Key takeaways
- AI in behavioral health is under real scrutiny: at least seven states now restrict AI therapy, the FTC has opened an inquiry into companion chatbots, and lawsuits over chatbot harm have been filed and settled.
- The safest uses are operational: billing, eligibility checks, scheduling, reminders, and alerts that route issues to a human.
- The riskiest uses are AI that delivers therapy, makes treatment decisions, presents itself as a clinician, or handles a crisis without a person involved.
- Before you sign, ask vendors about HIPAA, a BAA, 42 CFR Part 2 and a QSOA, human oversight, crisis escalation, data use, and state law compliance.
- Verity uses AI for automation, such as billing and eligibility workflows and alerts that route patient issues to the right staff member, while clinicians stay in charge of care.
AI in behavioral health is moving faster than the rules around it. Vendors promise AI that engages patients, answers questions, and even "provides support" between sessions, while states ban AI therapy, federal regulators open inquiries, and families sue chatbot makers. For executives and clinical directors at CCBHCs, SUD and OTP programs, outpatient clinics, and youth programs, the question is not whether to use AI. It is which uses are safe, which are hype, and what to ask a vendor before AI touches your patients. This guide covers the current scrutiny, a framework for safe and risky uses, and a vendor question checklist.
How Verity helps: AI for automation, with clinicians in charge
Verity uses AI where it is most useful and least risky: taking repetitive work off staff so they have more time for patients.
- Billing and eligibility.Verity AI Billing automates insurance eligibility verification and discovery.
- Alerts that reach a person.Workflow automation routes issues such as missed medications or low satisfaction to the right staff member, who decides what to do.
- Engagement built on proven methods.Verity for Patients uses points, streaks, quests, and a prize shop, contingency management with incentive caps and eligibility rules, and hundreds of expert-built exercises. Every exercise built by clinical psychologists, with new content added regularly.
- Measurement clinicians can see.PHQ-9, GAD-7, and 500+ validated assessments feed dashboards and post results to the EHR.
- Compliant communication.Broadcast messaging by text, email, or app with consent-aware sending.
- Privacy and integration. Verity works with any EHR by API or SFTP, with 200+ integrations, signs a BAA, and signs a QSOA where 42 CFR Part 2 applies.
What can AI realistically save a behavioral health program?
The table below is an Illustration, not a statistic. It uses hypothetical numbers to show where automation saves time: by surfacing exceptions instead of asking staff to check every chart.
| Illustration, not a statistic | Manual review | Automated alerts |
|---|
| Active patients checked each week | 200 | 200 |
| Patients a person has to review | 200 | 20 flagged exceptions |
| Minutes per review | 3 | 3 |
| Staff hours per week | 10 | 1 |
| Staff hours per year | About 520 | About 52 |
In this example a person still reviews every flagged patient. The AI decides what to surface, not what to do. That distinction is the core of safe AI in behavioral health.
Why is AI in behavioral health under so much scrutiny?
Because consumer chatbots have been linked to serious harm, and regulators have responded quickly.
- State laws. In August 2025, Illinois enacted a law that prohibits using AI to make mental health and therapeutic decisions, while allowing AI for administrative tasks and supplementary support for licensed professionals, with fines of up to $10,000 per violation. Nevada's AB 406, effective July 1, 2025, prohibits offering AI systems that provide or claim to provide professional mental or behavioral health care. Utah added disclosure rules for mental health chatbots, and California's SB 243 added safety and disclosure duties for companion chatbots starting January 1, 2026.
- More states in 2026. Colorado, Maine, Rhode Island, Tennessee, and Vermont enacted laws in 2026 restricting AI therapy chatbots, some of which also limit how licensed clinicians can use AI.
- The FTC. On September 11, 2025, the FTC issued orders to seven companies that run consumer AI companion chatbots, asking how they test for harm, protect minors, and handle complaints.
- The FDA. On November 6, 2025, the FDA's Digital Health Advisory Committee met on generative AI mental health devices. As of that meeting, the agency had not cleared any mental health tool using generative AI, and panelists called for human escalation plans, one-tap crisis escalation, and adverse event reporting.
- Lawsuits. In November 2025, seven lawsuits were filed against OpenAI alleging ChatGPT contributed to suicides and harmful delusions. In January 2026, Character.AI and Google agreed to settle lawsuits brought by families of teens who self-harmed or died by suicide.
- Professional guidance. The American Psychological Association's November 2025 health advisory says AI chatbots and wellness apps should not be used as a replacement for a qualified mental health care provider.
State laws are changing quickly and vary in scope, so confirm current requirements with counsel in each state where you operate.
Which AI uses in behavioral health are safe, and which are risky?
| Use | Risk level | Why |
|---|
| Eligibility checks, billing, claims, scheduling | Lower | Administrative; errors are financial and fixable |
| Reminders and appointment messages | Lower | Templated content with consent controls |
| Alerts that flag patients for staff review | Lower to moderate | A person makes the decision; watch for missed flags |
| Draft notes or summaries a clinician reviews | Moderate | Useful, but needs review, consent, and accuracy checks |
| AI-generated messages sent to patients without review | High | Can give clinical advice no one approved |
| Chatbots that provide therapy or present as clinicians | Highest | Restricted or banned in several states; no FDA-cleared generative AI tools as of late 2025 |
| AI handling crisis or suicide risk without a person | Highest | Patient safety risk; the core issue in current lawsuits |
A simple test: if the AI's output reaches a patient as clinical guidance, or decides on treatment, without a licensed person approving it, treat it as high risk.
What should you ask AI vendors before you sign?
Use these questions in every evaluation, and get the answers in writing.
Privacy and compliance
- Will you sign a BAA? Which subcontractors, including AI model providers, touch our data, and do they have BAAs too?
- Do you handle 42 CFR Part 2 records, and will you sign a QSOA? HHS's updated Part 2 rule had a compliance date of February 16, 2026, so ask how the product supports it.
- Is our patient data used to train your models or anyone else's? Can we opt out?
- Where is data stored, how long is it kept, and how is it deleted?
- Which security reviews or audits can you share?
Human oversight
- Exactly what does the AI decide, and what does a person decide?
- Does any AI-generated content reach patients without staff review?
- Can clinicians override, edit, or turn off each AI feature?
- Is every AI action logged and auditable?
Crisis and safety
- What happens if a patient mentions suicide, self-harm, or a crisis in any message or entry?
- Who gets alerted, how fast, and what does the patient see?
- How do you test for harmful or wrong outputs, and how are incidents reported to us?
Evidence and claims
- Is the product FDA-cleared? If not, does it make any treatment claims?
- What evidence supports the outcomes you advertise, and was it independent?
- Does the product describe itself to patients as a therapist, counselor, or clinician?
State law and youth
- Which state AI laws have you reviewed for our locations, and how does the product comply?
- How do you handle minors, including age checks, parental access, and disclosures?
A board-ready summary
- The landscape: AI therapy is restricted in a growing number of states, under FTC inquiry, not FDA-cleared in generative form, and the subject of lawsuits.
- Our position: use AI for administrative work and alerts that support staff; do not use AI to deliver therapy or make treatment decisions.
- Controls: BAAs and QSOAs, no training on our data without approval, human review of anything clinical, and tested crisis escalation.
- Oversight: an inventory of every AI feature in use, reviewed by compliance and clinical leadership each year.
AI governance checklist
- Inventory every AI feature in current tools, including ones vendors added quietly.
- Classify each use as administrative, decision support, or patient-facing.
- Confirm a BAA, and a QSOA where Part 2 applies, for every vendor and subcontractor.
- Require human review for anything clinical that reaches a patient.
- Test crisis escalation paths with real scenarios.
- Check state AI laws for each location and update consent and disclosure language.
- Train staff on what AI tools do and do not do.
- Review incidents and vendor changes at least quarterly.
Edge cases to plan for
- Youth programs. Minors are the focus of the FTC inquiry and of California's companion chatbot law. Apply the strictest standard to youth programs.
- SUD and OTP programs. Part 2 records need consent and a QSOA, and AI subcontractors must be covered. See Verity for substance use treatment programs.
- Multi-state and telehealth programs. The law where the patient is located may apply. Map your patients' states, not only your offices.
- AI features added to existing tools. EHRs and messaging tools are adding AI. Review release notes and turn off features you have not approved.
- Staff using consumer AI. Clinicians pasting notes into public chatbots is a privacy risk. Set a clear policy and offer approved tools.
Common mistakes and how to fix them
- Buying the demo, not the controls. Fix: ask for the oversight, logging, and escalation design before features.
- Assuming a BAA covers everything. Fix: confirm Part 2, model providers, and data training terms separately.
- Calling a chatbot "support" to avoid rules. Fix: judge the tool by what it does, not what it is called.
- No owner. Fix: name one leader accountable for AI governance across clinical and compliance teams.
- Ignoring state law changes. Fix: review new laws with counsel at least twice a year.
Recap
AI in behavioral health is safest when it takes work off staff and leaves clinical decisions to people: billing, eligibility, reminders, and alerts that reach a human. It is riskiest when it delivers therapy, presents as a clinician, or handles a crisis alone. Ask vendors about BAAs, Part 2 and QSOAs, human oversight, crisis escalation, and state law before you sign. Book a demo to see how Verity uses automation to support your team while clinicians stay in charge.
Does Verity integrate with your EHR?
Verity works with any EHR and connects to 200+ systems through API or secure file transfer, so appointments, attendance, and assessment results flow automatically.
See all integrations →
Frequently asked questions
Is AI therapy legal in the United States?
It depends on the state. Illinois and Nevada restricted AI therapy in 2025, and Colorado, Maine, Rhode Island, Tennessee, and Vermont enacted restrictions in 2026. Others regulate chatbot disclosures and safety. Several, including Illinois, still allow AI for administrative work. Check current rules with counsel in every state where your patients live.
Are there FDA-approved AI therapy chatbots?
Not in generative form, as of the FDA's November 2025 advisory committee meeting. Reports from that meeting said the agency had not yet cleared any mental health tool using generative AI. Panelists recommended human escalation plans, crisis escalation, and adverse event reporting for any future product.
What are safe uses of AI in behavioral health?
The safest uses are administrative and supportive: eligibility verification, billing, scheduling, reminders, and alerts that flag patients for staff review. In each case a person makes the clinical decision. Uses become riskier as AI output reaches patients as clinical guidance or makes treatment decisions without review.
What should a BAA and QSOA cover for an AI vendor?
A BAA should cover how the vendor and every subcontractor, including AI model providers, protect health information and report breaches. A QSOA is needed when the vendor handles 42 CFR Part 2 SUD records. Also confirm in writing whether your data can be used to train models.
Does Verity use an AI chatbot or AI therapist?
No. Verity uses AI for automation, such as billing and eligibility workflows and alerts that route patient issues to the right staff member. Patient engagement comes from gamification, contingency management, expert-built exercises created by clinical psychologists, and validated assessments, with clinicians reviewing results and making every treatment decision.