Therapists want AI to take over their paperwork. They do not want it doing their therapy. In practice surveys, documentation and charting rank as the top driver of clinician burnout, ahead of caseload and pay. Studies of AI scribe tools, which draft the clinical note for a clinician to edit and sign, show measurable drops in documentation time, after-hours charting, and burnout (JAMA Network Open, 2025). The pitch clinicians consistently reject is the opposite one: AI that acts as the therapist with no licensed person in the loop. This piece lays out what therapists actually want from AI, what the research supports, and the four questions a tool has to answer before it belongs in a practice.
I am a licensed clinical social worker. I also build software for clinicians. So I spend a lot of time in both conversations, and the gap between them is the most important thing happening in this field right now. The pitch is “AI will see the patient.” The demand is “AI, please do my notes, fight my insurance denials, and otherwise stay out of the room.” A tool that gets that backwards does not just miss the market. It earns the distrust clinicians are already carrying.
What clinicians are actually asking for
The number one request is documentation relief, and it is not close. When Tebra surveyed mental-health professionals about what drives their burnout, documentation and charting came in first.[1] Solo practitioners report spending somewhere between five and ten hours a week on documentation alone — before scheduling, billing, or insurance correspondence enters the picture — and a majority say they finish that documentation outside scheduled work hours, on their own time, at night.[2] There is a line that circulates among clinicians, in various forms, that captures it exactly: nobody became a therapist to fight insurance denials at midnight.
This is where the evidence on AI-as-tool is genuinely encouraging, and worth being honest about. Randomized and observational studies of ambient AI scribes — tools that, with consent, draft the clinical note so the clinician edits instead of composing from scratch — have reported real reductions in documentation time, after-hours charting, and self-reported burnout.[3] One multi-system study of more than 1,400 clinicians found lower cognitive burden and less after-hours documentation after adoption; a separate pilot found clinicians spent meaningfully less time composing notes than matched controls.[3][4] The effect sizes vary a lot by how well the tool is implemented — which is the whole game — but the direction is consistent. Admin relief is the use case the research supports and the use case clinicians are begging for. The two things finally agree.
So the first thing therapists want from AI is simple: be the secretary, not the clinician. Eat the intake summary, the note scaffold, the measure scoring, the scheduling friction, the parts of the day that have nothing to do with why anyone went into this work.
What clinicians do not want, stated plainly
The hard line is just as clear, and a tool that crosses it loses the room instantly. Here is what the people who would actually use these tools keep saying.
AI should augment, never replace. The resistance to AI-as-primary-therapist among licensed providers is close to consensus. The argument is not nostalgia. It is that change in therapy happens inside a relationship between two human nervous systems, inside an accountable clinical frame, over time — and a text generator does not have a relationship with anyone. I have written separately about why the alliance literature pushes back on the replacement framing, so I will not re-run it here. The short version: the relationship does the work, the clinician carries the responsibility, and the tools that earn their keep make the clinician more effective without pretending to be the clinician. And the evidence base for the replacement pitch is thinner than the marketing implies — a 2025 systematic review of generative-AI mental-health chatbots found only a small minority of studies had undergone genuine clinical-efficacy testing at all.[9]
The validation trap is a clinical problem, not a UX quirk. General-purpose chatbots are tuned to be agreeable. They affirm. They smooth. They tell you what you want to hear, because that is what keeps the conversation pleasant. Clinically, that is a serious problem, and clinicians say so constantly: a lot of the work is challenging a client, holding a boundary, naming the avoidance the client has organized their life around not seeing. An endlessly agreeable model does the opposite. It can collude with the exact defenses the therapy is built to interrupt. This is now documented, not just clinical intuition: a 2024 study found that five leading AI assistants systematically matched users’ stated beliefs over accurate answers — the structural property researchers call sycophancy.[6] In 2025, licensed psychologists reviewing real-world chatbot responses catalogued systematic violations of mental-health ethics standards, over-validation of the user’s beliefs among them,[7] and the American Psychological Association issued a formal health advisory warning that consumer chatbots are built to affirm rather than to challenge.[8] That is not a knock on the technology. It is a statement about who should be setting the clinical frame — and it is never the model.
Patient trust is fragile, and it is the whole asset. The stories that travel fastest in clinician circles are the trust-rupture stories: a client who finds out their therapist ran session content through ChatGPT and immediately stops trusting the work; the iPad propped up to record that changes the entire feel of the room; the description of a tool as a “glorified ChatGPT wrapper” that the clinician now has to defend. Once a client suspects a machine is doing the part they came to a human for, the alliance takes a hit that is expensive to repair. Clinicians know this in their bones, which is why they are cautious in a way the hype cycle reads as resistance and is actually professional judgment.
The data cannot quietly leave the building. The privacy fear is specific and correct: who can access the recording, the transcript, the most vulnerable forty-five minutes of a person’s life? Even with “review and edit” rights, if the raw material has already been shipped to a third-party model the clinician cannot see, the horse is out. This is the fear that should be doing the most work in any clinician’s evaluation, and it is the one most vendors answer with a logo wall instead of an architecture.
The four conditions a tool has to meet
When another clinician asks me how to evaluate an AI-assisted tool, I give them four questions, in this order. They are the difference between a tool that earns trust and one that creates work and risk you cannot see yet.
One: is the infrastructure under a Business Associate Agreement appropriate for PHI? Protected health information cannot transit infrastructure that is not under a BAA. That is regulation, not preference.[5] Most consumer chat APIs are not covered for PHI. If a tool cannot answer this question cleanly, the conversation is over before it starts.
Two: is the AI scoped to specific tasks, or pitched as a substitute for judgment? “Drafts your note for review” is a scoped task. “Assesses the client” is a substitution. The first is a tool. The second is a liability wearing a tool’s clothes.
Three: is the clinician the editor and signer of every clinical artifact? The model produces a starting point. The licensed clinician reviews it, corrects it, and owns it. The day the clinician becomes a rubber stamp on machine output is the day the de-skilling worry the field keeps voicing — if I stop synthesizing my own sessions, do I get worse at it? — becomes real. Governed, not autonomous. The arms can move; the head decides; the clinician signs.
Four: does the client’s data stay under the clinician’s control? This is the one most tools fail. The right answer is not “trust our security page.” The right answer is an architecture where the practice’s own material stays inside infrastructure the practice controls, and the only data that ever touches our protected infrastructure is the between-session clinical material the tool exists to handle — held under an appropriate BAA, isolated from everything public, never resold, never used to train a model. Two doors, and you should be able to see exactly what goes through each one.
Why I'm building for the admin work, not the session
This is the part where most software companies tell you their AI is the breakthrough clinician. I am telling you the opposite, on purpose, because it is what the work actually needs.
The independent clinician today is stuck between two bad options. On one side is the worksheet add-on — a thin tool that does one small thing and leaves the other ninety percent of the admin untouched. On the other side is the corporate platform: the VC-backed marketplace or all-in-one suite that runs the back office but, in exchange, tends to own the client relationship, the data, and increasingly the economics of the practice itself. Neither of those is built for the clinician who wants to keep their own practice, their own clients, and their own clinical voice.
That gap — a practice platform that does the admin so the clinician can do the therapy, without taking the practice hostage to do it — is the thing I think the next several years of credible clinical software is about. It is the ground VibeCheck is built on. The between-session layer that keeps the work alive in the 167 hours a week the client is not in the room. The note scaffolding, the measure scoring, the intake synthesis the clinician edits and signs. The AI assist that lives inside the clinician’s workflow and answers to the clinician, not the other way around. PHI held under an appropriate BAA on infrastructure isolated from this public site; the practice’s own back-office material kept under the practice’s own control. The clinician is the editor of everything clinical. The crisis path routes to 988 and a clinician, never to a model output.
That is not the exciting pitch. It is the honest one, and it is the one therapists are actually asking for. The dominant mood among clinicians right now is not “this is amazing.” It is “this could finally take the paperwork off my hands — just do not screw up the part that is actually the therapy.” That sentence is the entire product spec. Everything I am building answers to it.