AI in medicine: what can be handed over and what never can
The conversation here does not start with possibilities but with a boundary: diagnosis, prescriptions and patients' personal data are never handed over under any circumstances and regardless of how good the model is. Everything else is the work around medicine, and more money leaks away there than people think.
The conversation here does not start with possibilities but with a boundary: diagnosis, prescriptions and patients' personal data are never handed over under any circumstances and regardless of how good the model is. Everything else is the work around medicine, and more money leaks away there than people think.
---
The main point first
Among those who answered my survey there are doctors and dentists. The subject is live for them, but the conversation has to start with the boundary rather than with possibilities.
What is never handed over under any circumstances:
Diagnosis. A model does not diagnose, does not assess a patient's condition and does not confirm somebody else's diagnosis. An answer that looks plausible is more dangerous here than no answer.
Prescriptions. Dosages, regimens, interactions. The mistake is invisible from the look of the result: a number looks like a number and a drug name like a drug name.
Patients' personal data. It does not go out. Not anonymised "just in case", not in fragments.
Anything you sign for professionally. The decision is made by a specialist, and responsibility is not shared with a tool.
That boundary does not soften over time and does not depend on how good the model is. Everything below is about the work around medicine, not about medicine.
Where the tool genuinely helps
The common feature: you supply the material, and it is about a process rather than a patient.
Documents and internal rules
Internal instructions, appointment protocols, preparation checklists for procedures, staff handouts.
Analysing existing documents for contradictions and gaps is the safest and most useful task: the model compares what you gave it against itself.
Explaining the complex in plain language
One of a doctor's most frequent and least liked tasks: explaining to a patient what is happening without jargon.
What matters here: you supply the content, the model works on the form. Not "explain what an implant is" but "rewrite this text of mine so somebody without a medical education understands it, adding nothing".
Patient handouts
How to prepare for a procedure, what to do afterwards, when to seek help. Standard texts that in a clinic either do not exist or are written in the language of an appliance manual.
And the same principle: the content is yours, the form is the model's work. Every statement is checked by you before it goes to a patient.
Answers to standard questions
Not medical ones — administrative. Opening hours, prices, how to book, what to bring, how to get there.
That is what eats a receptionist's time and requires no medical qualification at all.
Transcription and structuring
Dictated notes after an appointment, meeting recordings, going through patient feedback in bulk.
Administrative routine
Reporting, planning, website copy, handling reviews.
Separately: why this concerns the whole clinic rather than the doctor
The biggest loss in a medical business is not clinical.
Enquiries outside working hours. Somebody wrote in the evening, the answer came in the morning — they have already booked somewhere else. That loss appears in no report: an enquiry left unanswered exists nowhere.
Expensive services. Asked "how much does an implant cost", a person gets a number. A number with no explanation of what makes it up and what they get for it frightens people off — particularly on large amounts, where the decision takes a long time.
Patients who did not come back. Work with the existing base usually does not happen: once a year the owner remembers, asks somebody to ring round, and it fizzles out.
Social media with no link to booking. Content goes out and no path from a post to a booking has been built.
All four are administrative tasks, and they get solved without a single medical statement.
A real case
The client is Zakhar Fomin, owner of a dental clinic. He gave permission to use his name and the figures.
The task: a system that answers standard questions round the clock, qualifies the enquiry and books an appointment. A smart receptionist, in effect.
Sold for $1 900 a year, paid in full. As the second step, a CRM at $3 000–3 600.
Why he agreed. The comparison was not against zero but against an existing cost line: an in-house team — a marketer, a social media specialist, a remote administrator — at around $720 a month plus taxes, holidays and the human factor.
And his main objection, stated outright: the fear that a bot would damage the reputation and lose an expensive implant patient.
How that gets closed. Not by promising the bot is clever but by the architecture: it recognises the status of an enquiry, qualifies it tactfully and hands a hot request to a live person. A large fee never stays on automated messaging to the end.
That is the rule which matters more in medicine than anywhere: hot leads go to a human.
The verification protocol
For anybody who does use the tool on texts around the clinic.
Separate the verifiable from the reasoning. Names, dosages, timescales, standards, references — checked against the source every time, even when they look obviously right.
Set the material frame. "Work only from the attached text; if the answer is not in it, say the data is insufficient." Without that line the model will fill the gap with something plausible.
Nothing reaches a patient without a specialist's review.
And the safest mode: ask not "tell me about" but "rewrite this more clearly, adding nothing". When all the material is yours, there is nothing to invent.
How this gets monetised
For a doctor. Time freed from administrative routine. Plus handouts and explanations that improve how well recommendations get followed.
For somebody working with clinics. This is one of the most solvent niches, and getting into it requires no medical knowledge — because what is being bought is not medicine but organisation: handling enquiries, working the existing base, content, analytics.
How to approach a clinic. Not with an offer of automation. With a question: how many enquiries arrive in the evening and at weekends, and what happens to them. The owner almost never knows the exact answer, and that not knowing is your sale.
The price range. From my scrape of 848 listings: a system or pipeline is $490–1 460, against a $30–75 median for one-off work. And the $1 900 in the example above is the price of one specific project, not market statistics.
What is available for this on the platform
More than 140 AI assistants across 14 categories. Relevant to a clinic: client retention, sales, marketing, HR, legal.
The legal category — thirteen assistants. Particularly important for medicine: handling personal data, consent, and the requirements around advertising medical services.
Gen AI, "file → text" — transcription and recognition. $0.15 per generation.
The express course "Perplexity as an investigative tool" — a method for checking facts. For a profession where a mistake costs reputation, it matters more than any other. $9, included in Basic.
AI Workflow — if you need a process: handling enquiries, working the base, regular reminders. Scheduled runs come with the Full plan.
Registration is free and opens three days of full Basic access.
Where to start today
One action, and it is not about medicine.
Take one patient handout you already have and ask for it to be rewritten more clearly — adding nothing and removing nothing in substance.
Read the result as a doctor and check every statement.
That is the safest task available: the content is yours, the checking is yours, and the model's work is only the form. That is where people start.
---
The second approach: a pipeline rather than a chat
Everything above is work in dialogue with an assistant. There is another format, and it suits volume better.
Cascading prompts in spreadsheets
Module 1 of the programme — "Earning with ready templates" — holds Google Sheets with an integrated model: a large task split into micro-steps, each living in its own cell and taking the previous one's result as input. A formula can be dragged down a thousand rows and a thousand tasks get processed at once. How that works step by step is covered separately: [Thirty posts in one run](/en/blog/thirty-posts-one-run).
What is available for medicine:
The audience analysis template — patient segments, their fears and objections. The HR template — job descriptions and internal rules for clinic staff. Neither touches diagnosis.
Module 1 is included in the Basic plan.
The same chains, visually: AI Workflow
When a chain needs images, video, sound or calls to external services, a spreadsheet is not enough — AI Workflow takes over: the same cascade, but on a canvas where each block's result becomes the next one's input. The quote is calculated before the run and an assembled chain is saved permanently. A detailed breakdown: [Build once, run always](/en/blog/build-once-run-always). Scheduled runs come with the Full plan.
---
How this looks on a real project
The administrative part described above was built for a dental clinic — Fomin Studio, the Fomin Method. The work was delivered as eleven sections, and it shows how this is put together.
The audience was split into four segments. The anxious — priority one, because the barrier here is fear rather than price. The aesthetically driven — a high-value way in, with decisions maturing over months. The functional, with jaw joint problems — an open niche where nobody has found the patient a cause. Ambassadors — the loyal core who bring their families.
A principle worth remembering: a segment is a way in, not a cage. People migrate: a functional patient becomes an aesthetic one, a satisfied one becomes an ambassador. The bot identifies the segment from the first few sentences and reclassifies on new signals.
The running offer — a free CT scan. The patient leaves with the scan, a photo record and a treatment plan explained, even if they decide to be treated elsewhere. That is not a discount but risk removal: diagnostics like that are normally paid for separately.
And the medical frame written into every piece of material: we do not promise an outcome, we promise a cause and a plan · links between oral and general health are presented as an observation rather than a promise · anything acute goes straight to an in-person doctor · online is a second opinion rather than a diagnosis, as medical regulation requires · no invented deadlines.
The metric used to judge the content was reactions and shares rather than views. A telling figure: the phrase "it doesn't hurt", delivered as a slogan, produces just 1.12 shares. What works is not the promise but the story.
---
What to read next
[Local AI models for sensitive data](/en/blog/local-models-sensitive-data) — the only permissible option where data requirements apply.
[What AI agents are, in plain words](/en/blog/what-are-ai-agents-basics) — how a smart receptionist is built.
[One deal end to end: from a listing to $1 900](/en/blog/full-deal-breakdown) — the whole path of selling a system to a clinic.
[Checking a fact with AI search](/en/blog/fact-checking-with-ai) — the compulsory protocol for this profession.
[Best open LLMs 2026](/en/blog/open-source-llm-guide) — the full guide to open models.