about the rest of the AI market
Healthcare billing is a preview, not an outlier. Any industry with an adversarial paperwork layer is about to get the same treatment. Legal discovery. Insurance underwriting. Procurement. Tax preparation. Anywhere two parties negotiate through documents, both sides are about to arm themselves, and the aggregate cost of the negotiation will rise while the underlying transaction stays flat.
The uncomfortable part for AI boosters is that this counts as success by most vendor metrics. Higher throughput, measurable revenue lift, clear ROI for the buyer. If you’re a hospital CFO, $942 million distributed across your peer group looks like the tool paid for itself many times over. The cost lands on someone who never signed the contract and never evaluated the software.
That’s the pattern I keep running into in this space: AI tools are getting genuinely good at zero-sum tasks well before they get good at positive-sum ones. Writing a more persuasive claim is easier than reading a scan correctly. Drafting an appeal is easier than coordinating care. So the money goes where the capability is, and the capability is currently in the part of healthcare that produces nothing.
My verdict
If you’re buying AI for a healthcare organization, ask the vendor a blunt question: what is the model optimizing, and who pays when it succeeds? Billing tools will answer with revenue numbers, which is the honest answer, and you should decide with open eyes whether that’s the deployment you want to be known for.
And if you’re a patient, the useful takeaway is that the AI in your care pathway is probably not in the exam room. It’s in the argument about the exam room, on both sides, and it’s very good at its job.
đź•’ Published: