The conversation about AI mental health chatbots tends to run in two registers. One is enthusiasm: “24-hour access to evidence-based CBT at the cost of an app.” The other is skepticism: “an algorithm cannot replace a therapist.” Both contain elements of truth. Neither does justice to what the peer-reviewed literature actually shows.
I’ve evaluated a lot of AI systems over the years, and the honest answer is almost always in between the marketing and the dismissal. The chatbot literature is a good example.
A 2024 systematic review of AI-powered CBT chatbots, published in PMC, identified ten studies - on Woebot, Wysa, and Youper - across depression, anxiety, chronic pain, and maternal mental health. The review reported large improvements in mental health symptoms across the chatbots reviewed, with high user engagement.
A randomized controlled trial of Woebot in young adults with subclinical depression and anxiety found Woebot more effective at reducing depressive symptoms over a two-week period than self-help materials prepared by the World Health Organization. A 2025 JMIR Mental Health narrative review of CBT-based chatbots concluded that the field has accumulated meaningful evidence for clinically relevant effects on depression and anxiety in specific populations.
Wysa-specific research has documented PHQ-9 reductions of more than 5 points among users who engaged with the chatbot at least twice per week over two weeks - a clinically meaningful magnitude - though with fewer randomized controlled trials than the Woebot literature.
The same systematic reviews are explicit about the limitations.
The clinically defensible posture is neither blanket recommendation nor blanket avoidance.
Chatbots may be appropriate as adjuncts - supervised, structured, and integrated into the treatment plan - for some patients with mild-to-moderate symptoms, particularly where between-session structured CBT homework is already part of the treatment. The clinician retains responsibility for case formulation, modality selection, alliance, and safety.
Chatbots are not appropriate as substitutes for clinician care for moderate-to-severe presentations, for trauma work, for personality-disorder work, for crisis-presenting patients, or for patients whose treatment depends on the alliance the chatbot literature acknowledges it cannot fully replicate.
The clinician’s role in either case is to select carefully, integrate explicitly, and monitor engagement and outcomes with the same measurement-based-care discipline used for the broader treatment plan.
For a practice, the chatbot question is rarely “should we recommend this app.” It is “what role does between-session structured engagement play in our clinical model?”
A clinician-supervised, between-session engagement workflow - structured patient-reported outcomes, CBT homework adherence tracking, mood and sleep monitoring, and clinician review of the data - produces a coherent clinical and reimbursement structure. Where coverage exists, RTM (CPT 98975–98981) reimburses the structured between-session review. The patient receives more care. The clinician is paid for reviewing the data. The alliance, the work, and the economics align.
This is a different model from “recommend an app.” It is a model where the practice integrates evidence-based digital tools into a clinician-supervised treatment plan, retains clinical responsibility, and operates within a measurement-based-care discipline that supports clinical, audit, and value-based-payment objectives simultaneously.
The peer-reviewed evidence on AI chatbots is real, limited, and population-specific. The clinically defensible use is supervised, integrated, and bounded by careful patient and presentation selection. Practices that take this seriously can use these tools as legitimate adjuncts without abdicating clinical responsibility. Practices that treat them as either silver bullets or threats miss the careful place in between - where the actual clinical and economic value lives.
#DigitalTherapeutics #MentalHealthAI #CBT #BehavioralHealthTech #MeasurementBasedCare #RTM #ReliefAI