Key Takeaways:

• Your claims determine whether FDA treats your app as a device. Under FDA’s January 2026 general wellness guidance, an app that helps users manage stress can stay outside device regulation, while one that claims to treat an anxiety disorder can’t.
• FDA hasn’t authorized a generative AI mental health device, and it’s writing the rules now. An advisory committee met on the topic in November 2025, FDA released a discussion paper on generative AI devices in August 2026, and draft guidance specific to mental health chatbots is on its fiscal 2027 agenda.
• The device route means controlled clinical data, and clearance alone doesn’t guarantee payment. Payers want evidence comparable to drugs, and Medicare pays only for devices in specific FDA classifications.
• States now regulate AI therapy directly. Seven states restrict it, and their rules differ, so plan a state-by-state strategy alongside the FDA one.

If you’re a digital health founder building an AI therapy app, one decision shapes your timeline, budget, and funding story: do you position the product as a wellness tool, or do you build the clinical evidence to take it through FDA as a medical device? Make that call before you write a line of production code, because your claims decide which side of the line you land on, whatever the technology underneath.

The ground is moving under that decision. FDA is actively writing its approach to generative AI, seven states now restrict AI-delivered therapy, and the companies that went furthest down the regulated path have taught some expensive lessons. This guide covers where the regulatory line sits, what FDA is signaling about generative AI, which pathways fit, how to design clinical validation, what payers and states require, and how to choose a route.

Does Your AI Therapy App Need FDA Clearance?

It does if you claim to diagnose, treat, or mitigate a mental health condition. It doesn’t if you stay within general wellness claims for a low-risk product.

FDA’s general wellness guidance, revised on January 6, 2026, keeps a two-factor test: the product is intended only for general wellness use, and it presents a low risk to users. One category of wellness claims never mentions a disease, covering uses such as relaxation or stress management, mental acuity, self-esteem, and sleep management. Another allows disease-related language only in a narrow form. FDA’s own example is software that tracks sleep, work, and exercise routines which, as part of a healthy lifestyle, “may help living well with anxiety.”

The line is the intended use you communicate. An app that coaches breathing and stress management sits comfortably in wellness territory. A claim that a product helps treat an anxiety disorder puts it on the device side, no matter how simple the software is.

What Enforcement Discretion Actually Means

Enforcement discretion isn’t an approval or a permanent exemption. It means FDA has said it doesn’t intend to enforce device requirements for a defined class of low-risk software. Your claims, onboarding copy, app store listing, and sales deck all count as evidence of intended use, so a wellness product can drift into device territory one marketing sentence at a time.

A Practical Self-Check

• List every sentence about outcomes in your app, store listing, website, and investor deck.
• Flag any that name a diagnosis (depression, generalized anxiety disorder, PTSD) or promise symptom change.
• Flag any feature that screens for, detects, or responds to a clinical condition with specific guidance.
• If anything is flagged, assume device regulation applies and plan accordingly.

Where Does FDA Stand on Generative AI Therapy?

FDA hasn’t authorized a generative AI mental health device, and it’s actively working out its approach. That’s the most important planning fact for an AI therapy founder.

The November 2025 advisory committee. FDA’s Digital Health Advisory Committee met on November 6, 2025 to discuss generative AI-enabled digital mental health devices. According to FDA’s meeting summary, the agency told the committee it had authorized more than 1,200 AI-enabled devices and that none yet involved generative AI for mental health conditions. The discussion worked through a hypothetical prescription chatbot for major depressive disorder and emphasized human oversight and monitoring across the product’s life cycle.

The August 2026 discussion paper. On August 18, 2026, FDA released a discussion paper on regulating generative AI-enabled medical devices. It isn’t guidance, but it previews FDA’s thinking, including a two-axis view of risk and competency-based evaluation modeled on how human clinicians are trained and tested. The comment period closed October 19, 2026.

Guidance on the way. FDA’s fiscal 2027 guidance agenda includes a planned draft on evidence for generative AI conversational devices for mental disorders. When it arrives, it’ll be the closest thing to a rulebook for this category.

FDA has already shown it can clear a patient-facing language model. In December 2025, it cleared UpDoc, prescription insulin management software for type 2 diabetes, in a design that analysts describe as a cage around the model, with a dose calculator at its core. That’s diabetes, not mental health, but it suggests the shape reviewers are comfortable with: a narrow task, bounded outputs, and the model kept away from unchecked clinical decisions.

For you, the timing cuts both ways. There’s no finished rulebook yet, so expect FDA to apply existing device frameworks and probe output variability hard. Being early is risky and valuable at once: a clear regulatory story sets you apart, and you’ll help write the precedent.

Which FDA Pathway Fits an AI Therapy App?

For software that delivers behavioral therapy for a psychiatric disorder, the established route is Class II with special controls, reached through 510(k) if you have a predicate or De Novo if you don’t.

The existing classification is 21 CFR 882.5801, computerized behavioral therapy devices for psychiatric disorders, a Class II category whose special controls cover clinical performance data, software verification and validation, hazard analysis, and labeling. The risk those controls target is ineffective treatment that lets a condition worsen.

Precedent shapes which door you use. EndeavorRx, a prescription digital therapeutic for ADHD, went through De Novo in 2020, creating a classification that later products could follow. A 2025 analysis of FDA labeling found 13 software-only prescription digital therapeutics cleared as of May 2025, eight through 510(k) and five through De Novo. Rejoyn, a prescription app for major depressive disorder symptoms, used the 510(k) route in 2024 by pointing to an earlier behavioral therapy predicate.

The Catch for Generative AI

Every cleared product above follows a fixed therapy protocol. A generative model produces different words each time, so demonstrating that it reliably delivers a validated therapy is harder, and its technological differences may rule out a clean 510(k) predicate. Plan for a De Novo-style conversation with FDA, and use the Q-Submission process early to test your intended use and study design before you invest in the trial that will support your submission. Dogtown’s guide to software as a medical device covers the classification questions.

The pathway also decides your reimbursement options. As the payer section below explains, Medicare pays only for devices in specific classifications, so a product that creates a new classification through De Novo may sit outside existing payment codes until CMS expands them.

How Should You Design Clinical Validation for an AI Therapy App?

Design the study for two audiences at once. FDA wants proof the software does what you claim, safely, and payers want evidence they could compare with a drug.

The bar has been uneven in the past. A Health Affairs analysis of 20 FDA-authorized prescription digital therapeutics found that 40% had no clinical studies referenced in FDA’s decision summaries, and only two were supported by studies that met every marker of high quality, including randomization, blinding, and multiple sites. Design above that floor, because payers and clinicians read the evidence too.

What the Best Generative AI Trial Shows, and What It Leaves Open

The most cited study to date is Dartmouth’s Therabot trial, published in NEJM AI in March 2025. It randomized 210 adults with clinically significant symptoms of depression or generalized anxiety, or at high risk for eating disorders, to four weeks of Therabot or a waitlist control. Therabot users improved more than the waitlist group in all three symptom areas, and Dartmouth reports average reductions of 51% in depressive symptoms, 31% in anxiety, and 19% in eating-disorder concerns.

It’s a serious study, and its design shows where reviewers will push. The control was a waitlist, not an active comparator, and the authors themselves call for an active control in future research. Participants were recruited online and may skew toward people comfortable with AI. And people with active suicidality, mania, or psychosis were excluded, so the trial says little about how a generative chatbot handles the users at greatest risk.

A Design Checklist

• Control arm. Use an attention-matched or active comparator, not only a waitlist.
• Independent assessment. Have outcomes assessed by people outside the development team.
• Validated measures. Pick instruments built for the construct you’re measuring, and interpret therapeutic alliance scores carefully, since those scales were designed for human relationships.
• Safety endpoints. Prespecify how you detect and handle crisis language, how you escalate to humans, and how you report adverse events.
• Population. State exactly who the product is for. Adolescents raise distinct concerns, as the next item shows.
• Adversarial testing. In a JMIR Mental Health pilot, 10 therapy and companion chatbots endorsed fictional teenagers’ clearly harmful proposals in 19 of 60 opportunities (32%), and none rejected them all. It’s a small study, but it shows the failure mode reviewers will probe.
• Model change control. Freeze the model version, prompts, and safety layers for the trial, log every output, and document any change, since a generative product that changes mid-study weakens its own evidence.

If you’re running the study with a research partner, the app has to be trial-ready from day one. Dogtown built the apps for Moffitt Cancer Center’s smoking cessation interventions, which were deployed in two NCI-funded randomized clinical trials, and our clinical trial app development work covers what that build requires.

What Evidence Will Payers Want?

Payers want effectiveness evidence that looks like what they see for pharmaceuticals, and FDA clearance is the entry ticket. In focus groups with 21 U.S. payers, participants said they wanted evidence of effectiveness similar to traditional drugs, viewed mental and behavioral health products as the most likely to be reimbursed, and named the exclusion of nonprescription products as a coverage barrier. Several said they need evidence beyond what FDA requires.

The commercial history is a warning. Pear Therapeutics held three FDA authorizations and still reported $12.7 million in revenue against $136 million in expenses before its 2023 bankruptcy. Woebot, one of the most studied therapy chatbots, shut down its app in June 2025, and its founder told STAT that the decision was largely attributable to the cost and challenge of meeting FDA’s requirements for marketing authorization. Kintsugi, a voice-biomarker company, spent four years and $16 million on regulatory work and shut down before filing its De Novo request, according to Innolitics.

Medicare’s path is clearer than it was, and narrow. CMS created payment for digital mental health treatment devices in 2025 for products classified under 21 CFR 882.5801, and for 2026 it expanded HCPCS codes G0552 through G0554 to also cover ADHD devices classified under 21 CFR 882.5803. The billing practitioner furnishes the device as part of a behavioral health treatment plan, and the device-supply code remains priced by Medicare’s regional contractors.

The practical takeaways: a wellness app can’t use these codes, a cleared app can bill only within its indication, payment flows through clinicians, and a device in a new classification needs CMS to act before these codes apply. Write your indication for use with your reimbursement path in mind.

Which States Restrict AI Therapy?

State law is now as important as FDA policy for an AI therapy product, and the states don’t agree.

• Restrictions on AI-delivered therapy. Illinois and Nevada acted in 2025, and in 2026 Colorado, Maine, Rhode Island, Tennessee, and Vermont followed.
• Different mechanisms. According to a Legal500 analysis, Nevada bars presenting AI as able to provide psychotherapy and making AI mental health chatbots available. Colorado bars presenting AI as a therapist and will stop health plans from paying for AI psychotherapy. Tennessee bars presenting AI as a therapist. Illinois, Maine, Rhode Island, and Vermont require psychotherapy to come from a qualified professional.
• Permission with conditions. Utah expressly permits AI mental health chatbots, with disclosure and data protection requirements.
• California. Its companion chatbot law, SB 243, took effect in January 2026, and a further law on minors’ use of companion chatbots was signed in September 2026. Governor Newsom vetoed SB 903, which would have prohibited offering psychotherapy through AI chatbots, on September 30, 2026.

These laws turn on how you present the product and who delivers care. A clinician-in-the-loop design, in which licensed professionals direct treatment and the AI supports them, fits most of these frameworks better than a standalone chatbot marketed as a therapist. Map your launch states, review how each defines therapy, and be ready to gate features or positioning by state. FDA authorization doesn’t by itself settle these state questions, so get legal review in each market.

Wellness Route or Device Route: How Do You Choose?

Choose the wellness route if your business model is consumer subscription or employer wellness and you can live with nonclinical claims. Choose the device route if you need clinician prescribing, health system adoption, or Medicare billing.

Factor	Wellness route	Device route (Class II)
Claims allowed	General wellness, limited disease-related phrasing	Condition-specific treatment claims
Premarket evidence	None required by FDA	Controlled clinical data plus software validation
Reimbursement	Not eligible for Medicare’s digital mental health treatment codes	Possible within the cleared indication and covered classifications
State law exposure	Restrictions on presenting AI as therapy still apply	Still applies, and clinician-directed designs fit best
Main risk	Claims drift pushes you into device territory	No cleared generative AI precedent yet, and evidence costs

Capital is available, but concentrated. Digital health startups raised $14.2 billion in 2025, and AI-enabled companies took 54% of it. Rock Health noted that removing the top nine companies would pull 2025’s total below 2024’s, and by 2026 it had stopped labeling AI as a distinguishing category because it had become so common. With AI now table stakes, a credible regulatory and evidence plan is what sets a company apart.

A Hybrid Approach Many Teams Should Consider

You don’t have to choose once and forever. A common sequence is to launch a wellness product with strictly nonclinical claims, collect real-world usage and safety data, and use it to scope a clinical study for a separate, device-labeled indication. Keep the two product lines clearly separated in claims, labeling, and architecture so one doesn’t contaminate the other.

The wellness route still carries obligations. A consumer health app outside HIPAA falls under the FTC’s Health Breach Notification Rule, which treats unauthorized sharing of health data, such as with advertising partners, as a breach. Dogtown’s guide to mental health app development covers the privacy and safety features either route needs.

An AI Therapy App Readiness Checklist

• Intended use written. One indication statement drives your claims, labeling, and marketing.
• Claims audited. Every outcome statement across app, store listing, website, and deck has been checked against the wellness guidance.
• Route chosen. Wellness, device, or a separated hybrid, with the reasons documented.
• FDA engaged early. A Q-Submission has tested your intended use and study design before the main study enrolls.
• Study designed above the floor. Active or attention-matched control, independent assessment, validated measures, and prespecified safety endpoints.
• Crisis handling built. Detection, escalation to humans, and adverse event reporting work before the first user sees the product.
• Adversarial testing done. The model has been tested against harmful requests, including those typical of minors.
• Model changes controlled. Versions, prompts, and safety layers are frozen and logged for the study.
• State map complete. Each launch state’s rules on AI therapy and companion chatbots are reviewed, with feature or positioning gates where needed.
• Reimbursement path mapped. Your classification and indication line up with the payment codes or commercial coverage you plan to use.

Frequently Asked Questions

Is an AI therapy app a medical device?

It is if it’s intended to diagnose, treat, or mitigate a mental health condition. Apps limited to general wellness claims, for products that pose low risk, fall under FDA’s wellness policy, which says FDA doesn’t intend to regulate them as devices. Your labeling and marketing define intended use, so a claim like “helps treat an anxiety disorder” makes it a device, while “may help living well with anxiety” as part of a healthy lifestyle can stay in wellness territory.

Has FDA cleared a generative AI therapy chatbot?

No. FDA told its advisory committee in November 2025 that none of its authorized AI devices involved generative AI for mental health conditions. It has cleared a patient-facing language model in another area, insulin management for type 2 diabetes, and it’s developing its approach through an August 2026 discussion paper and planned draft guidance on generative AI conversational devices for mental disorders. Check FDA’s device databases for updates before you finalize your plan.

What’s the difference between FDA-cleared, De Novo, and FDA-approved for a therapy app?

FDA-cleared products come through 510(k) by showing substantial equivalence to a predicate. De Novo classifies a novel low- to moderate-risk device that has no predicate, as it did for EndeavorRx. FDA-approved refers to premarket approval, the pathway for the highest-risk devices. Most digital mental health therapeutics to date are cleared or De Novo-authorized, not approved.

Do I need a randomized controlled trial?

For a Class II behavioral therapy device, plan on clinical data from a well-controlled, prospective study that supports your indication. A randomized design with an active or attention-matched control is the strongest position with both FDA and payers. Confirm your study design with FDA through a Q-Submission before you enroll patients.

Will Medicare pay for my AI therapy app?

Only if it’s an FDA-cleared or De Novo-authorized device in a classification CMS covers, used within its indication as part of a clinician’s behavioral health treatment plan. For 2026, that means computerized behavioral therapy devices for psychiatric disorders and digital therapy devices for ADHD. Wellness apps can’t use these codes, and commercial payers set their own evidence bars, generally asking for evidence comparable to drugs.

The Bottom Line

The founders who handle this well treat regulation as a product decision made at the start. They write the intended use first, pick the evidence they can afford to generate, map the states where they’ll operate, and build the claims, architecture, and study design to match. If you’re weighing a device route, see how Dogtown approaches regulated builds on our digital therapeutics app development page, and if it helps to talk through your intended use and evidence plan, request a free consultation.