The AI Stack Therapists Actually Use in 2026 (Notes, Practice Management, Where to Draw the Line)

Walk into a group practice at 7pm in 2026 and the clinician offices are dark. Five years ago every one of them would have had a therapist still hunched at a screen, writing the progress notes from the day's eight sessions while the kids ate dinner without them. That tail — the unpaid third shift every full-caseload therapist used to work — is finally gone.

It didn't go away because therapy got faster. The hour is still the hour. It went away because a narrow stack of clinical AI tools got good enough at the documentation, the intake summary, the treatment-plan boilerplate, and the insurance-shaped progress note that the clinician now signs out the chart between sessions instead of after midnight.

This is the stack I see on practicing therapists' laptops and phones in 2026 — private-practice LMFTs, group-practice LCSWs, community-mental-health psychologists, and the psychiatry-adjacent therapy world. The kit differs at the edges. The center is the same.

The session didn't change. The chart did.

The therapy hour is still 50 minutes of two humans in a room (or on a screen). The AI doesn't sit in that chair and shouldn't. Where it sits is upstream and downstream: intake forms, the progress note, the treatment plan, the insurance authorization, the outcomes measure, the no-show follow-up email, the client psychoeducation handout. Seven jobs around the session. AI moved from "interesting demo" to "the practice would fall apart without it" in six of them in eighteen months.

1. AI scribes for therapy — the breakout category

If you adopt one thing on this list, this is it. Ambient AI scribes listen to the session (with consent), draft the progress note in the format your payer expects, and let you sign out the chart in the five minutes between clients instead of in your kitchen at 10pm.

• Mentalyc — the household name in private-practice therapy AI. Trained specifically on therapy sessions; produces SOAP, DAP, and BIRP notes; flags the elements an auditor looks for. The reason every Facebook therapist group has a thread about it.
• Upheal — strong in the telehealth-native practice. Integrated session recording, transcription, AI progress notes, and treatment-plan suggestions in one workflow. The pick if you're already doing video sessions and don't want a second tool just to record them.
• Blueprint Health — measurement-based-care first. The note is good; the differentiator is the outcomes dashboard (PHQ-9, GAD-7, etc.) that auto-administers between sessions and surfaces trends. The pick if you actually want to know whether your client is getting better.
• Eleos Health — community-mental-health and large-group-practice play. Behavioral-health-tuned ambient documentation plus quality-improvement analytics across a clinician team. The choice when the CFO is signing the check, not the LMFT.

The trap: every one of these produces a draft. Sign nothing without reading it. A hallucinated suicide-risk assessment that you signed is a board complaint waiting to happen, and the 2026 malpractice carriers are starting to ask whether you reviewed every AI-generated note before co-signing.

2. Practice management — the non-clinical half of the job

• SimplePractice — the dominant private-practice EHR. AI features keep getting layered on: intake-form summarization, smart scheduling, treatment-plan templates. If you're solo or small-group, this is probably already your spine.
• TherapyNotes — the SimplePractice alternative. Stronger billing workflow, lighter telehealth. AI features arrived later but are catching up.

Neither replaces a scribe — they live alongside one. The progress note from your scribe gets pasted into the EHR; the EHR handles the claim.

3. The general-purpose layer

• ChatGPT and Claude — for the client-psychoeducation handout at a 6th-grade reading level, the referral letter that needs to be diplomatic, the treatment-plan first draft, the supervision-prep notes. Use the business/enterprise tier. Never paste a client name or identifying detail into a consumer endpoint — the HR & recruiting stack lessons on PII apply double here.
• Perplexity AI — for the cited public-knowledge question. "Latest evidence base for ERP in pediatric OCD" beats a Google rabbit hole every time.
• Otter.ai — for the non-clinical meeting (consultation group, supervision, case conference). NEVER for client sessions — use a clinical scribe and obtain consent. Otter has no HIPAA BAA on the consumer tier.

The AI Writing Tools That Don't Sound Like a Robot post covers the broader writing surface for client comms.

A warning about consumer chatbots

Woebot, Wysa, Replika — these are *consumer* tools that some clients will arrive having used. Know what they are. They are not therapy. They have utility as between-session skill-practice for some clients (Woebot is the most clinically grounded of the three) and they are actively harmful for others (Replika in particular has produced documented cases of parasocial attachment that interfere with treatment).

The right move in 2026 isn't to ban them. It's to ask about them on intake, the same way you'd ask about any other thing the client is doing between sessions, and document the answer.

What AI still can't do in this work — and a therapist should never let it

• Be the therapist. The alliance is the intervention. The model can transcribe; it cannot hold a person who is grieving.
• Assess risk. The model can produce a checklist; it cannot read the slump of someone's shoulders when they say "I'm fine." Suicide and homicide risk assessments are a human judgment under license. Period.
• Diagnose. The progress note draft can suggest a code. You assign it. Your license is on the form, not the model's.
• Hold confidentiality. The duty is yours. Read your scribe vendor's BAA. Read where the audio is stored, for how long, and what happens to the transcript. If your vendor can't answer those three questions in a sentence each, do not use them.
• Have the goodbye. Termination is a clinical intervention, not a follow-up email.

The parallel with the doctors' stack is exact: the AI deletes the part of the work that was never the work, and gives the clinician back the hours to be present in the part that was. Therapists who use this stack well in 2026 are not seeing more clients. They are seeing the same caseload and going home for dinner.

Try Them Yourself

• AI Tools for Doctors 2026 — yesterday's post; the adjacent clinical stack
• AI Writing Tools That Don't Sound Like a Robot — for client psychoeducation and difficult letters
• Perplexity vs ChatGPT Search — picking the right citation-grounded research assistant
• Healthcare category — the full clinician toolkit
• ChatGPT vs Claude — picking the general-purpose layer

Start here: Mentalyc or Upheal for the progress note, SimplePractice or TherapyNotes for the chart spine, ChatGPT or Claude (business tier only, no PHI in prompts) for the writing surface, Perplexity for the evidence question between clients. Five tools the model handles, four roles the therapist keeps — the alliance, the risk call, the diagnosis, and the signature on the chart.