Shadow AI in a Clinic: The Adoption Decision Your Staff Already Made For You
Your staff already made the call on AI. Not you, not a partners' meeting, one paste into a personal chat window at a time. In a clinic each of those pastes is an unlogged disclosure of personal health information. A short exposure check shows where you stand, and the fix is smaller than a policy document.
In this article
Somewhere in your practice, right now, someone is pasting a piece of a patient encounter into an AI tool you have never heard of. Nobody asked permission. Nobody needed to, and nobody meant any harm by it.
The decision that already got made
Every AI rollout plan starts from the same assumption: climbing the adoption ladder is a choice still ahead of you. It already happened. Not officially, not at a partners’ meeting, but for real, one paste into a personal chat window at a time. A clinician turns rough notes into a readable letter with a free account. Someone asks a chatbot to summarise a long specialist reply nobody had time to read. Someone drafts a difficult message to a family. None of it shows up in a rollout plan, because none of it went through one.
This has a name, shadow AI, and the reason it matters is not that anyone did something wrong. It is that the tools people reached for were never built for what they are being asked to hold.
None of this is hypothetical. A receptionist pastes a patient’s message into a personal account to draft a calmer reply. A physiotherapist runs their scrappy session notes through a chatbot before writing the insurer report. Someone uploads a scanned consult letter to ask what a paragraph actually means. Every one of these is a real task, done fast, with a tool that has no idea it is holding a person’s health information, and with no record anywhere of what it was told or how long that will be kept.
Before the fix, the honest question: how much of this is already happening in your practice, and would you know if it were?
Decision helper
What is your shadow AI exposure?
1. Does your practice have an approved AI tool staff are actually expected to use?
2. If someone pasted a patient's history into an AI tool to draft a letter faster, would you know?
3. How does a clinician get help with something sensitive, a long consult letter, a difficult message home?
4. What happens if someone is found using an unapproved AI tool for work?
My read
Why banning it does not work
The instinct, once this becomes visible, is to shut it down. Block the sites, send a memo, done. It rarely survives contact with how people actually work.
A ban removes two things at once: your visibility into what is happening, since nobody reports what they know is against the rules, and the real relief clinicians were getting from it. It does not remove the behaviour. The notes still need finishing, the letter is still due, and the tool is still one tab away on a personal phone you do not control. Most bans just push the same paste into a private browser window instead of a work one. It is the same problem behind why an officially provided AI tool goes unused, pointed the other direction: people route around whatever gets in the way of finishing the task, whether that is a clunky sanctioned tool or a rule with nothing to replace it.
What actually closes the gap is not prohibition, it is a faster, equally good, sanctioned way to do the same thing.
One sanctioned tool with Canadian residency, for the task staff actually reach for AI to do.
There is no reason left to go around it.
A staff notice says AI tools are not allowed. No approved alternative exists.
The notes still need finishing tonight. The same paste happens anyway, on a phone you cannot see.
What actually reduces the risk
Two small things do most of the work.
A one-page acceptable-use note, not a policy document. Plain language: what is fine to paste into a general AI tool, staff notices, generic drafts, anything with nothing identifying in it, and what never leaves the clinic, which is anything at all about a patient. That line is easier to remember than any policy, and it is the right line.
One sanctioned tool for the task your staff already reach for AI to do, which in almost every practice is note-writing, wired with the same human checkpoint discipline you would want on anything trusted with a patient’s information. It does not need to cover everything on day one, it needs to cover the one thing already happening with the least visibility.
The honest part: not every task needs a product built for it
This is not a pitch for building AI into everything you own. Some of what your staff paste into a general AI tool today is genuinely fine there: a staff notice, a first draft of a policy, a summary of something public. The line is not the tool, it is the data, and in a clinic the line is unusually easy to draw. The moment a patient is involved, it needs a path with real accountability behind it, not because the AI is dangerous, but because nobody is watching what happens to it afterward and the obligation is yours either way.
You did not skip the AI decision. Your staff made it for you, one paste at a time, at nine at night, without anyone noticing there was a decision to make.
The decision that is still yours
Your team already adopted AI. What do you do about it?
Ban it
The paste still happens, just in a private window you cannot see.
Ignore it
Patient histories and notes keep leaving, with no record anywhere and the obligation still yours.
Channel it
One page of plain rules, one sanctioned tool for the riskiest task.
What to do Monday
Do not commission a full AI policy or an audit of everyone’s browser history. Both take too long, and both feel adversarial to people who were only trying to get home before ten. Write the one page. Name the one sanctioned tool for the one task with the most exposure. Start with that single workflow, the same restraint that works for choosing your first real AI project works here too.
Not sure where your practice’s exposure actually is? Book a free consult and we will map the one task worth fixing first, and what a sanctioned, PHIPA-ready path for it would look like. See more on PHIPA and Guardrails.
Frequently asked questions
What is shadow AI?
It is AI tools your staff are already using for work without approval or visibility, most often a personal ChatGPT or Claude account used to draft a letter, summarise a consult, or turn rough notes into something presentable. It is almost never malicious. It usually means a clinician had twelve notes outstanding at nine at night and nobody had given them a sanctioned alternative.
Is it bad if my team uses ChatGPT or similar tools without permission?
It depends entirely on what goes into it. Drafting a generic policy or a staff notice is low risk. The real exposure is a patient history, a consult note, or anything identifying leaving your control with no record of where it went or how long it is kept. In Ontario that is not a policy breach in the abstract, it is a disclosure of personal health information that you, not the tool, are accountable for. The tool is not the problem. The data is.
Should we just ban AI tools until we have a policy?
No, and it rarely works even as a stopgap. Banning removes your visibility and the relief clinicians were already getting, but it does not remove the behaviour. The notes still need writing tonight. A sanctioned alternative with Canadian residency, at least as fast as the public one, closes the gap. A ban alone just moves the paste into a private window.
What is the fastest way to reduce shadow AI risk?
Write one page, not a policy document, that says plainly what is fine to paste into a general AI tool and what never leaves the clinic, then name one approved tool for the task your staff are actually reaching for. In most practices that task is note-writing. That closes most of the gap in an afternoon, and it is also the honest starting point for the privacy impact assessment you will need anyway.
Found this useful?
Share this with your network on LinkedIn, it helps more than you think.
Enjoyed this read? Get the next one in your inbox.
When there is something worth your time, you will be first to know. No spam, unsubscribe anytime.
Keep reading
What Should We Refuse to Sign? Five AI Contract Terms Decided Once, at Signature
Data ownership, training rights, exit, price, and liability are settled the day you sign and renegotiated almost never. In a clinic one of them is not just commercial, it is a privacy obligation you already hold. Here is what each clause decides, what the standard wording quietly grants, and the version worth asking for instead.
Read article
What Are We Exposed To, and What Is Our Policy? The One Page Every Practice Owner Should Write
Your staff are already using AI, so the policy question is not whether to allow it. It is whether the rules exist in writing or in each person's private judgement. Exposure comes in three shapes, and in a clinic the one most policies address is not the one most likely to reach a patient.
Read article
Agent Washing: How to Tell a Real AI Agent From an Automation With a New Sticker
Every clinic product renamed itself an agent this year, and the word stopped carrying information. A five-scenario quiz trains your eye, and five questions expose what a vendor actually built, because the label decides the price you pay, the way it fails, and the oversight your practice owes it.
Read articleHave software that should be smarter?
Let’s map a free AI-transformation roadmap for your product.