AI notes for therapists and counsellors: the tool that stays in the room

Therapeutic relationship confidentiality is not a compliance requirement that sits above other requirements — it is the foundation on which therapy works at all. A client who discloses suicidal ideation, childhood trauma, or a history of abuse does so because they believe that information will remain within the room. The moment a client has reason to doubt that, the therapeutic relationship changes.

This makes AI documentation in mental health practice a genuinely different conversation from AI documentation in, say, a business meeting or a GP surgery. The stakes of data exposure are not primarily legal — they are clinical. The question is not just whether your notes comply with GDPR; it is whether your use of technology would cause a client to disclose less.

AI notetakers can help mental health practitioners. The documentation burden in private practice and NHS mental health settings is substantial and contributes to burnout. A tool that reduces the time spent writing session notes, risk assessments, and formulations from memory at the end of a full clinic day is genuinely valuable. But the tool must be one that is architecturally compatible with therapeutic confidentiality — not just legally compliant, but clinically trustworthy.

The mental health documentation burden

A full-time psychologist or therapist seeing 8–10 clients per week in private practice, or 15–20 in an NHS IAPT or CMHT setting, generates significant documentation outside of session time:

Session notes. Written after every appointment, these capture the presenting issue that day, the therapeutic content of the session, the client's response, risk indicators reviewed, and the plan for the next session. In CBT-based approaches, this may also include homework agreed, thought records reviewed, and formulation updates. Writing a thorough session note from memory an hour after the session — or, more realistically, at the end of a day in which seven or eight sessions have happened — produces notes that are thinner than the clinical reality warrants.

Risk assessments. When a client presents with risk indicators — suicidal ideation, self-harm, risk to others — a structured risk assessment must be documented contemporaneously. The quality of that documentation matters for clinical decision-making, supervision, and, in the worst cases, legal review. Risk assessments written from memory at the end of a busy day are a documented patient safety concern.

Clinical formulations. A comprehensive clinical formulation — the theoretical framework explaining a client's presenting difficulties — is one of the most cognitively demanding documents in clinical psychology. It should reflect the nuance of the client's presentation across multiple sessions.

Supervision notes. Clinical supervision is both a professional requirement and a significant source of learning. The supervision conversation — what the supervisor said, the cases discussed, the clinical reasoning that emerged — is valuable material that most practitioners summarize very briefly because note-taking during supervision is awkward.

All of this happens after a working day spent in intensive relational work. The documentation system for mental health practitioners was designed without regard for the cognitive load that system imposes.

Why cloud AI tools are inappropriate for clinical mental health use

The professional and legal framing for mental health data is clear:

GDPR Article 9 defines mental health information as special category data requiring the highest protection standard. A client's disclosure about their mental health history, their suicidal ideation, their trauma history — all of this is Article 9 data. Transmitting it to a cloud AI server for transcription processing without a robust lawful basis, a DPIA, and explicit client disclosure is not compliant. More importantly: it is not the kind of data handling a client would agree to if they understood what was happening.

Professional body guidance. The British Psychological Society, BACP, and UKCP all have guidance on technology use in practice that emphasizes the practitioner's responsibility to ensure client data is handled in ways the client would consent to if fully informed. Cloud AI processing of session content is a significant disclosure that the professional guidance requires practitioners to consider carefully.

The US CLOUD Act. For UK practitioners using US-headquartered AI tools, the US CLOUD Act permits US government agencies to compel US companies to disclose data held on their servers, regardless of where those servers are physically located. For a client whose safety or legal situation is sensitive — a survivor of domestic abuse, a client with asylum claim complications, a client involved in legal proceedings — this is not a theoretical risk.

Therapeutic relationship damage. Beyond legal compliance: if a client discovered that their session content had been transcribed by a cloud AI service they hadn't consented to, the damage to the therapeutic relationship would likely be irreparable. The risk is not just regulatory — it is clinical.

What on-device processing changes

An AI tool that processes audio entirely on the practitioner's device changes every one of these considerations.

No data leaves the device. The client's words, the session content, the risk disclosures — none of it transmits to a server. There is no cloud processing event. There is no third-party data processor. There is no US company receiving the audio. The only destination for the session content is the practitioner's phone.

The consent conversation is honest and simple. Before using any recording tool in a clinical context, client consent is required. With an on-device tool, the consent conversation is: "I use a note-taking app on my phone to help me produce accurate session notes. It works entirely on my phone — the recording never goes anywhere else. Is that something you're comfortable with?"

This is a truthful statement that clients can understand and make an informed decision about. Compare it to the disclosure required for cloud AI processing: "I use an AI transcription service that processes audio on remote servers, operated by a US company, under their data retention policy." The first conversation supports informed consent. The second conversation is not one most practitioners have, because if they did, many clients would decline.

No DPIA required for a third party. On-device processing means no third-party data processor. No DPA to negotiate. No cross-border transfer to assess. The practitioner's phone is already within their information governance boundary — it holds their calendar, their email, their clinical scheduling system. An on-device AI tool fits within that existing framework rather than creating a new exposure to assess.

The Kuulo workflow for mental health practitioners

Session notes. With client consent, record the session on-device. Kuulo transcribes in real time on the Neural Engine of the iPhone's Apple Silicon chip — no internet connection required. After the session, generate the structured note: presenting issue, therapeutic content, client response, risk indicators reviewed, plan for next session. Review, edit, and copy to your clinical record.

The benefit over retrospective note-writing is not just speed — it is accuracy. The note generated from a recording of the actual session contains what the client actually said, not a memory reconstruction that compresses and loses nuance. For risk documentation in particular, the difference between "client expressed passive suicidal ideation" and "client reported fleeting thoughts of not wanting to be alive, denied intent or plan, safety planning reviewed" is clinically significant. The recording captures the latter.

Risk assessments. Generate a structured risk assessment note from a session recording in which risk was discussed. The template prompts for: nature of risk, frequency and intensity, access to means, intent and plan, protective factors, safety planning agreed, clinician's assessment, and next steps. Contemporaneous, structured, accurate.

Supervision. Record the supervision session with supervisor consent. Generate attributed notes that separate the supervisor's clinical observations from the supervisee's case presentations. A 60-minute supervision session becomes a 5-minute structured summary — cases discussed, clinical recommendations made, learning identified — without the awkwardness of typing notes while your supervisor is talking.

A note on professional registration and AI use

No UK professional body has issued categorical guidance prohibiting AI notetaking in clinical practice. The relevant professional standards — BPS Ethics Code, BACP Ethical Framework, GMC Confidentiality guidance for clinical psychologists in NHS settings — require practitioners to ensure client data is handled appropriately and that clients are informed.

On-device AI processing is, in principle, consistent with these standards when:

• Client consent is obtained before recording
• The practitioner can truthfully describe to the client how the tool works and where data goes
• Data is stored securely on the device and not shared inappropriately
• The tool does not substitute for clinical judgement in risk assessment or formulation

This is a developing area of professional guidance. Practitioners should consult their professional body's current guidance and, where uncertain, discuss with their clinical supervisor or professional indemnity provider.

What is clear is that on-device processing — where audio never leaves the practitioner's device — creates a significantly cleaner professional position than cloud processing. The practitioner can describe accurately what happens to client data, and that description is consistent with the principles underlying all professional guidance on confidentiality: client information is held by the clinician, protected by the clinician, and disclosed only when the clinician decides it should be.

The documentation burden and practitioner wellbeing

The connection between administrative burden and practitioner burnout in mental health services is well-documented. Writing session notes from memory at the end of an emotionally demanding clinical day is a specific form of burden — it requires cognitive precision at the moment when cognitive resources are most depleted.

A tool that generates a structured draft from the recorded session — a draft the practitioner reviews, corrects, and signs off rather than writes from scratch — is a different cognitive task. The review of an accurate draft is faster, less taxing, and produces better documentation than retrospective reconstruction from memory.

This is not about replacing clinical judgment with AI. The formulation, the risk assessment, the therapeutic decision-making — these remain entirely the practitioner's. What AI handles is the transcription of what was said and the first-pass structuring of the session note. The clinician's expertise goes into reviewing, correcting, and supplementing what the AI produced — applied to a draft that is already 80% of the way there, rather than to a blank page.

For a practitioner seeing 40 clients per week in an NHS IAPT service, the difference between 40 session notes written from scratch and 40 session note drafts reviewed and corrected is several hours per week — hours that currently erode evenings and contribute to the burnout that is taking experienced therapists out of the NHS at an unsustainable rate.

The tool is free to start. The architecture is the only one consistent with the confidentiality standards that underpin therapeutic practice. The only remaining question is whether to try it on the next session note that needs writing.