The intake is where every later problem in a chart is born. The payer denial at month four traces back to a treatment plan with no measurable goals. The board complaint defense collapses because the consent never mentioned AI. The 90837 recoupment starts with a biopsychosocial that never established medical necessity. Maria's intakes cost her three and a half hours of documentation per new client, all unpaid past the 90791 itself. This lesson walks the complete therapy intake workflow with AI at every stage where AI belongs, the clinician verification step at every stage where judgment lives, and the signature points that turn drafts into legal records. By the end you will have a replicable intake cycle, first contact to signed treatment plan, that you can run for every new client and defend to any auditor.

The Intake as a Five-Station Assembly Line

Think of the intake cycle as an assembly line with five stations; that is the controlling analogy for this lesson. Station one: first contact and screening triage. Station two: the consent packet. Station three: the biopsychosocial assessment, the BPS, built from the 90791 diagnostic interview. Station four: the diagnostic formulation against DSM-5-TR criteria. Station five: the treatment plan, signed by you and, in most settings, by the client. At each station there are three questions: what does the AI produce, what does the clinician verify, and where does a signature convert a draft into an attestation. If you cannot answer all three for a station, you do not have a workflow there; you have a habit, and habits do not survive audits.

The analogy matters for a second reason: every station has a quality gate, and a defect that passes one gate compounds at every later station. A screening call that misses active suicidal ideation produces a BPS that underweights risk, a diagnosis that misses severity specifiers, and a treatment plan that fails to justify the session frequency you intend to bill. Payers read charts backwards: the concurrent-review nurse starts at the treatment plan and asks whether the BPS supports it. Build the workflow so each station's output is verified before the next station consumes it.

One framing rule: AI drafts and the clinician decides, at every station. The AI never conducts the assessment, never assigns the diagnosis, never determines risk, never selects the level of care. After the Illinois WOPR Act and Nevada AB 406, AI performing the clinical acts of assessment and diagnosis is the statutory line, and your intake is the document trail proving you stayed on the right side of it.

Station One: First Contact and AI-Drafted Screening Triage

First contact is a voicemail, a portal inquiry, or a referral from Headway or Alma. The clinical question is fit and urgency: is this person appropriate for your practice, and do they need care faster than your next opening? The AI touchpoint is drafting, not deciding. Your intake form or screening call notes go to an AI assistant under a BAA, with a prompt you save as a reusable block: "From the attached screening responses, draft a triage summary with these sections: presenting concern in the caller's words, requested service, stated urgency indicators (quote any language about self-harm, harm to others, recent hospitalization, or substance use verbatim and flag it), insurance and scheduling constraints, and open questions for the clinician. Do not assign a risk level. Do not recommend acceptance or referral. Quote, do not paraphrase, any risk-adjacent language."

Read that prompt again, because its restrictions are the workflow. The AI surfaces and quotes; you decide. If the form says "I have been thinking about ending things," the AI's job is to put that sentence, verbatim, at the top with a flag. Your job is everything after: the same-day callback, the risk screening you conduct yourself, the decision about crisis resource, higher level of care, or your Tuesday 4 PM slot. The verification step at station one is reading the original screening responses, not just the summary, whenever the AI flagged risk language or your own pattern recognition itches. The signature point: your triage disposition note ("reviewed screening, no acute risk indicators, offered intake appointment 6/9") is the first dated entry in the record, and you write or approve every word of it.

Where standardized screeners enter: send a PHQ-9 and GAD-7 with the intake paperwork, and where responses indicate alcohol or drug use, an AUDIT-C or DAST-10. The AI can transcribe item responses into EHR fields and compute the arithmetic total, but you verify the score against the raw items before it enters the record, and you, not the software, interpret what a PHQ-9 of 18 means for triage. Item 9, the self-harm item, gets your eyes on the raw response every single time, regardless of the total.

Station Two: The Consent Packet, Including the AI Addendum

Before the diagnostic interview, the consent packet has to be complete, and a complete packet has three layers. Layer one is your standard informed consent to treatment: nature of services, limits of confidentiality, mandated reporting, fees, cancellation policy. Layer two, where any part of care happens by video or phone, is telehealth informed consent, which must address the technology used, the limits and risks of remote care, emergency procedures including the client's physical location at each session, and your state's requirements. Layer three is the AI consent addendum: a plain-language disclosure that you use an AI documentation assistant, what it does (drafts notes and summaries for your review), what it does not do (make clinical decisions, communicate with the client, determine diagnosis or risk), the vendor relationship including the BAA, the client's right to decline, and what declining changes about their care (nothing clinical; you write notes the slower way).

The AI touchpoint here is administrative: AI drafts the addendum language from your policy, populates the packet, and drafts the plain-language explanation for a client who asks what the addendum means. The verification step is legal: confirm the addendum matches the tool you actually use (a Mentalyc addendum does not cover an Upheal deployment), the telehealth consent matches your state's requirements, and nothing in the packet promises confidentiality your tools cannot deliver. The signature points here belong to the client: signed informed consent, signed telehealth consent where applicable, signed AI addendum, all dated before the interview begins. An intake conducted before consent is signed is a chart defect you cannot repair; an AI scribe running during a session the client never consented to record is a board complaint with your name pre-filled.

Practical sequencing: send the packet electronically with the screeners, confirm signatures before the appointment, and open the intake by verbally confirming the AI addendum ("you signed the form about my documentation assistant; any questions?"). That sixty-second confirmation, documented in your intake note, is the difference between consent on paper and consent you can testify to.

Station Three: The Biopsychosocial From the Interview Transcript

The 90791 diagnostic interview is yours: sixty to ninety minutes of history-taking, mental status examination, risk screening, and rapport-building that no software performs. The AI touchpoint comes after, and it is the biggest time recovery in the cycle. With consent in place, the transcript or your post-session dictation goes to the AI with a structured prompt: "From this intake transcript, draft a biopsychosocial assessment with sections: identifying information, presenting problem and its history, psychiatric history including prior treatment and hospitalizations, medical history and current medications, substance use history, family psychiatric history, developmental and social history, trauma history, current functioning (occupational, relational, ADLs), mental status examination observations as stated by the clinician, strengths and protective factors, and clinician-stated risk screening results. Use only information present in the transcript. Mark any section with insufficient information as [NOT ASSESSED OR NOT IN TRANSCRIPT] rather than inferring. Quote the client directly for presenting problem and any risk-related content."

The instructions doing the heavy lifting are the insufficiency markers and the no-inference rule. A BPS's most dangerous failure mode is the plausible fabricated detail: the AI infers "no family psychiatric history" because none was mentioned, when you simply ran out of time to ask. "[NOT ASSESSED]" is an honest gap you close at session two; a confident "denied" you never elicited is a falsified record. Your verification pass is a section-by-section read against the transcript: every quote checked, every "denies" confirmed as something the client actually denied rather than never asked, every screener score (PHQ-9 and GAD-7 from the packet, PCL-5 if trauma screening indicated it, AUDIT-C or DAST-10 where substance use surfaced) verified against the raw instruments. The mental status exam deserves special attention: the MSE is your observation, and the AI may only restate observations you dictated, never generate "well-groomed, cooperative, euthymic" boilerplate you did not record.

The signature point at station three: the completed BPS, edited and signed by you, dated for both the interview and the signature. In supervised settings the supervisor co-signs, which means the supervisee's verification pass must be documented well enough that the co-signature is informed rather than ceremonial.

The intake chart is read backwards by everyone who matters: the auditor starts at the treatment plan and asks whether the assessment supports it. Build each station so the next one can stand on it, and sign nothing you have not read against the source.

Station Four: The DSM-5-TR Diagnostic Formulation

Diagnosis is where the AI's role shrinks to almost nothing, on purpose. Matching assessment data against DSM-5-TR criteria, selecting the principal diagnosis, specifiers, and any provisional or rule-out designations is a clinical act reserved to you. The permitted AI touchpoint is organizational: after you have decided, the AI drafts the formulation paragraph connecting your BPS findings to the criteria you determined were met. A defensible prompt: "I have diagnosed F43.10, PTSD. From the attached BPS, draft a diagnostic justification paragraph organizing the documented symptoms under the DSM-5-TR criterion clusters for PTSD (intrusion, avoidance, negative alterations in cognition and mood, alterations in arousal and reactivity), the duration criterion, and functional impairment, using only symptoms documented in the BPS." The diagnosis appears in your prompt because you made it first; the AI is formatting your reasoning, not generating a candidate diagnosis for you to ratify.

The wrong version of this station is asking "what diagnosis fits this presentation?" and accepting the answer. That inverts the workflow: the AI performed the clinical act and you performed the formatting. Beyond the WOPR-line problem, it produces worse clinical work, because LLMs pattern-match toward the statistically common formulation and will under-diagnose comorbidity, miss specifiers, or anchor on the presenting complaint while quieter data points elsewhere. Your verification step at station four is the criteria check itself: open the DSM-5-TR, walk the criteria, confirm each criterion the paragraph cites maps to a documented symptom in the BPS. If the paragraph cites a symptom the BPS does not contain, the paragraph is wrong even if the diagnosis is right, and an auditor will treat the gap as an unsupported diagnosis.

The signature point: the diagnosis enters the record under your signature, typically inside the BPS or the assessment section, and becomes the medical-necessity anchor for everything billed afterward. Every future progress note, treatment plan review, and prior authorization argues from this diagnosis. Write it like the foundation it is.

Station Five: The Treatment Plan and Its Signatures

The treatment plan converts diagnosis into a contract: problems, goals, objectives, interventions, frequency, projected duration. Payers deny against this document more than any other. The AI touchpoint: "From the attached signed BPS and my diagnosis of F43.10, draft a treatment plan with two problem statements, each with one long-term goal, two measurable short-term objectives with target dates and a measurement method (name the instrument: PHQ-9, GAD-7, or PCL-5, with current baseline score and target score), the evidence-based intervention I will use (I have selected trauma-focused CBT), and proposed frequency of weekly 90837 sessions with clinical justification drawn from the documented severity and functional impairment." Notice what you supplied: the diagnosis, the modality, the frequency. The AI assembles your decisions into payer-ready structure, including the part most hand-written plans omit: the baseline score and numeric target that make an objective auditable. "Client will reduce PCL-5 from baseline 52 to below 33 within 12 weeks, measured every fourth session" survives review; "client will process trauma" does not.

Your verification pass on the plan draft checks four things. Measurability: every objective has an instrument, a baseline, a target, and a date. Necessity linkage: the stated frequency is justified by documented severity and impairment from the BPS, because weekly 90837 for a mild presentation is the pattern that triggers high-utilization review. Modality honesty: the named intervention is one you are trained in and will actually deliver, because the plan becomes the standard your progress notes are audited against. Client voice: goals reflect what the client said they wanted, in language they would recognize.

Signature points, plural: you sign and date the plan; the client signs where the payer expects evidence of collaborative treatment planning (most Medicaid programs do, and a client signature is the cleanest evidence); a supervisor co-signs for pre-licensed clinicians. The fully signed plan, atop the signed consents, signed BPS, and documented diagnosis, closes the intake cycle: one to two weeks of calendar time and, with the AI touchpoints in place, roughly half the documentation hours Maria spends today.

The Verification Economy: Where Your Saved Time Actually Goes

A workflow lesson owes you honest arithmetic. The AI does not eliminate intake documentation; it converts composition time into verification time at a favorable exchange rate. Drafting a BPS from scratch costs Maria 75 to 90 minutes; verifying an AI draft against the transcript costs 20 to 30, and verification produces a better document because she reads critically instead of generating under fatigue. The treatment plan drops from 45 minutes to 15. Across one new client that is roughly 90 to 120 minutes recovered, all of it from composition, none from judgment, because the judgment minutes (interview, MSE, risk screening, criteria walk, frequency decision) were never on the table.

The discipline that keeps the exchange rate favorable is refusing to skim. Clinicians who skip verification do not save additional time; they defer cost into the audit, the complaint, or the resubmitted claim, at a brutal interest rate. The clinicians who get burned by AI documentation are almost never burned by the AI; they are burned by their own signature on a draft they did not read. The signature is a legal attestation, not a formatting step, and at five stations in this cycle it is the load-bearing wall.

One more note for group practices: the intake cycle is the easiest workflow to standardize, because its stations are identical for every new client. Jordan's 25-clinician practice gets more compliance value from one shared, written intake cycle than from any per-clinician improvisation, and the malpractice carrier's AI questionnaire is far easier to answer when the workflow exists on paper.

The Applied Problem: The Intake Cycle Runbook

Your artifact is the Intake Cycle Runbook: a one-page, five-row table you will use for every new client and hand to every clinician you supervise. Each row is a station; the columns are AI Produces, Clinician Verifies, and Signature Point. Build it now.

Row one, Screening Triage: AI produces the triage summary with verbatim risk quotes and computed screener arithmetic (PHQ-9, GAD-7, AUDIT-C/DAST-10 where indicated); clinician verifies raw responses behind every flag, reads item 9 directly, and makes the disposition; signature point: the dated, clinician-authored disposition note. Row two, Consent Packet: AI produces populated documents and plain-language addendum text; clinician verifies the addendum matches the actual tool and BAA and the telehealth consent matches state rules; signature points: client signs informed consent, telehealth consent where applicable, and the AI addendum, before the interview. Row three, BPS: AI produces the structured draft from transcript with [NOT ASSESSED] markers and no inference; clinician verifies section by section against the transcript, every "denies," every screener score against raw items, and an MSE containing only the clinician's stated observations; signature point: clinician signs, supervisor co-signs where required. Row four, Diagnosis: AI produces the justification paragraph for the diagnosis the clinician already made; clinician verifies every cited criterion against the DSM-5-TR and the BPS; signature point: diagnosis entered under clinician signature. Row five, Treatment Plan: AI produces the structured plan from clinician-supplied diagnosis, modality, and frequency, with baselines and numeric targets; clinician verifies measurability, necessity linkage, modality honesty, and client voice; signature points: clinician, client, supervisor as required.

Now pressure-test the runbook. First, can a colleague run a complete intake from this page alone? If not, the missing instruction goes in the table. Second, does any row let AI output reach a signature without a named verification act in between? If so, the row is broken; fix it before a chart proves it. Third, does it handle exceptions: risk flags at station one route to same-day clinician callback before any scheduling, and AI-addendum declines at station two route to the manual-documentation track with no change in clinical care. Add both as footnotes.

"Done" looks like this: the runbook is saved and dated in your AI governance file, your next intake runs against it, and at the end you can point to five signatures, each sitting on top of a verification act you performed, each defensible in the audit that is coming for somebody in your consultation group this year.

Key Takeaways

• The intake cycle is a five-station assembly line: screening triage, consent packet, biopsychosocial assessment, DSM-5-TR diagnostic formulation, treatment plan. Every station answers three questions: what the AI produces, what the clinician verifies, and where a signature converts a draft into a legal attestation. Defects compound forward, and auditors read the chart backwards from the plan.
• At screening, AI drafts the triage summary and must quote risk language verbatim without assigning risk levels or recommending dispositions. The clinician reads raw responses behind every flag, reads PHQ-9 item 9 directly every time, and makes the triage decision. AUDIT-C and DAST-10 join the set where substance use is indicated.
• The complete consent packet has three layers: standard informed consent, telehealth informed consent where care is remote, and the AI consent addendum disclosing what the assistant does and does not do, the BAA, and the client's right to decline without clinical consequence. All three are client-signed and dated before the diagnostic interview begins.
• The BPS is drafted by AI from the transcript under a prompt that forbids inference and requires [NOT ASSESSED] markers for gaps. The clinician verifies section by section: every quote, every "denies," every screener score against raw items, and an MSE containing only the clinician's own observations.
• Diagnosis is the clinician's act. The AI's only role is drafting the justification paragraph after the clinician has diagnosed, organizing documented symptoms under the criteria the clinician determined were met. Asking AI "what diagnosis fits?" inverts the workflow, crosses the statutory line, and yields pattern-matched, comorbidity-blind formulations.
• The treatment plan needs instrument-anchored objectives: named measure, baseline score, numeric target, target date. The clinician supplies diagnosis, modality, and frequency; AI assembles the structure; the clinician verifies measurability, necessity linkage, modality honesty, and client voice before clinician, client, and supervisor signatures close the cycle.
• AI converts composition time into verification time, recovering 90 to 120 minutes per intake, but only if verification actually happens. Your artifact, the Intake Cycle Runbook, is the one-page table that makes every verification gate and signature point explicit enough to hand to a supervisee.