It is Tuesday, 9:54 PM, and Maria, an LCSW in Oakland who saw eight clients today, has seven half-finished progress notes, an unimported PHQ-9, and a 72-hour documentation window closing on a 90837 that an auditor will one day read line by line. She is about to spend two unpaid hours doing work an AI scribe could draft in minutes, and she is also one bad decision away from letting that same AI do work it must never touch. This lesson gives you the honest inventory: the seven categories of behavioral health work where generative AI reliably earns its subscription fee, and the clinician-only sub-tasks buried inside each one, the diagnosis, the risk level, the medication recommendation, the custody opinion, the mandated-report call, that no model gets to make. By the end you will be able to sort your own week into "delegate the drafting" and "never delegate the deciding," and you will build a one-page Delegation Map you can tape next to your monitor.

The Sous-Chef, Not the Chef

Hold one picture in your head for this entire lesson: a generative AI tool in your practice is a sous-chef. A good sous-chef is fast, tireless, and remarkably consistent at prep work. It chops, it portions, it plates, it keeps the station clean, and it does all of that at 9:54 PM without complaint. What a sous-chef never does is taste the dish and decide it is done, change the recipe for a guest with an allergy, or sign the health inspection form. The chef does that, because the chef carries the responsibility, the training, and the license. In your kitchen, you are the chef. The model preps; you taste, you decide, you sign.

This distinction is not a metaphor for politeness. It is the structural line that runs through every regulation now landing on this field. The Illinois WOPR Act (Wellness and Oversight for Psychological Resources Act, 2025) prohibits AI from providing therapy. Nevada AB 406 prohibits AI-delivered behavioral healthcare. New York's AI companion law regulates the chatbot side of the market precisely because lawmakers watched products drift from "wellness app" toward "therapist substitute." Every one of those statutes draws the same line the sous-chef analogy draws: AI can support the work of a licensed clinician; it cannot be the clinician. When you sort tasks this lesson's way, you are not just being efficient. You are staying on the legal side of a line that three states have already drawn in statute and others are drafting.

The practical question, then, is not "should I use AI?" Nearly one in three psychologists already use AI at least monthly, per the APA Practitioner Pulse Survey's 2026 wave, and your colleagues are not waiting for permission. The practical question is "which tasks are prep work, and which tasks are tasting the dish?" That is what the seven categories answer.

Category 1: Intake Summaries, the Highest-Yield Starting Point

An intake generates the most raw material of any encounter in your practice: a 90791 diagnostic evaluation can produce sixty to ninety minutes of history, presenting problem, family background, prior treatment, substance use history, and current functioning. Turning that into a coherent intake summary is exactly the kind of structuring work generative AI does well. Given a transcript or your dictated shorthand, an AI scribe such as Mentalyc, Upheal, Eleos Health, Twofold, or Heidi can produce a chronological history, organize presenting concerns by domain, pull out the prior treatment timeline, and format it into your EHR's intake template. What took forty-five minutes of after-hours synthesis becomes a fifteen-minute review-and-correct pass.

Now mark the clinician-only sub-tasks inside this category, because intake is where they cluster densely. The diagnosis is yours. The model can list symptoms the client reported; it cannot decide that the picture is F43.10 (PTSD) rather than F43.12 or an adjustment disorder, because diagnosis requires clinical judgment applied to a human you observed, not pattern-matching on a transcript. The mental status exam is yours: an MSE is an observation, and the model did not observe anyone. The initial risk formulation is yours, completely. If suicidal ideation surfaced in the intake, the model may transcribe what was said, but the screening instrument, the risk level, and the safety planning decision belong to you. We will hammer this again because it cannot be hammered enough: AI never scores the CSSRS and never assigns a risk level. It can format your completed risk documentation after you have made the determination. The sous-chef can plate the dish you cooked; it cannot decide the dish is safe to serve.

Used inside that boundary, intake summarization is probably the single highest-yield place to start with AI in a behavioral health workflow. The volume of structurable material is highest, the formatting burden is heaviest, and the deliverable, a clean intake summary, is reviewed in full anyway before it enters the record.

Category 2: Progress Note Drafting, Where Maria Gets Her Evenings Back

This is the category that sells the subscriptions, and for good reason. Solo and small-practice clinicians doing 25 to 35 client hours a week routinely bleed 5 to 10 unpaid hours into documentation. An AI scribe that listens to a session (with documented client consent, a topic with its own lessons later in this program) or takes your post-session shorthand can draft a structured SOAP, DAP, or BIRP note in seconds. For Maria, that thin 9:54 PM note, "patient reports continued anxiety, processed trauma material, will follow up next week," becomes a draft with the session's actual clinical content organized into the right fields: the specific intervention used, the client's response, the plan connected to treatment goals.

And this is where the payer dimension matters. A thin note is not just a tired note; it is a medical-necessity problem. A 90837, the 60-minute psychotherapy code, draws payer scrutiny precisely because it pays more, and a United Healthcare high-frequency 90837 review will ask what in the documentation justifies the time and intensity. Maria watched a peer absorb a $14,200 recoupment letter last quarter over exactly this. An AI draft built from the actual session content gives the note the specificity an auditor expects, the named modality, the observable response, the functional impact.

But note the clinician-only sub-tasks here too, because the draft is not the note. The model cannot supply the verifiable details only you hold: the actual time in session that justifies the 90837 versus the 90834, the specific PHQ-9 score and its delta from last administration, whether the intervention you named is the one you actually delivered. The model also cannot decide what belongs in the progress note versus what belongs in your separately protected psychotherapy notes under 45 CFR 164.501, a privacy decision with legal consequences. And the signature is yours alone. Your signature on a note is a legal attestation that the contents are true, not a formatting step. You read every word before you sign, you correct what is wrong, and you never sign a note for a session that is not yours. That rule gets its own lesson two stops from here; plant the flag now.

The model drafts the sentence; the clinician owns the truth of it. Everything AI does well in behavioral health lives on the drafting side of that line, and everything dangerous lives in forgetting where the line is.

Categories 3 and 4: Treatment Plan Structuring and Prior Authorization Letters

Treatment plans are structure-heavy documents: presenting problems mapped to goals, goals broken into measurable objectives, objectives tied to interventions and review dates. Generative AI is excellent at this scaffolding. Give it your clinical formulation, the diagnosis you made, the goals you and the client agreed on, and it will produce a plan formatted to your EHR template with objectives phrased in the measurable, time-bound language utilization reviewers want to see. What it must not do is originate the clinical content. The model does not choose the treatment goals, does not select the modality, and absolutely does not recommend medications. Medication recommendations are clinician-only, full stop, and for non-prescribers they are outside scope entirely. The sous-chef formats the recipe card; the chef wrote the recipe.

Prior authorization and medical-necessity letters are the adjacent win, and for group practices they may be the bigger one. Jordan, who runs a 25-clinician group practice in Sacramento, has a billing manager drowning in Anthem prior-auth letters bouncing back "denied for lack of medical necessity documentation," even when the PHQ-9 trajectory clearly justifies continued care. A medical-necessity letter is persuasive structured writing built on clinical facts, exactly the genre generative AI handles well. Feed it the facts: diagnosis, symptom trajectory with actual measurement scores, functional impairments, treatments tried, response to treatment, risk of deterioration without continued care. It returns a letter organized the way a concurrent-review nurse reads.

The clinician-only core: every clinical fact in that letter must come from you and must be true. The model cannot supply a PHQ-9 delta it never saw, and if it invents one, you have just submitted a false claim under your signature. The hallucination problem, an AI confidently inventing scores, diagnoses, and events, is serious enough that it gets the entire next lesson. For now: AI drafts the argument; you supply and verify every fact inside it.

Categories 5, 6, and 7: Psychoeducation Handouts, Billing Code Suggestion, and Scheduling Reminders

Category five is psychoeducation. Need a one-page handout explaining the CBT model of panic to a 16-year-old, a sleep hygiene sheet at a fifth-grade reading level, or a caregiver explainer on what a PHQ-9 score means? Generative AI produces serviceable drafts in seconds and revises tone, reading level, and language on request. This is among the lowest-risk categories because the output is general education, not an individualized clinical record. The clinician-only sub-task: you verify clinical accuracy before anything reaches a client, and you decide whether the handout is appropriate for this client at this moment. A trauma psychoeducation sheet handed to the wrong client at the wrong stage is a clinical decision gone wrong, and the model has no idea where your client is in treatment.

Category six is billing code suggestion, a classification task. Given a session description, AI can suggest the likely CPT code: 90832, 90834, or 90837 by psychotherapy duration, 90846 or 90847 for family work, 90853 for group, 96127 for brief behavioral screening where covered. As a prompt to check your own coding, useful. As the decider, dangerous. Code selection is a billing attestation under your name, and the difference between a 90834 and a 90837 is a fact about minutes in session that only you can attest to. Treat the suggestion as a junior biller's first pass that you always verify against what actually happened in the room.

Category seven is scheduling reminders and administrative communication: appointment reminders, waitlist messages, rescheduling templates, intake-packet cover emails. Lowest stakes of all seven, real time savings in aggregate, and still one clinician-only thread: any message touching clinical content, and any decision about contacting a client whose record carries safety concerns or confidentiality constraints (an unsafe home, a custody dispute, a 42 CFR Part 2 record), stays with you. Even the lowest-risk category keeps a human decision at its center.

The Clinician-Only List: Five Decisions That Never Move

Run the seven categories back and you will notice the same five decisions kept surfacing inside them. Pull them out and name them, because these are the load-bearing walls of safe AI use in behavioral health, the things that remain clinician-only no matter how good the model gets or how confident the vendor demo sounds.

One: diagnosis. The model can organize reported symptoms; it cannot decide what they mean. Two: risk level. AI never scores the CSSRS, never assigns a suicide risk level, never makes the Tarasoff-type duty-to-protect determination (in California, duty to protect under Civil Code §43.92, not duty to warn). It structures and formats documentation after your determination, never before and never instead. Three: medication recommendations, the territory of prescribers exercising clinical judgment, not language models exercising autocomplete. Four: custody-related opinions. Anything that could be read as an opinion in a custody matter carries forensic weight and board exposure; no AI-drafted sentence containing one should survive your review unexamined, and no AI should generate one in the first place. Five: mandated-report decisions. Whether a disclosure triggers a mandated report under a statute like California WIC §11166 is a legal judgment call made by the mandated reporter, you, with your knowledge of the client, the statute, and the context. The model is not a mandated reporter. It has no duty, no liability, and no judgment.

Here is why this list does not shrink as the technology improves: every item on it is an act of professional judgment that the law locates in a licensed person. Liability, duty, and attestation attach to a license. A model cannot hold a license, cannot be deposed, cannot lose the right to practice, and cannot sit across from a board investigator. When responsibility cannot transfer, the decision cannot transfer. The drafting can; the deciding cannot. That asymmetry, not any limit on model capability, is the permanent architecture of this field.

Why the Line Sits Where It Sits: Three Tests for Any New Task

The seven categories will not cover everything. Next month a vendor will pitch you a feature this lesson never named, and you need a way to sort it yourself. Use three tests, in order.

First, the observation test: does the task require having observed the client? An MSE, an affect description, a risk assessment, a judgment about whether the client minimized, all require eyes and clinical presence in the room. The model was not in the room in any clinically meaningful sense, even if it transcribed the audio. A transcript captures words; it does not capture the flat affect that contradicted them. Anything resting on observation is yours.

Second, the attestation test: does the output become something you sign or certify as true? Notes, billing codes, time in session, medical-necessity letters, mandated reports, all are attestations under your license. AI can draft them; it cannot make them true. You verify every fact before your signature converts a draft into a legal record. Third, the consequence test: if this output is wrong, does a person get hurt or does a license get exposed? A typo in an appointment reminder costs an awkward email. A wrong risk level costs a life. Tasks fail toward the clinician in proportion to their consequence. Run any new AI feature through observation, attestation, consequence, and you will land the task on the right side of the line without waiting for this program to publish a lesson about it. And when a task fails all three tests, delegate the drafting freely, because that is exactly the prep work the sous-chef exists to do.

The Applied Problem: Your One-Page AI Delegation Map

Your artifact for this lesson is the AI Delegation Map: a single page, two columns, taped where you can see it when you open the EHR at 9:54 PM. Left column: "AI Drafts (I Verify)." Right column: "Clinician Only (AI Never Touches)." It sounds simple. The value is in building it from your actual week, not from this lesson's generic list.

Step one: inventory your real workload. Pull up last week's calendar and list every documentation and admin task you actually performed: intakes, progress notes by CPT code, a treatment plan review, a prior-auth letter, two psychoeducation conversations you improvised, the reminder texts. Most clinicians find ten to fifteen distinct task types. Step two: sort each task into a column using the three tests. Use a prompt like this against your own list, with a general-purpose AI tool and no client information of any kind: "I am a licensed behavioral health clinician building a delegation map for AI use in my practice. For each task on this list, tell me which parts are drafting and formatting work, and which parts require clinical observation, legal attestation, or carry serious consequence if wrong. Do not soften the clinician-only side." Then argue with the output, because that argument is the learning.

Step three: the verification pass. Check the right column against this lesson's five fixed items: diagnosis, risk level (including any CSSRS scoring and duty-to-protect determination), medication recommendations, custody-related opinions, mandated-report decisions. If any of the five is missing from your clinician-only column, add it explicitly, even if it feels obvious. Audit findings and board complaints are made of things that felt obvious. Then check the left column for one trap: any "AI drafts" entry that involves a fact only you can verify (minutes in session, a PHQ-9 score, what homework was actually assigned) gets a handwritten asterisk meaning "verify the facts before signing."

Done looks like this: one page, every task from your real week placed in a column, all five clinician-only decisions named explicitly, asterisks on every attestation-bearing draft, and a date at the top, because you will revise this map as the program deepens. When a colleague at consultation group asks what you actually use AI for, you hand them this page instead of an opinion.

Key Takeaways

• The controlling frame is the sous-chef: AI does prep work (drafting, structuring, formatting) at speed and scale, but the chef, the licensed clinician, tastes, decides, and signs. The Illinois WOPR Act and Nevada AB 406 write that same line into statute by prohibiting AI from providing therapy or delivering behavioral healthcare.
• Seven categories of behavioral health work reliably benefit from generative AI: intake summaries, progress note drafting, treatment plan structuring, prior authorization and medical-necessity letters, psychoeducation handouts, billing code suggestion, and scheduling reminders. Each contains clinician-only sub-tasks that never transfer.
• Five decisions stay clinician-only permanently: diagnosis, risk level, medication recommendations, custody-related opinions, and mandated-report decisions. AI never scores the CSSRS, never assigns a risk level, and never makes the duty-to-protect call; it structures documentation only after the clinician's determination.
• The clinician-only list is permanent because liability, duty, and attestation attach to a license, and a model cannot hold one. When responsibility cannot transfer, the decision cannot transfer, regardless of how capable models become.
• Progress note drafting is where the payer stakes live: a thin 90837 note invites a high-frequency review and a recoupment letter. AI drafts add specificity, but the verifiable details, minutes in session, the PHQ-9 delta, the modality actually used, must come from you and be true.
• Sort any new AI feature with three tests: observation (did it require seeing the client?), attestation (do I sign this as true?), and consequence (who is hurt if it is wrong?). Tasks fail toward the clinician in proportion to their consequence.
• Your signature is a legal attestation, not a formatting step. Read every word of every AI draft before signing, correct what is wrong, and never sign for a session that is not yours; the next two lessons show exactly how drafts go wrong and why judgment cannot be delegated.