Working ethically with AI: what every counsellor needs to consider

Two years ago, most counsellors could reasonably treat AI as a future problem. Today, it's already in your practice — whether or not you've noticed. It's in the transcription tool your supervisor mentioned. It's in the intake platform some agencies are trialling. It's in what your clients are using between sessions when they can't reach you. It may be in your practice management software.

The question is no longer whether AI belongs in therapy work. It's already there. The question is what you, as an ethical practitioner, need to think about now — before you use it, when you use it, and when your clients do.

BACP's incoming Ethical Framework (Autumn 2026, mandatory three months later) makes this explicit for the first time. Practitioners are expected to think through their use of digital tools, including AI, and be able to account for it. NCPS and UKCP guidance points the same way.

This piece is a practical framework for thinking about AI in your practice. It doesn't take a position on specific products. It won't tell you what to buy. What it will do is give you the categories, questions, and grey zones you need to make your own defensible decisions.

What “AI” actually means (and why the distinction matters)

The word AI covers several different technologies, each with different ethical implications. Being precise about which one you're using matters.

Large language models (LLMs)are the systems behind ChatGPT, Claude, and similar tools. They generate text based on patterns learned from vast training data. In therapy contexts, they're being used to draft session notes, structure written intake responses, summarise sessions, and generate letters. They're probabilistic — they don't “know” anything, they generate plausible-sounding text.

Transcription and voice-to-text turns spoken words into written text. Typically uses machine learning but is more deterministic than LLMs. Used for note dictation, session recording (with consent), and voice-controlled interfaces.

Matching and recommendation algorithmstry to pair clients with therapists based on data — presenting complaint, therapist specialism, availability. Used by platforms like BetterHelp and, increasingly, UK counselling directories.

Predictive risk scoringuses historical data to flag clients at higher risk of self-harm, non-attendance, or crisis. Marketed as “clinical decision support.”

These are ethically distinct. Transcription of your own dictated notes is very different from an algorithm choosing which therapist a client sees. Being loose with the word “AI” makes it impossible to think clearly about any of them.

Where AI legitimately helps therapy practice

There is a real category of use where AI reduces the administrative burden that pulls therapists away from clinical work. Done carefully, this is straightforwardly positive.

Note structuring. LLMs can take your rough notes and structure them into SOAP, DAP, or your preferred format. You provide the clinical content; the AI handles the formatting. The clinical judgement remains entirely yours.

Voice-to-text dictation. Speaking notes rather than typing them is faster for most practitioners and often produces richer records. Modern transcription is accurate enough for clinical use, provided you review the output.

Intake form processing.Structured extraction of information from open-text intake responses — pulling out named risk factors, previous therapy history, presenting concerns — for practitioner review.

Scheduling and administrative correspondence. Drafting appointment confirmations, cancellation acknowledgements, invoice reminders. Nothing clinical, just admin at scale.

Session summaries for the practitioner's own reference. Distilling a session's key themes for your own records or supervision preparation.

The common thread: you remain the clinical decision-maker.AI reduces the friction between your judgement and the record of it. It doesn't make the judgement itself.

Where AI has no legitimate place in therapy

There is an equally real category of use where AI shouldn't be involved at all.

Clinical decision-making.AI cannot decide whether a client is safe, whether a referral is warranted, whether the therapeutic relationship is progressing, or what intervention is appropriate. These decisions require the therapist's clinical judgement, drawn from training, supervision, and direct experience of the client.

Replacing the therapeutic relationship.AI-driven chatbots marketed as “therapy” or “mental health support” are not therapy. They lack the fundamental features of therapeutic work: sustained relationship, clinical judgement, ethical accountability, the ability to hold complexity over time. Some may be useful adjuncts for self-management. None are therapy.

Client-to-therapist matching by algorithm. Choosing which therapist a client sees is itself a clinical decision — one that considers presenting concern, prior therapy, cultural context, therapeutic modality, and often intuition. Algorithmic matching optimises for what can be measured. Much of what matters can't be.

Replacing supervision. AI cannot supervise your clinical work. It can prompt you to reflect. It cannot hold the ethical and professional accountability that supervision requires. The therapeutic professions have supervision structures for good reason.

Content that appears to come from you.AI-drafted messages sent to clients as if you'd written them are a form of misrepresentation. If AI wrote it, the client should know.

The grey zones (where reasonable practitioners disagree)

The interesting questions aren't at the extremes. They're in the middle, where thoughtful practitioners land in different places.

AI-drafted session summaries shared with clients. Some clients ask for written summaries. AI can produce these efficiently. The question is whether the client is told AI was involved, and whether the summary genuinely represents the session or introduces subtle distortions.

AI-suggested interventions or reflections.Some tools offer “here's what you might explore next session” prompts based on note content. Some practitioners find these useful for supervision preparation. Others see them as offloading the clinical thinking that should be theirs.

Between-session client-facing tools.Journaling apps, mood trackers, and reflection prompts increasingly incorporate AI. Whether these support or undermine the therapeutic frame depends heavily on how they're integrated and what the client understands.

AI-generated first-draft assessments.Producing an initial assessment framework based on intake responses, for the practitioner to then review and revise. Efficient — but risks anchoring the practitioner's thinking to the AI's structure.

There isn't a right answer to these. What matters is that you've thought about them, made a considered choice, can explain your reasoning to a supervisor, and are open to changing your mind as evidence emerges.

The BACP framework, briefly

The draft 2025 Ethical Framework (published Autumn 2026, mandatory three months after) addresses digital tools and AI directly. Practitioners are expected to:

• Understand how the digital tools they use handle client data
• Consider the ethical implications of AI use in their practice
• Be able to explain and account for their choices

The framework doesn't tell practitioners which tools to use or avoid. It expects practitioners to think through their choices and be able to defend them. That framing is worth keeping in mind: the goal isn't compliance with a checklist, it's ethical clarity about what you're doing and why.

Questions every counsellor should ask

Before adopting any AI tool in your practice, these are the questions worth answering — for yourself, your supervisor, and if asked, your professional body.

Where does the client's data go? Does the AI provider retain it? Do they use it to train their models? Is it processed in the UK, EU, or elsewhere? Who has access?

What am I asking the AI to do? Structure, transcribe, summarise, decide? Different tasks have different ethical weight.

Would I be comfortable telling my client I use this? If not, that's a signal. If yes, do they actually know?

Am I keeping clinical judgement?Is the AI supporting my thinking or replacing it? Would I make the same clinical decision without the AI's output in front of me?

How does my supervisor view this? Have you talked about it? Would you want to?

Can I explain my decision to a professional body? If a complaint arose or you were audited, could you articulate why you use this tool, how, and with what safeguards?

What happens if the tool fails or changes? AI providers pivot. Products get discontinued. Are you dependent on something that could disappear tomorrow?

Closing

AI in therapy work is not a problem to solve once. It's a set of ongoing judgements that will keep evolving as tools change, evidence emerges, and professional guidance firms up.

The therapists best positioned for this are the ones who engage with it deliberately — neither dismissing AI as automatically unethical, nor adopting it uncritically because it's convenient. The line the BACP framework draws, and the line most thoughtful practitioners land on, is that AI belongs in the administrative and infrastructural work of therapy, but the therapy itself remains the therapist's.

If you use AI in your practice already, you already have decisions to defend. If you don't yet, you soon will. Either way, the work is the same: understand the categories, ask the hard questions, keep the clinical judgement yours, and be able to account for what you've done and why.

Mark Devereux is the founder and CTO of Sessionly, UK-built practice management software for therapists and counselling agencies. This piece reflects his own view; Sessionly does not take a position on which AI tools practitioners should or shouldn't use.