Who Owns This Once It Is Live? The AI Job Nobody Budgets For

The tool went live in March. It worked, the numbers were good, everyone moved on. In November a locum reading back through a chart mentions, almost in passing, that the last few notes have looked thin. Somebody checks. The notes have been drifting for a while.

How long is a while? Nobody can say, because nobody was looking, because nothing had ever gone wrong in a way that asked to be looked at.

Why does nobody notice when AI gets worse?

Ordinary software fails loudly. A form breaks, an integration times out, a page throws an error, and the failure announces itself: someone files a ticket, someone fixes it, the clock on the outage is measured in hours. The system is designed so that being broken is a visible state.

AI does not do that. When an AI system degrades it keeps answering, in the same tone, at the same speed, with the same confidence, and the answers are simply worse. There is no error, no ticket, no alert. A failure that looks exactly like success is the one you will keep funding the longest, because nothing about it ever asks to be reviewed.

Nothing has to break for this to happen. The world moves underneath the system: a new clinician joins with a different consulting style, your EMR template changes, a payer updates its rules, your staff start phrasing things differently, the vendor ships a new model version on a Tuesday with no changelog you ever see. Each shift is small. None of them is an error. Together they are why a system that was 94 percent right in March is something else entirely by November, and why an AI that demos beautifully can still die in production.

What ownership actually means

Most practices believe they have an owner because a name comes to mind when you ask. That is not ownership, that is an association. Ownership is four specific things, and the useful exercise is checking how many you actually have.

Four rungs, and most projects sit on the first

1. 1

   Nobody

   The tool went live and everyone moved on. Degradation is discovered in a chart, at an unknown age.

2. 2

   A name, and nothing else

   Someone would be blamed if it went wrong, but they watch no number and have no scheduled moment to look. This is the most common state, and it feels like ownership.

3. 3

   A name and a number

   Somebody knows what good looks like and can tell you this month against last month. A slide is now visible while it is still small.

4. 4

   A name, a number, a date, and a switch Start here

   They look on a fixed day, and they can turn it off on their own authority. Only at this rung can the system be stopped before it does damage.

The fourth item is the one that gets negotiated away, and it is the one that matters most. An owner who has to convene a meeting to pause a misbehaving system is not an owner, they are a witness. Give them the authority to switch it off unilaterally and you will almost never see it used, because knowing the switch exists changes how carefully everyone treats what sits behind it.

What does the owner actually do?

Less than people fear, which is exactly why it goes undone: it is too small to feel like a job and too easy to skip.

On a fixed day each month the owner pulls a fixed number of the system’s outputs, twenty is plenty for most practices, and compares them against what a competent clinician would have produced. They record one number. They read the corrections their team made during the month, because a rising correction rate is the earliest signal there is. Then they make one of three calls: leave it alone, narrow what it is allowed to do, or stop it.

That is the whole job. It takes a couple of hours. It is the difference between a system you can trust in year two and one you quietly stopped trusting in month seven without ever deciding to.

Be honest with yourself about what it does not do. Twenty samples a month will reliably catch a slide, because a slide shows up across many outputs at once. It will not catch the rare, expensive, one-in-a-thousand answer, and no amount of sampling will. That case is what the approval gate is for, and it is the reason the gate should outlive the enthusiasm to remove it.

The reason it works is that it is scheduled rather than triggered. A triggered check never happens, because the trigger for silent degradation does not exist. This is the same argument as proving the thing works before you scale it, except that proof is not an event you hold once at launch. It has a shelf life, and the monthly sample is how you renew it.

An owner who cannot switch it off is not an owner. They are the person who will be asked why nobody switched it off.

Who should it be?

The instinct is to give it to whoever is most comfortable with technology. That instinct produces the second rung: a name, and nothing else.

Give it to the person who owns the outcome. If the tool drafts clinical notes, the owner is a clinician, because they are the only person in the building who can look at a note and know within seconds whether it is any good. That judgment is the entire job, and it cannot be delegated to someone who lacks it. Technical help is something they call on, the way they call an accountant without becoming one.

This is also the honest answer to the vendor question. A vendor cannot own it: they cannot tell you whether a note was clinically sound, and they have no incentive to recommend switching their own product off. When a vendor offers to handle monitoring, the question to ask is who inside your clinic would notice if they stopped. It is the same thinking as onboarding an AI the way you would onboard a hire, where the point of a probation period is that somebody specific is watching.

What it costs, and where it belongs

A couple of hours a month of your workflow owner’s time, plus the running cost of the software itself. That is a real line and a small one, and it belongs in the approval, not in a discussion six months later about why the thing drifted.

The mistake is not underfunding it. The mistake is that ownership hours are nobody’s hours, so they are free right up until the month they turn out to have been essential. Put the name and the hours in the same document as the build cost, where the full running cost of an AI system already lives.

What to do this week

Take the AI you already run, and write one name beside it. Then a number, then a date each month, then answer whether that person can switch it off. It takes ten minutes and it is not really a documentation exercise. It is a test of whether the thing was ever actually handed to anyone.

If naming the person requires a meeting, you have your finding, and it is a more useful one than any dashboard would have given you. That answer belongs in the same conversation as what you tell the partners, because a project with a named owner reads as governance and a project without one reads as a hope. In Ontario it is also closer to an expectation than a nicety: the privacy commissioner recommended in January 2026 that clinics have oversight with the authority to approve, pause or decommission an AI deployment, which is this rung four, written as guidance.

---

Running AI in your clinic with nobody watching it? Book a free consult and we will map the ownership: the person, the number, the cadence, and the switch, for whatever you already have live. See more on PHIPA and Guardrails.