Employers are beginning to evaluate AI mental health tools that can mimic empathy, remember details and remain available around the clock. Current evidence does not show that those systems can replace the clinical judgment, professional accountability or contextual understanding of licensed therapists.
That leaves HR and IT teams with a high-stakes procurement decision. Workplace wellness tools may collect sensitive health and employment disclosures, offer individualized guidance or interact with employees in severe distress, so any deployment requires product-specific evidence, strict data controls and a reliable path to human care.
What controlled studies show — and what they do not
Purpose-built systems have produced promising results under controlled conditions. Therabot, developed by Dartmouth researchers, was tested in a randomized trial involving 210 US adults with clinically significant symptoms of major depression, generalized anxiety or an eating disorder.
Among the 106 participants assigned to Therabot, symptoms fell over four weeks compared with a 104-person waitlist control group. Participants also reported a therapeutic alliance comparable to benchmarks from outpatient psychotherapy, but the Therabot clinical trial did not compare the system directly with licensed therapists, and researchers monitored conversations for safety.
The findings support more study, not autonomous deployment. Dartmouth researchers said generative AI still required clinician oversight, reflecting the broader need to restore human oversight after AI quality failures.
General-purpose chatbots present greater uncertainty because they are not designed or clinically validated for a defined treatment. A Stanford evaluation of therapy chatbots found greater stigma toward schizophrenia and alcohol dependence than depression, while some systems missed indirect suicidal intent or reinforced delusional thinking.
Research on LLM counselors also identified failures involving crisis management, contextual adaptation, discrimination and simulated empathy. American Psychological Association guidance advises against using generative chatbots and wellness apps as substitutes for qualified professionals or as crisis resources.
How employers should govern mental health AI
Buyers should demand evidence for the exact product and model version under review. Results from one chatbot, population or controlled trial do not validate every product marketed as an AI therapist, so the review should form part of an AI governance and risk assessment process rather than sit solely with benefits or procurement teams.
Vendors should document how their systems identify crisis language, disclose limitations and transfer users to qualified human support. Those safeguards should be retested after significant model, prompt or policy changes, and contracts should identify who remains accountable when escalation fails.
Data handling deserves equal scrutiny because mental health conversations may contain health, disability and employment information. Contracts should specify storage location, access, retention, deletion and whether conversations are used for model training; unmanaged consumer tools can turn routine inputs into a corporate data governance issue.
Employers should avoid collecting identifiable transcripts without a defined need. Multinational organizations should also require evidence across the languages, age groups and cultural settings represented in their workforce.
AI may support psychoeducation, journaling, routine reflection and administrative work, but it should not be presented as a diagnostic service, crisis resource or replacement for licensed care. Organizations should not deploy it without product-specific evidence, enforceable privacy controls and a qualified human accountable for failures.
Read more: The risks of replacing specialist employees too quickly extend beyond health care, with Gartner expecting some companies that cut customer service roles for AI to reverse those staffing decisions by 2027.