Medical Writing AI Software: What Small Agencies Can Use Safely

A small medical writing agency can safely use business-tier AI plans (ChatGPT Business, Claude Team, Microsoft 365 Copilot or Gemini in Google Workspace) for editing and restructuring public or client-approved text. Patient-level data never goes into any tool, and unpublished results, embargoed material or NDA-covered files go in only with the client's written permission, never through a personal account.

The question an agency owner is really asking is not "which tool is best" but "which tool, on which plan, for which kind of document". The same model can be fine for tidying a public review and wrong for a client's unpublished efficacy table. So the order of work is: classify the document, match the class to a plan, keep a never-paste list where everyone sees it, and put a person between every AI-touched reference and the client. Nothing here is a clinical claim or legal advice; contract terms are for your lawyer.

Follow me on Instagram@sagnikteaches

Start with the document, not the tool

Every job an agency handles falls into one of four data classes. Classify the document before you open any tool, and the tool choice mostly makes itself.

Connect on LinkedInSagnik Bhattacharya

Class 3 is where agencies get caught, because a results table looks like ordinary text. A pharma or device client may have unpublished data they regard as their most valuable asset, and an NDA that forbids disclosure to third parties. An AI vendor is a third party. Unless the contract says otherwise, treat any class 3 material as off limits and edit it in your normal tools.

Subscribe on YouTube@sagnikteaches

Tool tiers by plan: what you pay and what you get

The plan matters more than the brand. The core difference is whether your content is used to train the vendor's models and who can see it. The figures here are list prices checked this month; confirm them on the vendor pages before you budget.

For most four-to-ten-person agencies, the business tier is the right place to land. A one-person agency can use an individual paid plan with the training switch off for class 1 work, but should not use it for class 2 because there is no administrator, no central record of who used what, and no guarantee that a freelancer's personal login stays within your policy. The wider question of what Workspace plans promise is covered in whether Gemini is safe for confidential business data.

One practical point on retention. On consumer Claude plans (Free, Pro and Max), new chats are kept for 30 days with the training setting off and up to five years with it on; ChatGPT's temporary chats are kept for up to 30 days; Gemini temporary chats up to 72 hours. Those are consumer settings, they change, and they are not what a client NDA is written against. Business plans give you admin controls and a contract; consumer plans give you a toggle.

The never-paste list, in the words I would put on the office wall

• Patient names, initials with dates of birth, record numbers, or any case narrative that could identify a person.
• Unpublished results tables, interim analyses, statistical outputs or datasets, unless the contract names the tool for that use.
• A client's embargoed material before its release date.
• Reviewer comments that name or identify the reviewer, and any confidential peer-review material you hold for a client.
• Anything marked confidential by a client where you have not recorded permission to use AI.
• Passwords, login details for journal submission systems, and client contact lists.
• A client's proprietary product name or code name linked to unpublished indication or mechanism details.

A useful test for a doubtful item: if this text appeared in a stranger's chat history tomorrow, would the client be unhappy, or would I? If either answer is yes, it stays out. For a deeper checklist built around patient information, see patient data and AI: a confidentiality checklist.

NDA and contract terms to check before the first project

Many NDAs predate widespread AI use and say nothing about it. Silence is not permission. Four clauses are worth searching for in every client agreement, and a lawyer should advise on the exact wording.

1. Third-party disclosure. Does the agreement bar sharing confidential information with third parties without consent? Treat an AI vendor as one.
2. Subprocessors. Does the client require approval for any tool that processes their material? Keep a list of tools and plans you use, and offer it.
3. AI-specific wording. Some clients now include a clause prohibiting or conditioning AI use. If so, follow it exactly, even if it is stricter than your policy.
4. Return and deletion. What must you delete when the project ends? Chat histories in an AI tool may count; check how to delete them and by when.

A sentence you can add to your statement of work, for a lawyer to check: "The agency may use [named tools on named plans] for drafting support, editing and summarisation of non-clinical text. The agency will not enter patient-level data or unpublished study results into these tools without the client's written consent. All output is reviewed by a qualified writer before delivery." Short, specific and easy for a client's legal team to approve or amend.

A four-person agency sorts a month of 31 documents

Illustrative figures. A four-person agency writes for six clients: two device companies, one pharmaceutical client, a patient charity, a medical education provider and a hospital department. The owner classifies a typical month's work. She counts 31 documents: 9 class 1 (reviews of published literature, public guidelines), 14 class 2 (draft manuscripts, slide decks, internal briefs), 5 class 3 (unpublished results sections for the pharmaceutical client) and 3 class 4 (case reports with patient details for the hospital department).

Her tool plan follows. Class 1 and class 2 work, 23 of 31 documents, goes through a business plan: four seats of ChatGPT Business at $25 monthly comes to $100 a month, or $80 on annual billing (list prices; she checks them before buying). The five class 3 documents stay out of AI entirely, because the pharmaceutical client's NDA is silent on AI, and she emails the client's contact to ask. The three class 4 documents never enter any tool; the department's own information-governance policy applies and she follows it. The result: about three-quarters of the workload can use AI, for a software cost of $80 to $100 a month, and the rest is protected by a documented rule rather than by good intentions.

The email to the pharmaceutical client is short: "Our contract is silent on AI tools. We would like to use [tool] on a business plan that does not train on our content for editing support on the results sections. We will not do so unless you confirm in writing. If you prefer we do not, say so and we will edit manually." Three outcomes are possible, and each is fine, but the agency now has a record.

What AI does well in medical writing, with a safe prompt for each

When the data class allows it, AI is useful for the tasks that eat writers' time. Each prompt below is built to need no confidential content. The first uses public text; the others use placeholders or structure only.

Task: restructure a public review article into a 600-word plain-language summary.
Audience: patients with no medical background. Reading age 12.
Use ONLY the text I paste below. Do not add facts, statistics or claims about treatment.
Mark any sentence you were unsure about with [CHECK].

[public text pasted here]

A realistic output is smooth and readable, with the odd [CHECK] mark. What you fix: any sentence that turns a cautious finding into a promise ("may reduce" becoming "reduces"), and any invented number. The plain-language version is for the writer to check against the source, not to send on trust.

Here is the outline of a clinical study report section (headings only, no data).
Suggest a consistent order of subheadings and a house-style checklist for tables and figure legends.
Do not generate any results or numbers.

This one uses structure alone, so it carries no data risk. Other tasks of the same kind: build a style sheet from a client's published guidance, produce a checklist for abbreviation consistency, draft a query list from your own comments, or write a hand-over note after a project. For longer proposals and funding documents, using AI to write grant applications shows a comparable workflow with checks.

Reference checking: the human step that cannot be skipped

AI tools can produce a citation that looks perfect and is fictional, or that matches a real title with the wrong journal, year or authors. For a medical writer that is the single worst failure mode, because a reviewer or reader may act on the reference. The rule is simple: no reference leaves the agency unless a person opened the source and confirmed it supports the sentence it is attached to.

A reference table makes this manageable. Illustrative columns: number, claim in the text, citation as written, source opened (yes/no), identifier matched, checked by and date. Budget real time for it: at a minute or two per reference, a 40-reference project is an hour or more of checking. A short, illustrative example of what it catches: the AI cited a trial publication for a claim about a secondary outcome, but when the checker opened the paper, the outcome appeared only in a supplementary table with a different sample size. The text was corrected and the reference stayed.

Use AI for what it is reasonable at here: reformatting a verified reference list into a journal's style, or flagging inconsistencies in how author names appear. Do not use it to find or verify references. If you use a search tool that cites sources, still open each one.

No clinical claims: the editing rules that protect you and the client

Medical writers do not make clinical claims; they present the evidence a qualified author or client has approved. AI makes this boundary easier to cross by accident, because it writes fluently about treatment. Three editing rules help. Never accept AI-generated statements about efficacy, safety, dosing or indications unless the writer can point to the approved source for each one. Keep hedged language as the source has it ("was associated with", "in this population"). And search the final text for absolutes: "proven", "safe", "cures", "no side effects". If one appears, trace it to a source or delete it.

Journals and publishers have their own rules on AI assistance, including whether and how its use must be disclosed, and on what counts as authorship. The ICMJE recommendations, for example, say AI and AI-assisted tools should not be listed as authors, ask journals to require disclosure of AI use at submission, and ask authors to describe that use in both the cover letter and the submitted work. Those policies differ and change, so read the target journal's instructions for authors at kick-off and record the requirement in the project file. When a policy is unclear, ask the corresponding author rather than guessing.

Trying a new tool without exposing a client

New AI products appear every month and writers will want to try them. Give them a safe way in. Keep a small "sandbox" set of public documents, such as two open-access papers and a public guideline, and require that any tool is tested on those first. Record what the tool did well, what it got wrong and what the vendor's privacy page says about training, retention and where data is processed. Only then does the owner decide whether the tool joins the approved list and on which plan. Illustratively, a writer finds a literature-summary tool that looks impressive; on the sandbox papers it attaches two correct findings to the wrong trials, and its privacy page offers no business tier. It stays off the list, and nobody has lost a client's confidence finding that out.

Review the approved list at each renewal date. Vendors rename plans, retire products and change retention terms; a tool that was fine in spring may have a new data setting by autumn. Put the review in the calendar, name an owner, and note the date in the policy.

Freelancers need the same rules. If you subcontract writing, put the policy in the subcontractor agreement, supply the approved seat, and ask them to confirm in writing that they have not used other tools on client material. Your client holds you responsible for what a subcontractor does with their material, so the same checks apply down the chain.

A simple AI-use log that answers a client's question in a minute

Clients and journals increasingly ask whether and how AI was used. A log answers that without a scramble. Keep one row per project: project code, tools and plan used, tasks done with AI (for example "plain-language summary draft"), data class of the material, who checked the output, and the date of delivery. Illustratively: "P-2231, business plan, restructure of public review into summary, class 1, checked by [writer initials], delivered [date]". It takes thirty seconds per project and turns a hard conversation into a short email.

The log also shows you patterns. If one writer uses AI on class 3 material despite the rule, the log shows it before a client does. If a task never uses AI, you can drop the plan seat that supports it.

Three things that go wrong in real agencies

• The personal login. A freelancer pastes a client brief into their own free account at night because it is quicker. The fix is an agency-provided seat and a clear rule that client material never enters a personal account. Offer the seat; do not only prohibit.
• The browser extension. A grammar or "AI rewrite" extension reads everything typed into the browser, including a client's manuscript in a web tool. Add extensions to the policy: approved list only, and none that can read client documents.
• The shared screenshot. A writer screenshots a table to ask the AI to reformat it. The image holds the unpublished numbers. Treat images and uploaded files exactly like pasted text.

When a mistake does happen, record it, tell the client in line with the contract, delete the conversation where you can, and change the rule that allowed it. A short incident note beats a quiet fix.

A one-page policy outline you can adapt

1. Approved tools and plans: named, with the seat owner and the date last reviewed.
2. Data classes: the four above, with an example of each from your own client list.
3. Never-paste list: the one above, on a single printed page.
4. Client permission: where it is recorded and who asks.
5. Human checks: every reference opened, every claim traced to an approved source, final read by a qualified writer.
6. Disclosure: the wording you use for clients and the journal rules you check.
7. Incidents: who to tell, within what time, and what you do first.
8. Review: every six months, and whenever a vendor changes its terms.

If you would rather build the policy and the tool set-up with someone, that is what my AI implementation consultation covers. The broader picture of using AI in professional services is in the AI guide.

Questions agency owners ask about AI in medical writing

Is it safe to paste a client's manuscript into ChatGPT?

Only if three things hold: the client's contract or written permission allows AI tools, the tool is a paid business plan that does not train on your content, and the manuscript contains no patient-level data. Free and personal accounts are the wrong tier for unpublished client material. If any one condition is missing, do not paste it.

Do I have to tell clients I use AI?

Tell them, in your contract or statement of work, which tools you use, for which tasks, and what you never put into them. Many clients and journals have their own AI policies, and some require disclosure. Check each client's rules at kick-off and record the answer in the project file.

Can AI be trusted to find and format references?

No. Models can produce references that look correct and do not exist, or attach the right title to the wrong authors and year. Use AI to draft search terms and tidy formatting, but every reference must be opened and matched to the real record by a person. Keep a checked column in your reference table.

What should an agency AI policy contain?

One page: approved tools and plans, a never-paste list, who may use which tool for which task, how client consent is recorded, the human check required before delivery, and what to do after a mistake. Review it every six months because plan terms and privacy settings change.

Further reads

• Can Solicitors Use ChatGPT Without Breaching Confidentiality? — How another confidentiality-bound profession handles the same risk.
• Is It Safe to Put Client Documents Into AI Translation Tools? — Client documents and AI tools, from the translation side.
• AI for Testing Psychologists: Report Drafts Without Pasting Raw Scores — A never-paste rule set for clinical report writing.
• Patient Reviews and AI Replies: Staying Within Confidentiality — Confidentiality limits when replying to patient reviews.
• AI Note-Taking Tools for Therapists: A Confidentiality Checklist — Confidentiality checks for AI note tools in clinical settings.
• What to Ask Before Buying Any AI Tool for a Medical Practice — Vendor questions to ask before any AI purchase.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business pricing page and Temporary Chat FAQ (archived copies of 27 October 2026), Claude pricing page, Anthropic's consumer-terms update and commercial-training article, Gemini Apps Privacy Hub, Google AI subscriptions and Google Workspace pricing pages (archived copies, late October 2026), Microsoft 365 Copilot Business page and Copilot privacy documentation, and the ICMJE recommendations on authorship, checked 4 November 2026. Journal and publisher AI policies vary and must be read for each client. Nothing here is legal advice.