Briefing 21 · U.S. State Law

Colorado’s new AI-therapy line: assistant, not therapist.

U.S. state AI regulation · ~6 min read

Colorado just drew a line that a lot of AI products are quietly standing on. As of August 12, 2026, HB 26-1195, the Psychotherapy Artificial Intelligence Restrictions, makes it unlawful for AI to do the actual work of therapy in Colorado. It can help with the paperwork. It cannot be the clinician. If your product lives anywhere near mental-health conversations, the question is now which side of that line it sits on, and whether you can prove it.

What the law does

Signed June 3, 2026 and effective August 12, 2026, HB 26-1195 restricts how licensed, certified, or registered psychotherapy providers in Colorado can use AI in care. In plain terms, an AI system may not:

Engage in therapeutic communication directly with a patient. The conversation that is treatment has to be human.
Generate treatment recommendations or plans without human review. A clinician cannot hand the clinical judgment to a model and pass its output along unsupervised.
Detect a patient’s emotions or mental state without professional oversight. Affect and mental-state inference, on their own, are off limits.

What it does not ban is just as important. AI is still allowed for administrative work, and it can transcribe a session as long as the patient consents. The statute is enforced by Colorado’s Department of Regulatory Agencies (DORA) through the licensing boards that already govern mental-health professionals.

The line it draws: assistant, not therapist

Strip away the specifics and Colorado has codified a distinction the whole market has been blurring: the difference between AI that supports a licensed professional and AI that quietly replaces the professional’s judgment. Scheduling, notes, transcription, drafting: fine, with consent where required. Doing the talking, forming the plan, reading the patient: that is practicing therapy, and in Colorado a machine may not do it.

That is the same instinct now showing up across health-AI rules and the companion-chatbot laws: a human has to own the consequential moment. Colorado has simply made it explicit for one of the most sensitive interactions there is.

Who should read this even if they are not in Colorado

The direct duty falls on Colorado psychotherapy providers. But the reach is wider than the license:

Digital-health and wellness apps. A “supportive companion,” an “AI coach,” a mood tracker that offers guidance: features like these can drift from wellness into something a regulator would call therapeutic communication. The label on your app is not what decides it.
Employers and platforms offering mental-health tools. If you put an AI mental-health chatbot in front of employees or users in Colorado, you are now in a space with a bright line through it.
Vendors selling into clinical settings. Your healthcare customers now have to keep a human in the loop. Your product needs to make that easy to prove, not hard.

And it rhymes with the companion-chatbot wave we track separately: California’s SB 243, the state bills moving this summer, and even China’s national companion-AI rules all circle the same nerve, which is what an AI is allowed to do when a vulnerable person is on the other end.

What to do now

1. Classify the interaction, not the marketing. For each AI feature near mental health, decide honestly: is this administrative support, or is it therapeutic communication, a treatment recommendation, or mental-state detection? 2. Put a human in the consequential loop, on the record. Where a clinician must review or own the output, build that step in and log it, so oversight is evidence, not a promise. 3. Fix consent for anything that listens. Session transcription needs patient consent under this law; make sure your flow captures it. 4. Map your footprint by state. Colorado is first, not last. Know where your tool operates and treat this line as the direction of travel.

An honest limitation

We are describing an enacted Colorado statute (HB 26-1195) with an August 12, 2026 effective date, summarized from the bill and reputable reporting; the precise boundaries of “therapeutic communication” and “mental-state detection” will be worked out by DORA and the licensing boards in practice, so verify the current text and any guidance before you rely on a specific reading. This is also a licensure-based rule aimed at psychotherapy providers, not a general ban on wellness apps. The durable point stands regardless: when an AI system sits close to a vulnerable person’s care, the law increasingly insists a qualified human stays responsible, and it increasingly expects you to prove it.

This briefing is general information from Sentinel Assurance Group, not legal advice. Regulatory dates and requirements change, so verify against primary sources and counsel before acting. Last reviewed August 3, 2026.

See how a Gap Assessment maps your exposure →

Not sure if your AI crosses the line?
Find out in 30 minutes.

The free AI Risk Exposure call maps your AI features to the line Colorado just drew, and tells you plainly where a human needs to stay in the loop.