There’s a habit in technology commentary of treating the field’s relationship with AI as if it were a single relationship. It isn’t. Spend enough time listening to clinicians and three distinct postures emerge, often from the same clinician in the same week.
The first is structured skepticism. “I’m less worried about AI replacing humans and more concerned about a client’s private information being saved by a company I know nothing about.” This is the dominant register, and it’s well-grounded. It asks, before “does it work?”, the more specific questions: who keeps the data, for how long, with what protection, with what BAA, and used to train what. As someone who’s spent years building intelligent systems, I find this posture encouraging - it’s how engineers should think before deploying, too.
The second is pragmatic adoption. “Where I find AI genuinely useful is in reducing mental overhead around documentation and information organization. That’s significant for clinicians already running at capacity.” This posture recognizes AI as a tool - a meaningful one for documentation, organization, structured templates, and between-session engagement - when the implementation is clean. It’s not enthusiastic; it’s functional. And functional is exactly right.
The third is existential uncertainty. “Will AI replace therapists? Probably not for complex work. But will it reshape how much human care the system pays for in mild-to-moderate cases? Almost certainly.” This posture takes the longer view - five, ten years out - and asks what happens if a meaningful share of routine care migrates to lower-cost AI-mediated channels. It’s the question a strategist has to sit with.
All three postures are rational. None is the whole picture.
Why this matters for practice decisions
A practice owner has to make technology decisions now - what to adopt, what to avoid, what to verify, what to wait on - under conditions where the underlying technology is moving faster than most ethics guidance can update. BAA structures, FDA device classification, questions about whether session audio can be used to train a language model - these are live regulatory conversations with real clinical stakes.
The structured response is not to pick a posture and commit to it. It’s to triage decisions according to which posture best applies.
Triage
Some questions are skepticism questions. Is this vendor’s BAA enforceable? Where is the data stored, for how long, with what encryption? Is patient audio used to train the vendor’s models? What does “HIPAA compliant” actually mean in this contract? What happens to the data if the vendor is acquired? These belong to the skeptical posture, and the answers should be in writing before any patient encounter is recorded.
Some questions are pragmatic questions. Does this tool measurably reduce documentation time, with adequate clinician review and edit? Does it integrate with the EHR rather than fragmenting the workflow? Does the patient understand what is being recorded and why, and are they consenting freely? These belong to the pragmatic posture, and the answers depend on the specific tool, the specific implementation, and the specific practice.
Some questions are long-view questions. How will the practice’s clinical and economic identity hold up if mild-to-moderate care migrates to AI-mediated channels? What is the practice doing now to deepen the work that is least replaceable - the alliance, the complex case, the trauma work, the modality integration? These belong to the existential posture, and the answers are strategic, not operational.
Where the integrated stack fits
The technology conversation in behavioral health is sometimes framed as a binary: adopt AI or don’t. That’s the wrong frame. The technology questions are a portfolio, and the practice’s answers should differ by category.
Measurement-based care infrastructure, structured patient-reported outcomes, between-session engagement, RTM-supported reimbursement of structured between-session review - these are unambiguously beneficial when implemented with clinician oversight, BAA-covered data flow, and patient consent. The evidence base supports the value.
AI-assisted documentation is more conditional. Implementation research on ambient scribe technology shows measurable reductions in note-writing time and clinician burden when the workflow is designed around clinician review rather than blind acceptance - and when BAA, data-retention, and training-data terms have been explicitly negotiated.
Direct AI-to-patient therapeutic engagement (chatbots, autonomous CBT tools, free-text therapy agents) has a more limited and more careful evidence base. Specific Randomized Controlled Trials support specific tools for specific symptom severity ranges, with significant limitations the field is still working through. More on this in the upcoming Article 6.
The portfolio doesn’t have the same shape for every practice. But naming the categories explicitly (and triaging questions by the posture that best applies) is what separates deliberate technology decisions from decisions made by drift.
The honest version of the argument
Clinicians are not confused about technology. They’re holding three legitimate postures simultaneously because the underlying reality demands all three. The articles that follow get specific: what HIPAA and BAAs actually require, what AI documentation does and doesn’t do, what the chatbot literature actually says, what teletherapy and the alliance look like together in 2026, what cross-state licensing risk really is, and what an integrated, defensible technology stack actually looks like in practice.
More at reliefai.co
Sources & References
• U.S. Department of Health & Human Services. HIPAA Privacy and Security Rules.
• HHS Office for Civil Rights. Business Associate Contracts.
• Federal Trade Commission. Health Breach Notification Rule.
• Office of the National Coordinator for Health IT (ONC).
• American Psychological Association. Guidelines for the Practice of Telepsychology.
#BehavioralHealthTech #AIInHealthcare #ClinicalEthics #PracticeManagement #DigitalHealth #ReliefAI
