The session notes get written. The clinician stays in charge of what they say.

How it works

1. Consent first. Recording, AI drafting and any product-improvement use are three separate agreements. Each is off by default, each is specific, and each can be withdrawn. Declining is never a condition of care.
2. The session is transcribed with speaker labels and timestamps.
3. Facts are extracted — what was said, by whom, with a link back to the exact moment in the transcript.
4. A draft is written from those facts. Every sentence cites a fact. Every fact cites the transcript. No citation, no sentence.
5. A separate check runs — a different model, a different prompt, looking for claims the transcript does not support. The software that wrote the note never grades its own work.
6. The clinician reviews and signs. Click any sentence to see the moment it came from. Anything the software flagged has to be resolved first.

What it will never do

These are product boundaries, not settings. They came out of review by licensed clinicians, and changing one takes clinical sign-off.

• —Assign a risk level. The two clinicians who reviewed our drafts independently named this as the most dangerous thing software could write. We record what was said about safety, in the client's own words. The clinician assesses risk.
• —Suggest a diagnosis. A diagnosis appears only if the clinician said it aloud, or it is already in the client's record. Never inferred from a session.
• —Infer mood, affect or emotion. Emotion inference from speech is unreliable, and in some jurisdictions it is prohibited outright. If the clinician did not say it, it is not in the draft.
• —Judge progress or insight. The software supplies the evidence. The clinician draws the conclusion.
• —Claim something did not happen. Silence in a recording is not proof that an assessment did not occur. Mentalley can report what is present. It never asserts an absence.
• —Talk to a client. No patient-facing chat, no therapy AI, and nothing that asks for the clinician's attention during a session. Recording is passive.

The part most tools get wrong

If I have to reread the entire session because I don't know whether the AI missed an important risk statement, intervention, observation or clinical decision, then the draft could actually create more work. — Licensed clinician, design review

That reframed the whole product for us. A draft that looks finished but might have quietly dropped something is worse than no draft at all, because it has to be checked line by line against the session anyway.

So Mentalley is built to show its gaps. When something in the session touched on safety, self-harm, abuse, substance use or medication and did not make it into the draft, that is surfaced next to the note. The value is not the writing. It is not having to listen to the session again.