Everyone’s excited about Claude in healthcare. The CFOs aren’t.

Anthropic just published a healthcare page for Claude. Here’s what it didn’t say — and why it matters more than the launch.

The pitch: HIPAA-ready AI in healthcare that handles prior authorizations, claims appeals, clinical scribing, and patient triage. Faster. At scale. With fewer people stuck doing work that shouldn’t require people. The full announcement is here.

If you work in health IT, you’ve seen the reaction. The demos look real. The use cases are valid. And if you could actually get an AI model to sit inside your clinical workflow — catching documentation gaps before a claim goes out, flagging prior auth requirements at the point of order, helping write the appeal before the payer even sends the denial — that would be worth something significant.

But there’s a line on that page that deserves more attention than it’s getting.

“Trusted, HIPAA-ready AI for healthcare.”

HIPAA-ready is not the same as production-ready. And that gap is where most of this conversation currently lives.

The number CFOs are actually watching

Before we get into AI in healthcare, let’s talk about what’s been quietly getting worse.

In 2024, the average initial claim denial rate in US hospitals hit 11.8%. Four years ago, it was 10.2%. That’s not noise — it’s a consistent, multi-year trend moving in the wrong direction. For Medicare Advantage plans, the denial rate sits at 15.7%. For Medicaid managed care, 16.7%.

Now layer in the economics of a rural critical access hospital running on 2–3% operating margins — sometimes less.

Every denied claim is a cash flow event. The appeals process, which is largely manual and largely reactive, costs between $43 and $48 per claim to work through. Multiply that across a hospital’s denial volume and you get a number that, industry-wide, runs to $19.7 billion a year. Not in lost revenue. In administrative cost. To chase money the hospital already earned.

That’s the spreadsheet a CFO has open when someone walks in with an AI demo.

They’re not against the demo. They want it to work. But they’ve been burned enough times by tools that looked good in a conference room and fell apart in a live environment. Their question isn’t “can AI do this?” It’s “will it actually work in our systems, on our data, under our constraints?”

That’s a different question. And the answer depends almost entirely on something the AI vendor doesn’t control.

Why AI tools assume a world that doesn’t exist yet

Here’s the part that rarely makes it into the product announcement.

AI tools like Claude are genuinely good at what they’re built for. Give them a patient chart and a payer’s medical necessity criteria, and they can tell you whether the documentation holds up. Give them a denied claim and the appeals language from similar cases, and they can help write a response worth filing.

But they need the data to do any of that. In real time. From the right systems.

And in most hospitals, that data is locked.

The EHR holds the clinical record. The billing system holds the claim. The clinical decision support tool — if one exists — fires alerts that clinicians have learned to click through because there are too many of them and too few are relevant. None of these systems are sharing data in real time at the point of care. Most aren’t connected in any meaningful way at all.

FHIR is supposed to change that. It is changing it, slowly. But “FHIR-compatible” on a vendor’s specification sheet is not the same as bidirectional real-time data sync running in a production environment under clinical load. Anyone who has actually built this knows that the spec is the easy part. The hard part is what happens when your EHR version doesn’t fully implement the standard, or when patient data arrives in a format the AI wasn’t trained on, or when the sync is fast enough for reporting but not fast enough to be useful at the point of care.

The integration problem doesn’t come bundled with the AI tool. It has to be built. And it has to be built by people who understand both the clinical workflow and the underlying systems — not just one or the other.

The question worth asking before the demo

This isn’t an argument against AI in healthcare. The use cases are real and the potential is not small.

But the hospitals and health systems that will actually extract value from tools like Claude are the ones that solve the infrastructure problem first — or alongside. That means connecting the EHR to billing at the point of care, not retroactively. It means real-time data pipelines fast enough to matter clinically. It means FHIR implementation that goes beyond checkbox compliance into actual workflow integration. And it means doing all of this under HIPAA and HITRUST, in production, without adding complexity that makes already-stretched clinical staff less efficient.

That work is not glamorous. It doesn’t get a product page. It doesn’t demo well.

But it’s the difference between a proof of concept that impresses the board and a deployment that actually reduces denials.

The CFOs who have been watching that denial rate climb for four years aren’t skeptical of AI. They’re skeptical of AI that shows up without the infrastructure to back it up. And honestly, that skepticism is well-earned.

The question isn’t whether AI belongs in healthcare. It does.

The question is whether your systems are ready to let it work.

Sources