AI Progress Notes for Therapists: What to Automate and What to Keep

AI progress notes are drafts written by software from a session recording, transcript, or short summary. You review the draft, correct it, and sign it. Used well, they move the blank-page part of documentation off your evening. Used carelessly, they put words in the record that you did not mean.

The useful question is not whether to use AI for progress notes. It is which parts of the note software should draft, and which parts must stay yours.

What an AI progress note draft does well

Software is good at the parts of a note that are recall and structure:

• Chronology. What came up, in what order, without you reconstructing it from memory at 9pm.
• Structure. Putting content under the right SOAP, DAP, BIRP, or GIRP heading. If you need a refresher on the formats, see our therapy progress note template.
• Specific language. The client's own phrasing for a key moment, which is often more precise than a paraphrase written hours later.
• Interventions and homework. Naming what you did and what you agreed, so the plan section is complete.
• Consistency. The same headings and level of detail every session, which helps supervision and coverage.

What stays with you

Some parts of a note are clinical judgment, and judgment belongs to the clinician who was in the room:

• The assessment. What the session means clinically, whether the client is improving, and why.
• Risk. A draft can preserve what the client said about safety. You decide what it means, what you assessed, and what you did.
• Diagnosis and medical necessity. These carry billing and legal weight. They should never be accepted from a draft without thought.
• What to leave out. Not everything said in a session belongs in the record. A transcript-based draft can over-include detail that does not serve the client.

A simple rule: let the AI draft what happened, and write or confirm what it means.

A worked example

A client near the end of a session says: "Honestly, some nights I think everyone would be better off without me. I wouldn't do anything, though."

A weak draft compresses that to "Client reports low mood at night." The risk content is gone.

A good draft keeps the clinical shape: "Client reported passive thoughts that others would be better off without them, occurring at night. Denied intent or plan." It then leaves the assessment to you.

Your edit adds what only you know: "Risk assessed. No intent, plan, or access to means reported. Protective factors: daughter, upcoming trip. Safety plan reviewed and updated. Low acute risk." That sentence is the reason the note exists. No draft should write it for you.

Five checks before you sign an AI draft

1. Facts. Names, dates, medications, and measure scores are correct.
2. Attribution. Statements are assigned to the right speaker. A client quote written as your observation, or the reverse, changes the meaning.
3. Risk. Anything about safety is captured accurately, and your assessment and action are recorded.
4. Interventions. The techniques listed are the ones you used, named the way you would name them.
5. Minimum necessary. Sensitive detail that does not serve care is trimmed.

If a draft fails the first three checks often, the tool is costing you time instead of saving it.

Consent and recording

Recording a session is a clinical and ethical decision, not only a software setting. Tell clients what is recorded, what the AI does with it, who can see the result, and how long the audio and transcript are kept. Get consent in writing at intake and make it easy to say no.

Plan for the sessions you will not record. A good tool lets you type a short reflection after the session and drafts the note from that, so a client who declines recording does not mean a slower note for you. For the privacy side of choosing a vendor, read our HIPAA-compliant AI scribe checklist.

General scribe or therapy scribe

Many AI scribes were built for medicine and pointed at therapy later. They tend to default to medical SOAP framing, compress the relational content of a session, and treat each visit as a one-off. A scribe built for mental health supports therapy formats and treats risk language as first-class. We go deeper in why a mental health AI scribe beats a general one.

How Citt.ai handles AI progress notes

In Citt.ai you record in the room, capture a telehealth call with the browser extension, or upload a recording. You get a transcript with speakers separated, then a draft in SOAP, DAP, BIRP, GIRP, psychodynamic, or your own uploaded format. If you did not record, you type a reflection and Citt drafts from that.

Drafts stay drafts until you approve them. After you approve a note, later edits create a new signed version and the earlier one stays in the history. Approved notes then feed a summary of the client since the last session, so the documentation you sign also prepares your next hour. See AI therapy notes in Citt.ai or the therapist workflow.