The State of AI for Therapists in 2026

Two years ago, AI documentation tools for therapists were a curiosity. The early tools transcribed audio poorly, generated clinical language that read like a bad medical textbook, and made compliance officers nervous. Therapists who tried them were early adopters willing to tolerate unreliable output in exchange for the novelty of not writing SOAP notes manually.

Today, they are practice infrastructure. Mentalyc and Upheal have collectively processed tens of millions of clinical notes. The transcription quality on modern models is measurably better than what a therapist typing from memory produces. The note format flexibility has expanded to cover nearly every modality a practicing therapist encounters. And the compliance story — HIPAA, SOC 2, BAA availability — has standardized to the point where the question is no longer "is this tool safe to use?" but "which safe tool fits my practice?"

This is a meaningful transition. What follows is an honest assessment of where the market stands in mid-2026: what has matured, what remains genuinely uncertain, and what the next 18 months are likely to bring.

What has genuinely matured

Transcription accuracy is no longer the bottleneck

The earliest AI scribes spent most of their computational budget on transcription — converting session audio to text. The output quality was variable enough that therapists who saw complex clinical presentations (high affect, multiple speakers in couples sessions, significant accents, clinical terminology) found the transcription errors compounding into note inaccuracies.

That problem is largely solved. The current generation of purpose-built clinical transcription models handles clinical language, abbreviations, medication names, and multi-speaker scenarios reliably. In our 30-day testing, transcription errors that required substantive correction appeared in under 3% of notes by the third week. The residual errors were primarily proper nouns — client names, city names — and they were consistent enough that therapists adapted quickly.

The current bottleneck is not transcription. It is clinical calibration: generating a note that matches not just the session content but the therapist's clinical voice, their documentation standards, and their EHR's specific formatting requirements. The tools that have invested in template customization and format flexibility are advancing faster than those that optimized transcription alone.

Compliance infrastructure is standardized

Twelve months ago, a therapist evaluating an AI scribe needed to investigate each vendor's HIPAA compliance claims individually. BAA availability varied by plan tier. SOC 2 certifications were recent and sometimes provisional. Privacy policies were vague about what happened to session audio after processing.

The current market leaders have standardized this. Mentalyc and Upheal both offer signed BAAs on all paid plans, process and delete audio immediately after note generation, and hold SOC 2 Type II certification. Neither trains its models on user data. The compliance story is clear and verifiable.

This matters for the market because it removed the primary reason therapists in group practices and agency settings were blocked from adoption. Individual practitioners make tool decisions personally. Group practices and agencies require sign-off from compliance officers and legal teams. When the compliance story was inconsistent, institutional adoption stalled. Now that it has standardized, group practices are adopting these tools at scale — and that is where the volume of clinical notes actually lives.

Pricing has found its floor

The price range for AI documentation tools has settled into a predictable band: free tiers for minimal usage, $20–$70/month for solo practitioners, $50–$70/seat/month for teams. Upheal's $1/session model (capped at $69/month) introduced usage-based pricing into a category that was previously flat-rate only, and that pricing model has proven genuinely competitive for lower-volume practitioners.

What has not changed: there is still no tool that is both fully featured and meaningfully cheaper than this range. The "free forever" tools that appeared in 2024 have either pivoted to freemium (Upheal's free tier is genuinely useful, but it is also a funnel to paid plans) or quietly sunset. The cost of running clinical-grade transcription and note generation at scale is high enough that the economics of truly free products do not work without compromise on data practices or model quality.

What remains uncertain

The scope creep problem

AI documentation tools started as scribes. They are becoming something more ambitious. Both major platforms now offer features adjacent to clinical workflow that go beyond note generation: therapeutic alliance scoring (Alliance Genie in Mentalyc), treatment plan generation, goal tracking across sessions, CPT code computation, compliance checking.

Some of these features are genuinely useful. Alliance Genie, in our testing, provided meaningful data that prompted useful clinical reflection. The compliance checker in Upheal caught documentation gaps that would have been legitimately problematic in an audit.

Others are moving toward territory that raises clinical judgment questions. AI-generated treatment plan drafts are a time-saving tool when reviewed carefully by the clinician. But the documentation in some platforms now suggests a level of clinical authority — "AI-recommended treatment goals" — that conflates administrative assistance with clinical recommendation.

The distinction matters. A note-drafting tool that reduces administrative burden is a productivity product. A tool that suggests clinical interventions or treatment directions is something that requires a different level of scrutiny, clinical oversight, and liability analysis. The current generation is mostly still on the right side of that line, but the drift toward clinical feature expansion bears watching.

How licensing boards will respond

State licensing boards for mental health professionals have begun issuing guidance on AI tool use, but the guidance is early, inconsistent, and in most cases non-binding. The pattern so far: boards acknowledge that AI documentation tools are widely used, express concern about clinician review standards, and stop short of formal rules.

The formal rulemaking will come. When it does, the most likely requirements are: (1) explicit documentation that the clinician reviewed and verified AI-generated notes before signing, (2) client consent language that explicitly names AI documentation, and (3) possibly a sunset on the "AI-assisted" designation after a certain period to prevent clinicians from signing AI drafts without meaningful review.

None of these requirements would be unreasonable. Most responsible users of these tools already follow them. But therapists who have become dependent on AI documentation without maintaining the clinical discipline to actually review the output are at risk when formal requirements materialize.

The EHR integration gap

The most persistent friction point in the current market is the gap between AI documentation tools and EHR workflows. The best integrations — Mentalyc's native connections to SimplePractice, TherapyNotes, and Jane App — push completed notes into client records automatically. For practices using those three EHRs, the workflow is genuinely seamless.

For practices using other EHR systems — and there are hundreds — the workflow involves copy-paste. That is better than manual transcription, but it is not the zero-friction experience the marketing describes.

The deeper integration gap is bidirectional: notes can flow from documentation tool to EHR, but client context (diagnoses, active treatment goals, medication history) does not flow back from the EHR to inform note generation. A clinician who wants the AI to reference a client's ongoing treatment goals in the progress note has to manually set that context in a template — the tool does not read the EHR.

Bidirectional integration at the EHR level is technically feasible (FHIR APIs exist for this), but it requires EHR vendors to expose those interfaces and documentation tool vendors to build against them. Neither has prioritized this. It is the most likely area of meaningful product improvement over the next 18 months.

What comes next

The platforms that will win the next 18 months are those that solve two specific problems the current generation handles poorly.

Contextual continuity across sessions. Today's AI scribes treat each session as an independent input. A note generated for session 47 with a client does not automatically incorporate the treatment context from sessions 1 through 46. The longitudinal clinical record — the thread of a client's treatment over years — lives in the therapist's head and in the EHR, not in the documentation tool. The platform that builds genuine continuity — notes that reference previous goals, track symptom progression, and flag clinical patterns across a caseload — will be meaningfully ahead of tools that optimize only for individual session transcription.

Group practice workflow. Solo practitioners adopted first. Group practices and agencies are adopting now. But the tooling has not fully caught up to group practice workflow needs: shared note review, supervisor approval workflows, billing integration at the practice level, population-level outcome reporting. Mentalyc's Team plan and Upheal's Group plan are the beginning of this, not the completion.

The tools that exist today are genuinely useful for the problem they were designed to solve: reducing administrative documentation burden for individual therapists. That is a real problem worth solving, and the current generation solves it well.

The next problem — turning individual session documentation into a practice-level clinical intelligence layer — is larger, more technically complex, and likely more valuable to the practitioners who adopt it first.

We will be testing those tools when they arrive.

For current tool comparisons, read our Mentalyc review, Upheal review, and Mentalyc vs Upheal head-to-head.