AI Progress Notes Case Study: 81 Notes Signed by Therapists in 11 Days

Summary

A US behavioral health practice group put an AI progress-note assistant into daily use, and in the first 11 days in production 9 clinicians signed 81 AI-drafted notes. 68% of those notes were signed with no edits, a draft took a median of under one minute to review, and each draft cost about $0.14 in AI usage. Nirmitee built the assistant inside the practice's own EHR and billing platform, with the clinician deciding every word.

Who the client is

The client is a behavioral health practice group in the United States. In August 2026, 49 clinicians signed 1,721 session notes on the practice's own EHR and billing platform. Most appointments are 50-minute therapy sessions billed to commercial insurers and Medicaid.

The practice runs its own platform rather than an off-the-shelf EHR. That gave it full control over its data, and it also meant every new feature, including AI, had to be built and proven inside that platform.

The problem: the paperwork did not fit in the day

Every therapy session ends with paperwork that has to be finished before the claim can go out. Each session needs a progress note in the clinic's format (SOAP, DAP or BIRP). New clients need an initial assessment, every client needs a current treatment plan, and associates working toward a licence need supervision notes. The note is the clinical record, the supervisor signs it, and the claim depends on it.

• Notes ran late. Before the AI went live, half of all notes were signed more than 22 hours after the session, and one in three was signed in the evening, at night or at the weekend (1,558 notes, December 2025 to August 2026).
• The AI the platform had did not work. The platform already had code for an AI integration, but it had never been configured. It had never produced a single draft.
• Diagnosis codes were costing money. In a 10-day sample of the practice's clearinghouse reports, diagnosis coding was the top reason claims bounced: 46 rejections for an invalid diagnosis and 63 diagnosis-code corrections.

Behavioral health teams carry some of the heaviest documentation loads in healthcare. Our mental health software solutions start from that workload.

What we built, step by step

Nirmitee built an AI note assistant into the screens clinicians already used, then wrapped it in the controls a compliance lead needs before letting AI near a clinical record.

Step 1. Start where clinicians already work

The assistant lives inside the existing note screen, so there is no new app to log into. The clinician types a few lines about the session, presses one button, and a full draft appears in the clinic's format in about 48 seconds.

Step 2. Ground every draft in the client's own chart

The draft is built only from the clinician's summary and the client's own record: treatment plan goals and the previous note. Each section points back to its source. If the sources do not support a line, the assistant flags it as low confidence instead of writing it as fact.

Representative screen with invented sample data. Each section shows its sources and is accepted or edited on its own.

See how our healthcare AI agents keep a clinician in control of every step of a clinical workflow.

Step 3. Keep the clinician in charge

The draft is never the note. The clinician accepts it, edits it or throws it away. Possible risk signs appear as prompts for the clinician's judgement, and the assistant never makes a risk decision.

Step 4. Suggest diagnosis codes, never choose them

The assistant suggests ICD-10 codes, the standard diagnosis codes insurers use, each with a reason. It only suggests codes from the chapters the practice bills, and every code is checked against the official code list, so a code the model made up is dropped before anyone sees it. Nothing reaches the assessment or a claim until the clinician accepts it.

Step 5. Cover the rest of the paperwork

The same assistant drafts initial assessments, treatment plans and supervision notes. Only the assigned supervisor can finalise a supervision note, and the associate sees it only after that.

Step 6. Consent before any recording

For telehealth sessions, recording starts only when consent is on file for every person in the session. Audio goes to speech-to-text and nowhere else. The AI model only ever reads text.

Step 7. Give the practice the switches

The practice turns the assistant on or off by organisation, clinician and role, and an organisation switched off stays off whatever a user setting says. The practice sets how long audio is kept. Every instruction given to the model is versioned, can be rolled back, and is recorded in the audit trail.

Discover our healthcare product engineering approach to adding AI to a live clinical platform without disturbing sign-off or billing.

Step 8. Prove the safety rules with tests

91 automated tests (72 integration and 19 unit) prove that the assistant cannot change a signed note, has no code path into claims or billing, follows supervision rules and never writes clinical text into a log file. All 91 pass on the current code.

Step 9. Go live only with written approval

The model runs on Claude inside the practice's own Amazon Web Services account, under its existing business associate agreement. No change of vendor, model, region or data handling reaches production without the practice's written approval, and that rule applied to our own change first.

The outcome: the first 11 days in production

Between 19 and 29 August 2026, clinicians created 128 progress-note drafts and signed 81 of them. The numbers below come from the platform's own records for that window.

• Clinicians used it from day one. 16 staff tried the assistant, and 9 clinicians signed 81 notes with it: 12% of all notes signed in that window.
• Clinicians trusted the drafts. 63% of drafts were signed, 68% of signed notes needed no edits, and only 5 drafts (4%) were thrown away.
• Review is fast. A draft signed with no edits took a median of under one minute to review. A draft the clinician edited took a median of two minutes.
• It costs cents. One progress-note draft costs about $0.14 in AI usage. All AI use in the first 11 days cost about $25.

What this means for the practice: the formula

Clinician hours back per month equal notes per month, times the share of notes drafted with AI, times the minutes saved per note, divided by 60.

• Notes per month: 1,721 (August 2026).
• Share drafted with AI: 12% in the first 11 days. Our goal for month three is 60%.
• Minutes saved per note: 11. A therapy progress note written by hand takes about 15 minutes (published ranges run from 10 to 30). With the assistant, the clinician spends about 3 minutes typing a summary and 1 to 2 minutes reviewing the draft.

At today's 12%, that is 1,721 × 0.12 × 11 ÷ 60, or about 38 clinician hours a month. At 60% of notes it is about 189 hours a month, roughly four hours per clinician, for about $145 a month in AI usage.

Learn how our healthcare AI solutions are measured against the practice's own data from the first week.

What we learned

Three findings from the first 11 days changed what we build next, and they apply to any practice adding an AI scribe.

• Clinicians chose typing over recording. Every signed draft in the window came from a typed summary, not a session recording. Three typed lines felt quicker and safer. Build the typed path first.
• Early adopters were already fast. The clinicians who already finished notes quickly picked up the assistant first, so faster turnaround across the whole practice needs more than a draft. The next step is a prompt at the end of each session with the draft waiting.
• Diagnosis suggestions were mostly ignored. Of 215 code suggestions, 29 were accepted and 11 rejected, and the rest were never opened. They sit too far from where clinicians pick codes. Moving them is the next fix, aimed straight at the claims that bounce for diagnosis codes in the practice's revenue cycle.

Step-by-step guide: an AI note assistant your compliance team will sign off

1. Start inside the note screen. No new app, one button.
2. Ground every draft in the client's own chart. Each line points to its source.
3. Flag what the sources do not support. Show low confidence, never hide it.
4. Suggest, never decide. Diagnosis codes and risk signs are prompts for the clinician.
5. Lock signed records. The AI must not be able to touch anything already signed.
6. Keep the AI out of billing. No code path from the AI to claims.
7. Consent before recording. Proven per session, for every person in it.
8. Give the practice the switches. On or off by organisation, clinician and role.
9. Version your AI instructions. Treat them like code, with approval and rollback.
10. Prove the rules with tests. One test for every safety rule.
11. Measure from day one. Log every draft, review and edit so you have real numbers in week two.

Technology

.NET 8 API, Angular 20 front end, SQL Server, Claude on Amazon Bedrock for drafting, Amazon Transcribe for speech-to-text, and Amazon S3 for session audio.

Want an AI scribe your clinicians will actually sign?

If your clinicians are taking notes home, we can review your current platform and show where an assistant fits, what it should never be allowed to do, and how to measure it in the first two weeks. Our custom healthcare software development team builds this inside the system you already run. Book an architecture review with our team.