Use AI in your clinicwithout crossing a line

What you will have at the end.

A written boundary policy your team can apply, a prompt pattern that produces usable drafts, and one assistant handling a task where a mistake is cheap.

There is a version of this topic that promises your clinic will run itself. Ignore it. The useful version is narrower and duller: AI saves you real hours on writing, on first drafts, on summarizing, and on sorting. It does not practice, and it should never be anywhere near a decision about a patient.

Draw the boundary before you use anything

1. Write the three lines your team cannot cross

   Post it where people work. It has to be short enough to remember and specific enough to apply under pressure.

   —No patient information goes into any AI tool. Not a name, not a chart, not a photo, not a message from a patient pasted in for help replying, not a de-identified summary. The de-identified summary is the one people rationalize and it is still the patient's information.

   —No clinical output reaches a patient. Not advice, not triage, not an interpretation, not a reassurance. Anything a patient reads that touches their care is written and approved by a human who is qualified to write it.

   —Everything published gets read by a person first. Every social post, every page, every email, every image.

   In practice

   Most professional bodies have now published guidance. Read yours, because the boundary in your jurisdiction may be tighter than this and it is never looser.

Writing, where it genuinely helps

1. Use it for the draft, never the final

   The value is in the blank page problem. A draft you rewrite is far faster than a page you start. A draft you publish is the reason so much clinic content reads identically to every other clinic's.

   —Good uses: first drafts of service pages, social captions, ad variations, email replies to suppliers, meeting notes, turning a list into prose.

   —Poor uses: anything where being wrong matters and you cannot check it, anything with a statistic in it, anything addressed to a patient about their care.

2. Give it four things and it stops producing filler

   The difference between a useless draft and a usable one is almost entirely in what you provide. Four elements, every time.

   —The audience, specifically. Not patients. A forty-five year old who has avoided the dentist for a decade and is embarrassed about it.

   —The constraint. Length, reading level, what it must not say.

   —A sample of your own writing. Two paragraphs you wrote is worth more than any instruction about tone.

   —The facts. Everything true has to come from you. The model will invent a statistic with perfect confidence if you leave a gap.

   In practice

   Short prompts with good inputs beat long prompts with elaborate instructions. We have tested this repeatedly on ad creative and the short version wins.

3. Check every number and every claim

   Invented statistics are the single biggest risk in AI-written health-adjacent content, because they are plausible, specific and completely fabricated.

   —Any figure in the output either came from you or gets deleted.

   —Any cited study either has a link you have opened or gets deleted.

   —Any claim about outcomes gets checked against what you are actually permitted to say in advertising.

Visuals

1. Give the generator the brand, the text, and one line of direction

   The instinct is to write an enormous prompt describing layout, composition and hierarchy. It produces worse results than a short one. Give the brand, give the exact text, give one sentence of direction, and let the model compose.

   —Brand: your colors as hex values, your two fonts, and your logo attached as a file.

   —The exact text to place on the image, spelled as it should appear.

   —One line of direction. Flat vector with solid fills. Or a polished 3D cartoon illustration.

   —Nothing about positions, sizes or proportions.

2. Know what a clinic must never generate

   This list is not about taste. Each item is a real problem, either with an advertising policy, a professional body, or with patients recognizing a fake.

   —No generated before and after of a body. Ever, in any context.

   —No generated person presented as a patient or a member of your team.

   —No screen, chart or graph showing invented data. Generators fill screens with plausible numbers by default, and a fabricated result is a fabricated claim.

   —No third party logos, which appear spontaneously and are a trademark problem.

   —No generated clinical imagery. Anatomy comes out subtly wrong and a professional audience notices immediately.

3. Run two directions and compare

   Generate the same message in two visual directions rather than iterating one. Which direction performs varies by clinic and by service, and it is not predictable from taste.

   —Flat vector: clean, legible at small sizes, reads as professional.

   —3D cartoon illustration: warmer, more human, travels further on social.

   —Photorealistic: usually the worst of the three for a generated clinic image, because the fake details are exactly what people notice.

An assistant that sorts, not one that answers

1. Pick a task where being wrong is cheap

   The right first assistant in a clinic does classification, not conversation. It reads an incoming enquiry and decides which bucket it belongs in, so the right person sees it first. If it gets one wrong, a human notices in thirty seconds and nothing bad happened.

   —Good: sorting enquiries into new patient, existing patient, supplier, and spam.

   —Good: drafting a suggested reply for a human to edit and send.

   —Good: summarizing a week of enquiries into themes.

   —Bad: answering a patient directly without review.

   —Bad: anything that decides urgency. An assistant that judges who needs to be seen sooner is making a clinical decision, whatever you call it.

2. Build it with a human at the end, permanently

   Not as a training period that you remove later. Permanently. The design assumption is that the assistant prepares work and a person releases it.

   —The assistant writes to a queue. A person approves or edits.

   —Log every decision it made, so you can check the ones it got wrong.

   —Review the log weekly for the first month, then monthly.

   —One clear rule for what it does when it is unsure: escalate, never guess.

3. Tell people when they are reading generated text

   Not on every social caption. But if a patient is corresponding with something that is not a person, they get told, plainly, at the start. This is becoming a legal requirement in several places and it is already an ethical one everywhere.

When it goes sideways