You can buy it. Can you turn it off?
Healthcare AI has a procurement path with owners, dates and approvals, and no withdrawal path with the same, so the decision to restrict or switch off a failing tool has no owner, no trigger and no clock, and a tool that should be paused keeps running by default.
Ask the people who run your busiest AI tool a simple question: who is allowed to switch it off, on what evidence, and how fast? You will get three different answers, or a pause while everyone looks at everyone else.
The buying decision, by contrast, is fully documented. There is a request, a security review, a committee, a signature and a date. Procurement is a path with owners and timestamps at every step, because a purchase moves money and money demands a trail.
The withdrawal decision has none of that. Nobody signed up to be the person who takes a live clinical tool away from the clinicians using it. So the path out is missing exactly where the path in is thorough.
This is the gap that turns a monitoring plan into decoration. You can measure a tool’s drift beautifully, chart it, review it quarterly — and still not have answered the only question monitoring exists to serve: when the number crosses the line, who acts, and what do they do? A dashboard that cannot end in a decision is a more expensive way of watching something fail.
Name the failure precisely. A tool degrades. The metric moves. It gets noticed at the next committee, which is in three weeks. The committee agrees it is concerning and asks for more data, because no one in the room has the authority to restrict a tool a hundred clinicians now depend on, and no one wants to be the one who guessed wrong. So it stays live. Not because anyone decided it should — because switching it off was nobody’s job, and default is a decision too.
The frameworks already put the off switch where it belongs. The NIST AI Risk Management Framework, published by NIST in January 2023, gives its MANAGE function the response, recovery and decommissioning of a system — the acts of stopping, not just the watching. The Joint Commission’s September 2025 guidance with the Coalition for Health AI expects a defined path to restrict or withdraw a tool in use, not merely to approve one. Neither treats withdrawal as an emergency improvised on the day. The control mapping sits in the reference library.
The durable version writes the withdrawal path at purchase, in the same document as the purchase. Before a tool goes live, three things are named: the person who holds the authority to restrict or switch it off, the specific readings that oblige them to act rather than merely permit it, and the number of hours from trigger to action. The contract is where this belongs, because a control that lives only in a policy binds your staff, and a control that lives in the agreement binds the vendor too — the reason the Model Governance Rider requires a withdrawal trigger, not just a governance record.
Monday
Pick your highest-volume AI tool and try to write one sentence: who can restrict or withdraw it, what evidence obliges them to, and how many hours it takes. If you can write it from memory, you are ahead of almost everyone. If you cannot, you have found the thing to fix first — not more monitoring, but a name, a trigger and a clock attached to the tool you would least like to be unable to stop.
One operational argument a week
The Operations Edge lands each Monday: a hook, one thing to use before lunch, and the full argument here in the archive. No vendor sponsorship, ever.
The instruments behind the writing
Every framework in the series is published as a working file: registers, protocols, audit rubrics and unit-economics models, sized to be used rather than admired.
See the toolkits The library← Trained in March, hired in July
Published under the Healthcare AI Institute editorial standard.
Written and reviewed against the standard by a physician-executive whose career spans three national healthcare systems. Last reviewed on 2026-09-24.
Drafted as issue 12 of The Operations Edge, argued from operating experience rather than a dataset. No effect-size figure is used because none with a dated, named source applied; the two framework references are cited by publisher and date.
The Institute accepts no vendor sponsorship, holds no vendor equity and takes no referral fees.