What Clinicians Actually Do Inside a Clinically Supervised AI Therapy Workflow

Every AI mental health platform says it is “clinically supervised.” Few of them explain what that means in practice. What does the clinician actually do? When do they do it? What would they catch that the AI would miss? What happens if something goes wrong?

The phrase “clinically supervised” has been used so often as marketing shorthand that it has lost almost all descriptive content. Illinois passed legislation in 2025 explicitly banning AI mental health tools that imply clinical supervision without delivering it. More than 40 state bills were introduced in 2026 requiring licensed oversight of AI clinical tools. The regulatory convergence is not coincidental: the term had become meaningless as a consumer assurance, and policymakers are now requiring that it mean something specific.

This post describes what genuine clinical supervision of an AI mental health workflow looks like at each step of care: intake, prescribing, ongoing monitoring, and crisis response.

Why “Clinically Supervised” Needs Defining

The American Psychiatric Association’s 2024 position statement on AI in psychiatry describes AI as “augmented intelligence that supports, rather than replaces, clinician judgment.” That framing is useful but not operational. It does not tell a patient what their prescriber actually sees, reviews, or acts on inside the platform they are using.

The gap between the phrase and the practice is now a legal concern. Illinois’ Wellness and Oversight for Psychological Resources Act (WOPR Act, HB 1806, signed August 2025) specifies that licensed mental health providers may use AI tools only if they review and approve any therapeutic output, and that AI cannot independently make decisions or interact with clients therapeutically. Nevada and Utah passed similar legislation in 2025. Across 2026, over 40 state bills have introduced or proposed similar oversight requirements.

A 2024 JMIR Mental Health review of 793 state bills related to mental health AI found that 143 had substantial implications for clinical practice, with professional oversight requirements emerging as one of four dominant regulatory themes. The policy consensus is clear: human in the loop is not a slogan. It is a design requirement.

What the Prescriber Reviews Before Approving an Intake

In a clinically supervised AI model, the prescriber’s first task is reviewing the intake output before any treatment begins. At SiggyMD, this means reviewing a prescriber dashboard that includes:

• PHQ-9 results from the patient’s intake, providing a standardized severity measure.
• The AI-summarized patient history from the intake conversation, plus the full intake transcript. The prescriber can review what the patient actually said, not only the structured summary.
• Specific clinical considerations surfaced by the AI, flagging anything in the intake that warrants attention before prescribing.
• Three differential diagnoses, deliberately non-deterministic in line with FDA guidance on clinical AI that supports rather than replaces prescriber judgment.
• A treatment recommendation with a specific titration protocol. The prescriber reviews, adjusts, and approves.

Nothing moves forward without this review. A patient who completes an intake and pays for a plan still waits for prescriber approval before any medication is prescribed. That is not a delay in the model. It is the model.

What the Prescriber Approves (and What They Can Reject)

The prescriber at this stage is not rubber-stamping an AI output. They are making a clinical decision with structured AI support.

They can approve the AI’s recommended direction, modify it, replace it entirely, or flag the patient for additional clinical information before proceeding. If the intake raises concerns requiring a direct prescriber conversation before prescribing, that conversation happens before any prescription is issued.

The AI gathers structured clinical information and surfaces a proposed direction. The prescriber verifies the information is complete, assesses whether the direction is appropriate for this specific patient, and bears clinical and legal accountability for what is prescribed. Research suggests that when AI systems are clinically fine-tuned, risk-gated, and supervised, early evidence indicates they can improve patient engagement and symptom outcomes while maintaining therapeutic alliance. The supervision is not incidental to those outcomes. It is part of what produces them.

What the Prescriber Monitors Between Appointments

This is where most platforms that claim clinical supervision diverge most clearly from what genuine oversight looks like. The question is not only whether a prescriber reviewed the intake. It is whether they can see what is happening between visits.

In a supervised continuous monitoring model, the prescriber reviews longitudinal data from daily check-ins: mood trajectory, sleep quality, side effect timing and progression, adherence patterns, and functional indicators. This is not a weekly report. It is a clinical record that grows every day.

A prescriber reviewing two weeks of check-in data can see what the quarterly appointment cannot surface: sleep quality deteriorating before mood worsens, side effects peaking and resolving on schedule, adherence inconsistencies that explain a flat symptom response. The PHQ-9’s developers established that the instrument performs best as a longitudinal measure of treatment response when administered at regular intervals, with its sensitivity to change over time being one of its primary clinical applications. Daily check-in data makes this operationally possible at a scale that quarterly appointments do not.

What Triggers Prescriber Intervention

Genuine clinical supervision includes defined thresholds for intervention: specific patterns in the data that generate prescriber alerts and require a clinical response before the next scheduled visit.

These triggers are not limited to safety crises. They include: sleep quality dropping below a threshold for more than a defined number of consecutive days, side effect intensity increasing beyond expected ranges, symptom scores worsening after an initial period of improvement, adherence falling below a minimum consistency level. Each of these can occur without the patient flagging it as a problem. The clinical system surfaces the pattern before the patient recognizes it as significant.

This is the practical meaning of always-on care: not that the prescriber is reviewing every check-in manually in real time, but that the system monitors structured data continuously and surfaces patterns warranting clinical attention when they occur, so the prescriber can act at the right clinical moment rather than an arbitrary quarterly interval.

How Prescribers Handle Crisis Escalation

The most consequential test of clinical supervision is what happens when something goes wrong at 2am on a Tuesday.

In a genuine supervised model, an acute event—a patient expressing suicidal ideation, reporting a panic attack, or describing severe medication symptoms—triggers an escalation pathway that routes directly to a licensed prescriber with the full conversation transcript transferred. The prescriber does not receive a summary or an alert. They receive the actual clinical record of what happened.

After the acute event, the prescriber reviews the transcript to verify that the safety agent handled the interaction correctly and that the patient received appropriate de-escalation. The FDA’s Digital Health Advisory Committee has specified that AI mental health tools should have “clinically integrated escalation” with predefined human escalation plans connecting to qualified clinicians, not generic helplines.

If you are in immediate danger, call 911. Crisis escalation within a supervised platform supplements emergency services; it does not replace them.

Why This Differs From a General AI Chatbot

The distinction is not whether AI is involved. It is whether every clinical decision is reviewed, approved, and accountable to a licensed clinician who can see the full clinical picture.

A general AI chatbot generates responses based on user input, has no structured intake producing a clinical record, has no prescriber reviewing its outputs, has no defined escalation pathway to a qualified clinician, and in most states cannot legally claim to provide mental health treatment.

A clinically supervised AI care model collects structured clinical data, routes it to a licensed prescriber before any treatment begins, monitors ongoing data and surfaces clinical signals between visits, has defined thresholds that trigger prescriber intervention, and transfers full transcripts to prescribers during crisis events. The prescriber has accountability for every clinical decision.

The APA has been explicit that AI in mental health must function as augmented intelligence under clinician oversight, and that higher-risk prescribing decisions require human review. The regulatory convergence on this distinction reflects a clinical consensus that patients deserve to know which category of tool they are actually using.

How SiggyMD Implements Clinical Oversight

SiggyMD’s clinical oversight model is built around the specific prescriber responsibilities described above. The anonymous AI intake produces a structured dashboard the prescriber reviews before any treatment plan is approved. The prescriber reviews PHQ-9 results, the intake transcript, clinical considerations, three differential diagnoses, and a proposed treatment direction with titration protocol. Nothing is prescribed without this step.

Daily check-ins feed a longitudinal clinical record that the prescriber reviews continuously. Defined clinical thresholds trigger prescriber alerts. Crisis events route to in-house prescribers with full transcript transfer. The prescriber reviews the crisis response afterward to verify it was handled correctly.

“Clinically supervised does not mean the AI asks questions and a prescriber rubber-stamps the answer,” says Wendy Delgado, P.A., of the SiggyMD clinical team. “It means I have seen the full intake, I have approved the treatment direction, and I am watching what happens next. When the data shows something that needs attention, I am acting on it before the patient knows they should be concerned. That is what supervision actually looks like.”

What Members Are Saying

K.P., 35 — Anxiety, Depression: “I was nervous about how much a real doctor was actually involved. When I got my intake summary, I could see that a specific prescriber had reviewed my case and approved the plan before anything was prescribed. That changed how I thought about it. It was not just an algorithm deciding what I should take.”

A.W., 42 — Major Depressive Disorder: “My prescriber flagged a pattern in my check-in data before I mentioned anything to anyone. She reached out directly to ask about my sleep. I had not even connected that the change was significant. That interaction is why I trust the model.”

Member stories reflect real experiences. Names and identifying details have been changed to protect privacy. Results vary. SiggyMD is currently invite-only.

Bottom Line

“Clinically supervised” is not a marketing phrase. It is a design specification. The question patients should ask of any AI mental health platform is not whether a clinician is involved, but what that clinician can actually see, what they approve before treatment begins, and what triggers their involvement between visits.

The platforms that cannot answer those questions specifically are not clinically supervised in any meaningful sense. The ones that can explain the prescriber review process, the monitoring thresholds, and the crisis escalation pathway are operating in a different clinical category entirely. Regulators are now acting on that premise.

Sources

1. American Psychiatric Association. Artificial Intelligence in Psychiatric Care. APA. Accessed June 2026.
2. American Psychiatric Association. AI Prescribing: Considerations for Psychiatrists. APA. Accessed June 2026.
3. Blueprint AI. Breaking Down Current Legislation Regulating AI in Mental Health Care. Accessed June 2026.
4. Healio Psychiatry. The Hidden Dangers of AI Therapy Tools: What Clinicians Need to Know. September 2025.
5. JMIR Mental Health. Governing AI in Mental Health: 50-State Legislative Review. 2024.
6. Orrick. FDA’s Digital Health Advisory Committee Considers Generative AI Therapy Chatbots for Depression. November 2025.
7. Manatt Health. Health AI Policy Tracker. Accessed June 2026.
8. Kroenke K, Spitzer RL, Williams JBW. The PHQ-9: Validity of a Brief Depression Severity Measure. Journal of General Internal Medicine. 2001;16(9):606-613.