Yes, but only after you have vetted the tool. Before any client information goes in, check your lawful basis and special category condition, have a processor contract, know where the data is stored and for how long, update your privacy notice and confirm your ICO registration. If a tool cannot pass those, use it only on anonymised or invented text.
Counselling data sits in the strictest group the law recognises. What a client tells you about their health, relationships, sexuality or beliefs is special category data, and the people sharing it expect confidentiality that goes beyond ordinary business privacy. General GDPR guides cover the basics at length. The harder question comes after the basics: you already want to try a tool, so what do you check, in what order, and what do you write down? It is practical guidance, not legal advice, so take specific questions to your professional body or a data-protection adviser.
Start with the job, not the tool
The first test is whether the task needs client information at all. A great deal of counselling admin does not: a template for a first-session information sheet, a waiting-list message, a reminder text, a marketing page, a contract clause. If the task can be done with invented details, the data-protection analysis is short, because no personal data enters the tool. Safe AI uses that never touch client notes lists those jobs. Only when a task needs real client content, such as turning a session recording into notes, does the vetting below matter in full.
Write the job in one sentence before you look at any product: "Draft session notes from my voice dictation after each session" or "Summarise my own bullet points into a letter". That sentence defines what data goes in, which is the foundation of everything that follows.
Step one: your role, the tool's role
You are the controller: you decide why and how client data is processed. The AI vendor is normally your processor, acting only on your instructions. That distinction matters because a processor needs a written contract and may not use your data for its own purposes. Some tools blur it. A consumer chat app that trains on your inputs is using your data for its own purposes, so it behaves more like a controller, and sending client information to it needs a different, much harder justification. If you cannot say with confidence that the vendor is a processor, stop and ask them.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowStep two: lawful basis and special category condition
You need two things in place, and they are separate. A lawful basis for processing personal data (the usual ones for a practitioner are performance of a contract with the client, or legitimate interests), and, because the data is about health, an additional condition for special category data. Commonly cited options include provision of health or social care, which the ICO says applies only where the data is processed by or under the responsibility of a professional who owes a duty of confidentiality, or the client's explicit consent, which must be a clear, freely given and withdrawable choice. Which one fits your practice depends on how you work and whom you serve, so confirm it with your professional body or adviser. What matters here is that the basis for using an AI tool on notes should be the same basis you already rely on for keeping the notes, and that you have recorded it. If you would be relying on explicit consent only for the AI step, you need a genuine choice: a client who declines must receive the same service, and you must be able to work without the tool.
Step three: the processor contract, line by line
Ask the vendor for their data processing agreement and read it for these points. Keep the filled-in checklist in your records.
| What to confirm | Good answer | Walk away if |
|---|---|---|
| Instructions | Vendor processes only on your instructions | It reserves the right to use inputs to improve its products |
| Training | Client data is not used to train models, by default and in writing | Training is on unless you find a setting |
| Sub-processors | A named list, with notice of changes | No list, or "may change without notice" |
| Security | Encryption in transit and at rest, access controls described | Vague claims, no detail |
| Client rights | Help with access, deletion and objection requests | No process |
| Breach notice | A stated timeframe for telling you | No commitment |
| End of contract | Deletion or return of data, with a date | Data kept indefinitely |
As background, plans differ sharply by tier. Business and enterprise tiers of the main assistants do not train on your content by default, while individual plans put the control in each user's settings. A scribe built for clinical settings will usually be better prepared for these questions than a general chat app, but "usually" is not a contract, so ask for the document.
Step four: where the data lives, and moves
Ask where the data is stored and processed, and whether it leaves the country of storage for support, backups or the underlying AI model. Transfers of data out of the area your data-protection law covers need a recognised safeguard. Some vendors let you choose a storage region, but often only on enterprise tiers: OpenAI's own plan comparison, for example, lists data residency for ChatGPT Enterprise and not for Business. Write down the answer, including which region you chose and the date. If the vendor sends audio or text to a separate model provider, that provider is a sub-processor and sits on the same list. A tool that cannot say where your client data goes is not ready for it.
Step five: retention and deletion
Decide how long the tool keeps what you put in, and make that shorter than or equal to your own retention period for notes. Many scribes keep the transcript only until you save the note, or let you choose. Some say they do not keep audio at all; check that claim in the contract, not on the sales page. Test deletion: put in a dummy note, delete it, and ask the vendor how long until it is gone from backups. Add the answer to your retention schedule. A client who asks for their data to be erased needs you to know where all copies sit, so a tool with unclear retention creates a problem you cannot answer.
Step six: a short impact assessment
A data protection impact assessment is a structured risk check. The ICO expects one where processing is likely to be high risk, and its list of high-risk processing calls for a DPIA where innovative technology, which it says includes AI, is combined with another factor such as sensitive data. Session content in an AI tool fits that description, so assume you need one. Even if you conclude you do not strictly need one, writing a one-page version is the best evidence that you thought about it. Use four headings: what data and why; what could go wrong for the client; what you do to reduce it; and your decision. Keep it with the contract. The ICO's guidance on AI and data protection is under review after recent law changes, so check the live pages for the current position rather than relying on an older summary.
Step seven: registration and the privacy notice
The ICO says organisations, sole traders included, that use personal information must pay a data protection fee unless they are exempt, and a practising counsellor keeping client records generally falls in. Check the ICO's data protection fee pages to confirm whether an exemption applies and what you owe, as the amounts change. Your privacy notice must then tell clients about the AI tool: what it is for, who the processor is, where data is held, how long, and how to object. An illustrative paragraph to adapt:
Notes and AI. After sessions I use [tool name], a note-drafting tool, to turn my
dictated summary into a first draft of my session notes. The provider acts as my
processor under a written contract and may not use your information to train
its models. Data is held in [region] and deleted from the tool within [number]
days. I review and correct every note. If you would prefer I did not use this
tool for your notes, tell me and I will write them by hand.Check it against the real contract before you publish: every blank must be filled from the vendor's documents, not from memory.
A worked example: a sole practitioner choosing between two tools
Say an illustrative practitioner sees 22 clients a week and spends about 6 hours a week writing notes. She shortlists a general chat assistant on a business plan and a scribe designed for therapy and healthcare. She scores each on the seven steps, using 0 (fails), 1 (partial) or 2 (clear), with documents in hand.
| Step | Chat assistant, business plan | Therapy scribe |
|---|---|---|
| Role (processor) | 2 | 2 |
| Lawful basis fit | 2 | 2 |
| Processor contract | 2 | 2 |
| No training on content | 2 | 2 |
| Data location stated | 1 (no region choice below enterprise) | 2 |
| Retention and deletion | 1 (removed members' chats kept) | 2 |
| Client opt-out workable | 2 | 2 |
| Total | 12 of 14 | 14 of 14 |
Illustrative scores, but the pattern is realistic. The chat assistant loses points on location and retention, because on that kind of business plan a removed member's chats are kept indefinitely yet stay hidden from the owner, and data residency is reserved for the enterprise tier. The scribe is built for the use and documents it. She chooses the scribe for session notes and keeps the chat assistant for anonymised admin. The scorecard goes in a folder with the contract, the impact assessment and the dated privacy notice. That folder is her evidence if a client or the ICO ever asks.
Emailing a vendor the right questions
A short, specific email gets better answers than a form. Adapt this and keep the reply:
Hello,
I am a counsellor considering [product] for drafting session notes. Before I
decide, please confirm in writing:
1. Do you act as a processor under a data processing agreement? Please send it.
2. Is client content ever used to train or improve models, and how is that
switched off by default?
3. In which region is data stored and processed, and which sub-processors
handle it?
4. How long is audio, transcript and draft text kept, and how do I delete it?
5. How quickly will you notify me of a breach?
6. Can a client's data be deleted on request, including from backups?
Many thanks,
[your name]Vague or evasive replies are an answer in themselves. A vendor serving clinicians will have these ready.
Before and after: making an anonymised example properly
When you do need help on a real situation, such as drafting a letter or preparing a supervision question, anonymise first and check what is left. Original text: "[first name], 34, a primary-school teacher in a small village, told me she had a panic attack at the Harvest Festival after her mother's diagnosis." After a careless edit: "Client, 34, a teacher in a small village, had a panic attack at a festival after a parent's diagnosis." That is still identifiable to anyone who knows her community: age, job, place, event and trigger together form a fingerprint. A proper edit removes or generalises each: "An adult client had a panic attack in a public setting after a relative's diagnosis." Now the question you ask the tool, such as how to structure a supervision question or a grounding script, loses nothing it needs. The test is whether someone who knows the client could recognise them. If yes, change more. If you cannot make it unrecognisable and still useful, do not use the tool for it.
Three common set-ups, and where each lands
Admin only, no client data. You use a chat assistant for contract wording, reminder templates, a website page and a newsletter. No personal data of clients enters, so GDPR's weight falls on your own site and mailing list. Do the contract-line vetting only if you use client names in the mailing tool. This set-up needs the least paperwork and is where most practitioners should start.
Scribe for session notes. A tool records or transcribes sessions and drafts notes. This is the highest-risk set-up: you capture special category data in full, with audio possibly involved. All seven steps apply in depth, you need an explicit client conversation, and a workable opt-out. Some practitioners decide the therapeutic cost of recording outweighs the saving and use dictation after the session instead, which captures your summary rather than the client's words, and is a much smaller data footprint.
Post-session dictation into a drafting tool. You speak your own summary, without names, and the tool formats it. You still need the processor contract if the summary contains any identifying detail, but the footprint is smaller than for a full transcript, and many counsellors find the time saving close to the scribe's. Turning your own bullets into notes shows how that works in practice.
When a client says no, or asks what you hold
Plan for two requests. If a client objects to AI use, record that on their file, work without the tool for them and make sure the tool is never run on their material. A simple flag on the client record is enough. If a client asks for their data, a subject access request, you need to find everything, including copies in the AI tool. That is another reason to prefer tools with clear deletion and export. Practise once: choose a test client record and list every place it would appear, then check you can retrieve or delete each. If you discover a place you had forgotten, fix the process then, not when a real request arrives with a one-month deadline.
If something has already gone in
Suppose you realise last month that you pasted a client's session summary into a personal chat account. Act in a fixed order. Stop using that account for work. Delete the conversation and check the history and any memory feature. Switch off training in the settings. Write down what was entered, when, which client and what the tool's terms say about retention. Assess whether it is a personal data breach: ask whether the exposure is likely to risk the client's rights and freedoms. If it is reportable, the ICO expects to hear within 72 hours of you becoming aware where feasible, so ask your adviser straight away. Tell the client if the risk to them is high. Then close the gap: a rule, a training moment and a work account. Dealing with it quickly and in writing matters more than the embarrassment, and it is far better than hoping nobody asks.
Mistakes that catch counsellors out
- Pasting "just the headline". A short note such as "client disclosed an affair and is struggling at work" is still special category data, and a first name makes it personal.
- Trusting "anonymised" text. A distinctive detail, such as a rare job in a small town, can identify a client even without their name.
- Personal accounts for work. A free account on a phone has no contract and its own training settings.
- Switching on a built-in assistant by default. A software update can add an AI feature that was not covered by your earlier checks. Re-vet after each update that adds one.
- Forgetting the privacy notice. The tool is live for months before the notice mentions it.
The professional-body angle adds its own duties, set out in a checklist for counsellors before using AI, and for the trial itself a safe scribe trial plan shows how to build consent and an opt-out in from day one. For the broader GDPR duties of any small business using AI, the small-business guide sets out the basics.
Keeping the vetting alive after the first week
Vetting is not a one-off. Vendors change terms, add features and swap sub-processors, usually with an email you did not read. Set a six-monthly diary entry to reopen the folder: has the contract changed, is the sub-processor list the same, is the retention still what you wrote, has the tool added an AI feature that sends data somewhere new. Subscribe to the vendor's changes or security notices where they offer them. Note, too, the practical side: if the vendor is bought or closes, you need to export or delete your data quickly, so keep a note of how.
How to check your vetting would stand up
- Could you show it? Contract, impact assessment, scorecard, dated notice and registration, all in one folder.
- Could you explain it to a client in a minute? What the tool does, who has the data, how to say no.
- Could you delete it? Pick one dummy record, delete it and confirm it is gone.
- Is it reviewed? A diary entry every six months and after every major update.
If you would rather work through a particular tool with someone, that is what my AI implementation consultation covers. The principle is the same each time: decide the job first, check the tool against the seven steps, write down what you found, and tell clients plainly.
Questions counsellors ask about GDPR and AI
Is it ever fine to put client details into ChatGPT?
On a personal plan, no. On a business plan with a signed processor contract, no training on your content and a clear retention setting, it can be lawful, but session content is special category data and the risk bar is high. Most counsellors use it only for generic, anonymised tasks.
Do I need clients' consent to use an AI scribe?
Think of two separate questions: your legal basis for processing, and the client's agreement to the recording as part of the therapeutic relationship. Even where consent is not your legal basis, you should tell clients clearly and let them decline. Ask your professional body and a data-protection adviser how to word it.
What does a data processing agreement need to say?
It must set out what the tool may do with the data and only on your instructions, confidentiality, security measures, sub-processors, help with client rights requests, deletion or return at the end, and audit. If the vendor cannot supply one, the tool is not suitable for client information.
Do I need to update my privacy notice?
Yes, if client information goes through any AI tool. Name the type of tool and its purpose, say who the processor is, where the data is stored, how long it is kept and how a client can object. Do it before the tool is switched on, not afterwards.
Do I have to register with the ICO?
The ICO says organisations that use personal information, sole traders included, must pay a data protection fee unless they are exempt, and that usually covers a counsellor keeping client records. Check the ICO's data protection fee pages to confirm your position, because the amounts change.
Further reads
- Starting a Private Counselling Practice in the UK: What AI Can Do — Set up the practice admin that sits around data protection.
- Counselling Contract Template: Key Clauses and an AI First Draft — Contract clauses where AI and data use can be disclosed.
- Jane App for Counsellors: AI Scribe Cost, Consent and UK Data Location — Cost, consent and data location for one popular platform.
- AI for Counselling Practices: Contracts, Waiting Lists and Reminders — Contracts, waiting lists and reminders with AI.
- AI for Counsellors: Can It Draft a Professional Will and Handover Plan? — Plan who holds client records if you cannot.
- Childminder Privacy Notice: GDPR Wording and ICO Check With AI — See how a privacy notice and ICO check are drafted.
- Are ChatGPT, Claude, Gemini and Copilot GDPR-Compliant? — How the big AI assistants stand on GDPR.
- What's the Best AI Tool for Therapists? A Rule by Practice Size — A rule for choosing a therapy AI tool by practice size, with list prices, BAA checks and worked costs for a solo, three-clinician and nine-clinician practice.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ICO guidance on AI and data protection (under review after the Data (Use and Access) Act), on DPIAs, special category conditions, subject access, breach reporting and the data protection fee; OpenAI ChatGPT Business pricing page and help articles on Business data and removed-member retention, checked November 2026. Not legal advice.