Setup & safety

Which AI tool for which job — a clinician's map

Published July 2, 2026 · Updated July 2, 2026

TL;DR: The major AI tools are not interchangeable, and the differences that matter to a practice aren’t the ones the marketing leads with. There are really two categories — drafting engines (Claude, ChatGPT) that generate text from what you give them, and an answer engine (Perplexity) that searches and cites — plus two assistants you may already be paying for through your office software (Gemini, Copilot) and a design layer (Canva). Match the tool to the job, set each one’s privacy settings once, and keep the same PHI line across all of them. The one-page map at the end is the whole guide in a table.

Who this is for: any clinician who has one AI tool open and a nagging sense they’re using it for the wrong things — or six browser tabs and no idea which to trust with what. Nothing to buy; every tool here has a free tier that’s enough to evaluate it. 15 minutes.

This guide assumes the one rule from What AI actually is for a clinician: no patient-identifiable information goes into a general-purpose AI tool. Every recommendation below is patient-data-free by design.

Two categories, not six tools

The fastest way to stop misusing these tools is to see the category split:

Drafting engines (Claude, ChatGPT — and Gemini and Copilot, which are drafting engines wearing office-software badges) generate language by pattern. Give them source material, structure, and a clear brief, and they produce excellent drafts. Ask them for facts and they produce plausible-sounding text that may or may not be true. You supply the facts; they supply the labor.

Answer engines (Perplexity) run a live search first, then write a short answer from the results, with numbered citations you can click. That makes them the right shape for “what does the evidence say about X?” — and still not a substitute for reading the source. An answer engine narrows your search; it doesn’t finish it.

Hold that split and most “which tool?” questions answer themselves: drafting job → drafting engine; lookup job → answer engine; patient-specific job → neither, ever.

The five, one at a time

Claude — the long-form drafting engine

The job: substantial writing that has to sound like you and work from your material — patient-education handouts, service-page copy, SOPs, guide-length content, and the restructure-this-document work every practice accumulates.

Claude’s practice-relevant feature is Projects: a standing workspace where you upload your reference documents — your service descriptions, your writing samples, your intake templates (nothing patient-identifiable) — and every conversation in that project works from them. That’s the mechanism behind teaching AI your voice, which gets its own guide in this library.

Starter prompt — the patient-ed draft:

Write a one-page patient-education handout about [condition — e.g., plantar fasciitis] for a naturopathic/integrative clinic. Grade-8 reading level. Structure: what it is, why it happens, what generally helps, when to seek care. Plain English first, technical term in parentheses. No specific doses or product names. I am a clinician and will verify every factual statement before this is used.

That last sentence isn’t decoration — it states the workflow. The draft is the tool’s job; the facts are yours.

Where its PHI line sits: generic topics only, like every tool here. Check your training/model-improvement choice once (Settings → Privacy — the Safe Setup Checklist has the walkthrough), and note that a Project full of your practice documents is still a consumer tool: templates and public-facing material belong there, patient anything does not.

ChatGPT — the general drafting engine most people already have

The job: the same drafting work, plus the widest set of extras — and, notably for a practice, image generation is built in: workable illustrative images from a plain-language prompt, including legible text in the image.

If you or your staff already use ChatGPT, there’s no urgent reason to switch for general drafting; the discipline (source material in, verification after) matters more than the logo. Where it earns a specific slot on the map is quick illustrative images and one-off drafting on the account you already have.

Where its PHI line sits: same rule — and one setting to change first: consumer conversations can be used for training unless you turn “Improve the model for everyone” off (Settings → Data Controls). Never generate an image “of a patient,” even a fictional one presented as real — illustrative means illustrative.

Perplexity — the lookup tool (the only answer engine on this map)

The job: “What does the recent evidence say about…?”, “Which guideline covers…?”, “What are people asking about…?” — anything where the output you actually want is sources. Perplexity searches the live web, writes a short synthesis, and numbers its citations.

Starter prompt — the evidence scan:

What does recent peer-reviewed evidence say about [topic — e.g., vitamin D and bone density in postmenopausal women]? Prioritize systematic reviews, meta-analyses, and clinical practice guidelines. List every source. Note where the evidence is mixed or thin.

Then — this is the whole game — you open the sources. An answer engine’s synthesis is a reading list with a summary on top, not a finding. Nothing from it reaches a patient or your marketing without you having read the underlying paper. (Accuracy studies of AI search consistently find citation and synthesis errors; the tool is a time-saver on the search step, not the appraisal step.)

Where its PHI line sits: two cautions. Never search a real patient’s actual situation — a specific-enough query is a case description leaving your custody. And Perplexity’s data-retention-for-training setting is on by default (Account → Preferences → “AI data retention”); flip it before your first real session, because the opt-out only applies going forward.

Gemini — the one inside Google Workspace

The job: if your practice runs on Google Workspace, Gemini appears as a side panel inside Gmail, Docs, Sheets, Slides, and Drive (on Workspace plans that include it) — summarize a thread, draft a reply, restructure the document you’re already in, without switching tools. That convenience is the entire case; as a standalone chatbot it doesn’t do a practice job the two above don’t.

Where its PHI line sits: sharper than most, because your inbox is where patient emails live. The side panel summarizing administrative mail is a time-saver; pointing it at patient correspondence is a PHI-line violation with a nicer interface. On the settings side, the consumer Keep Activity setting governs whether chats can be reviewed by humans and used for training — Google’s own privacy page says plainly not to enter confidential information. Business Workspace accounts are governed by different terms; confirm what your plan actually covers before assuming.

Copilot — the one inside Microsoft 365

The job: the mirror image of Gemini for Microsoft practices — drafting and summarizing inside Word and Outlook (licensing varies: bundled allowances on consumer Microsoft 365 plans, an add-on license for business tenants). If your clinic lives in Word and Outlook, it’s the lowest-friction drafting engine you’ll get staff to actually use.

Where its PHI line sits: the account type decides the rules. Consumer Copilot has its own training opt-out in its privacy settings. Signed in with a work account, prompts fall under Microsoft’s enterprise data protection and aren’t used for training — but “confirm with whoever manages your tenant” beats assuming, and the PHI line holds regardless: Copilot summarizing a patient’s email thread is the same violation as pasting it anywhere else.

The design job — Canva

Guide one promised an answer for design, and for a non-designer clinician the answer is boring and correct: Canva, whose AI features (branded Magic Studio) now cover most of what a practice needs — Magic Design turns a written prompt into editable layout options, Magic Write drafts short copy inside the design, and Dream Lab generates images from text. Social cards, a handout’s layout after a drafting engine wrote its words, a simple one-pager: this is the tool.

Two practice-specific cautions. First, AI-generated images of “medical” subjects need judgment — anatomy renders confidently and wrong, exactly like text does, so keep generated imagery illustrative rather than instructional. Second, design polish doesn’t launder a claim: whatever the card says still has to meet your college’s advertising standards, sourced and un-guaranteed, same as every word you publish.

What this map deliberately leaves off: scribes

AI scribes — tools that listen to a consult and draft the note — are the loudest AI product in healthcare and they are not on this map, because they’re a different category with different rules. Everything above is patient-data-free by design; a scribe touches patient data as its whole function, which makes it a procurement decision: agreements (Canada: PIPEDA/provincial posture and a data-residency answer; US: a signed BAA), documented patient consent before recording, and your college’s current guidance. That’s a separate guide — until you’ve run that checklist, the answer to “can I just use my phone and ChatGPT as a scribe?” is no.

Do you need to pay for any of this?

Not to start. Every tool on this map has a free tier that’s enough to evaluate it against your actual work for a month. The sane sequence: run the Safe Setup Checklist on the accounts you create, work the free tiers, then pay for the one or two you reach for daily — typically one drafting engine and, if research is a big part of your week, the answer engine. Paying for all five is a subscription collection, not a workflow.

Install this: the tool-to-job map

Copy the table into your practice docs (or print it) and pin it next to the PHI Line policy from guide one. When someone asks “which AI should I use for—?”, the answer is a row.

# [PRACTICE NAME] — Which AI tool for which job
Companion to our AI Use Policy (the PHI line). Review: quarterly, with settings.

| The job | Open this | Why | The PHI line here |
|---|---|---|---|
| Draft a patient-education handout | Claude or ChatGPT | Drafting engines work from your brief | Generic topics only; clinician verifies every fact before use |
| Rewrite something in our voice / reading level | Claude | Strongest with your samples loaded (Projects) | Your own content — fine |
| "Is there evidence for…?" lookup | Perplexity | Cited answers you can click and check | Never a real patient's situation; read the sources yourself |
| Summarize a guideline or paper we have | Claude or ChatGPT | You supply the document, it supplies the summary | Published documents — fine |
| Website / service-page copy | Claude | Long-form, voice-consistent | Claims still need real sources + college ad standards |
| Social captions from existing content | ChatGPT or Claude | Repurposing is pattern work | Your own content; no new factual claims added |
| A social card or simple layout | Canva (Magic Design) | Design layer, editable output | No patient photos without a proper consent pathway |
| An illustrative image | ChatGPT or Canva (Dream Lab) | Built-in image generation | Illustrative only — nothing implying a real patient or outcome |
| An email to a patient | NONE of these | Patient-specific work happens in [PMS NAME], by a person | Hard stop |
| Summarize admin email / draft internal docs | Gemini (Workspace) or Copilot (M365) | Lives inside the account you already have | Work-account terms confirmed first; never patient correspondence |
| Notes from a recorded consult | None of these | That's scribe territory — different rules | Separate procurement checklist (agreements, consent, college guidance) |
| Business planning, pricing, quarter mapping | Claude or ChatGPT | Structured-thinking work | Aggregate / de-identified numbers only |

The repurposing prompt that pairs with row six:

Here is a blog post I wrote: [paste]. Turn it into three short social captions and one email paragraph. Keep my wording wherever possible. Do not add any factual or clinical claims that aren’t in the original.

That constraint — no new claims — is what makes repurposing the safest AI job in the building: everything factual was already verified once, upstream.

What’s next

The map tells you which tool; it doesn’t yet make any of them good. The next guide covers the highest-leverage move in this library: turning one verified idea into a week of patient-education content — the workflow, and the prompt that runs it.

Sources

Get the Safe Setup Checklist

The ten things to change before your first real AI work session — including the exact privacy settings per tool.