In 2018, the FDA let an AI diagnose patients on its own. Eight years later, adoption is still limited. The gate everyone points at opened, and almost nothing came through.

The system detects diabetic retinopathy in a primary care office, and no eye doctor reads the result. April 2018, the first autonomous AI diagnostic. If regulatory approval were the thing standing between healthcare and AI, that was the moment.

It is 2026, and that has not changed.

Here is what sat in between the clearance and the exam room. The billing code that let a clinic get paid for the exam arrived January 1, 2021, nearly three years later. The health systems that did adopt it describe a six-month startup: contracting, orders and results wired into the medical record, a referral path for an abnormal result. Operators needed roughly 40 exams before image capture got fast. And the detail I keep coming back to: handing the patient an abnormal result at the visit, instead of three to five days later, was associated with better follow-up with an eye doctor.

None of that is a regulatory problem. It is scheduling, training, integration, and who owns the result.

I write about AI governance in three layers. What the model's makers build and constrain. What the organization owns. The judgment of the person at the keyboard. This is that shape at the scale of an industry: the top layer opened, and the middle layer had not been built yet.

The systems layer gets the headlines because it is the layer that can be announced. A clearance. A pilot. A payment model. The layer that decides whether any of it reaches a patient is the one inside your building, and nobody outside can build it for you.

So when the next headline says the path is clear for AI in healthcare, the question worth asking is not whether the technology works. It is whether the organization has done its part. Who runs it? How does it fit the day? Who is accountable for the result? What happens when the answer comes back abnormal?

Personally, I am not against AI in healthcare. I want it built carefully and on controls somebody owns. Eight years with this one says the unglamorous work in the middle is what decides whether any of it reaches a patient.