Somebody has to own the approval step. Often nobody does.
While working on this I spent time with several AI tools that generate clinical and administrative output for provider organizations. In more than one, the approval step was not enforced anywhere in the workflow. The system produces the output. Nothing in the product requires a named person to approve it before it moves downstream.
Ask who owns that step and the answers do not converge.
The vendor points at the contract. The product is decision support. The output was never represented as final, and the agreement puts review before submission on the customer. That is the standard structure of these agreements, and it is not a loophole. It is the actual arrangement.
The practice points at the software. They bought a system that was supposed to handle this category of work. The whole basis of the purchase was that these would stop needing scrutiny.
The person working the queue points at the workflow. Nobody told him he was the review step. He was told the system drafts the output. His queue got faster and his headcount did not grow, which is a reasonable signal that review was not the expectation.
The system of record points at nobody. It was never designed to enforce someone else's review gate.
Here is the part worth sitting with. Every one of those positions is defensible. Nobody is lying. Nobody is dodging. Each party has a real reason to believe the responsibility sat somewhere else, and the reasons are individually correct.
That is what an accountability gap actually looks like. Not negligence. A chain where every link reasonably assumed a different link was holding.
It is also a design problem more than a people problem. The step has to belong to someone. A vendor can build it in. A system of record can enforce it. Absent both, it lands on the practice by default.
Worth knowing that the law is not neutral while everyone waits for this to settle. As things stand, the practice carries it. Vendor agreements can move it further onto you before a dispute ever reaches a courtroom.
So run one of these in your own operation. Take a real output your AI produced last week and walk it forward. Who approved it? Do they know that is their job? Could they name what they were checking for?
If the chain breaks anywhere, you found it before someone else did.
Xillium makes review, approval, and responsibility explicit around AI-generated work. What that looks like in practice is on our Solutions page.
