The biopsychosocial intake is the longest document you will write for any client, the one every later note leans on, and the one most likely to be written at 10:40 PM from a 75-minute interview you only half remember. It is also the document where AI support is most tempting and most dangerous at the same time, because a 90791 intake covers history, and history is exactly what a language model fabricates most fluently. This lesson walks you through drafting a BPS intake with AI support for a realistic new client: a 41-year-old presenting with comorbid generalized anxiety disorder (F41.1) and alcohol use disorder, moderate (F10.20), with structure pulled from the intake interview transcript and a completed AUDIT-C, and with the 42 CFR Part 2 disclosure question answered before the first word of the substance use history is typed. By the end you will have a biopsychosocial intake AI template with a Part 2 block built in, and a workflow that produces a thorough, accurate, defensible intake in a fraction of the late-night hours it costs you now.

What the BPS Intake Actually Does in the Chart

Start with the job description, because the intake's three jobs explain every rule that follows. Job one is clinical: the BPS is your case formulation's raw material, the organized account of biology, psychology, and social context from which diagnosis and treatment planning flow. Job two is administrative: the 90791 psychiatric diagnostic evaluation you billed requires a document that demonstrates a diagnostic evaluation actually happened, with history, mental status, diagnosis, and disposition. Job three is legal: the intake is the baseline record against which everything later is measured. When a board, an auditor, or an attorney reconstructs the case three years from now, the intake is where they start, and gaps or fabrications there contaminate every note that cites it.

Think of the BPS intake as the foundation pour of the chart, the controlling analogy for this lesson. Every later document, progress notes, treatment plan updates, prior-auth letters, sits on it. A foundation poured with voids (missing substance use history, no trauma screen, no medical history) makes everything above it unstable. A foundation poured with the wrong material (AI-invented history the client never gave) is worse than a void, because nobody can see the defect until weight lands on it: the day the chart is subpoenaed, the day the client disputes a detail, the day a new clinician treats from a history that never happened. AI can run the concrete pump. It cannot decide what goes into the mix, and it must never be allowed to quietly substitute filler.

Our worked client for the lesson: D.M., 41, self-referred after a partner's ultimatum, presenting with persistent worry, sleep disruption, and drinking that has grown from social to daily. Intake interview recorded with documented consent through a BAA-covered AI scribe; AUDIT-C completed in the waiting room, score 7 of 12, well past the positive threshold. Your diagnostic conclusions after the interview: generalized anxiety disorder, F41.1, and alcohol use disorder, moderate, F10.20. Those diagnoses are yours, made from criteria and clinical judgment; the transcript and the AUDIT-C are evidence you weighed, not conclusions the software reached.

The Part 2 Question Comes First, Not Last

Before drafting a word of D.M.'s substance use history, answer the regulatory question that determines how that history must be guarded: does 42 CFR Part 2 apply? Part 2 is the federal confidentiality rule for substance use disorder records, updated by the 2024 final rule, and it is stricter than HIPAA in ways that matter at the level of individual paragraphs in your chart. The applicability test for a practice like yours turns on whether you hold yourself out as providing substance use disorder diagnosis, treatment, or referral for treatment, that is, whether you are a Part 2 program or part of one. A general outpatient therapy practice that encounters SUD incidentally may sit outside Part 2; a practice that advertises addiction treatment, runs an IOP, or bills SUD-specific service lines like H0015 likely sits inside it. This is a determination to make deliberately, in writing, with competent guidance if the answer is unclear, because the consequences attach to every SUD record you create from then on.

If Part 2 applies, D.M.'s record needs the disclosure block: the written prohibition-on-redisclosure notice that must accompany disclosures of Part 2 records. The worked artifact, adapted from the regulatory language, reads: "This record which has been disclosed to you is protected by federal confidentiality rules (42 CFR Part 2). These rules prohibit you from making any further disclosure of records that identify a patient as having or having had a substance use disorder, either directly or by implication, unless further disclosure is expressly permitted by the written consent of the individual whose information is being disclosed, or as otherwise permitted by 42 CFR Part 2. A general authorization for the release of medical or other information is NOT sufficient for this purpose. 42 CFR Part 2 restricts any use of the information to investigate or prosecute, with regard to a criminal charge, any patient." Confirm the current required wording against the regulation text in force when you build your template, because the 2024 final rule revised the notice language.

The practical scenario the playbook flags: D.M. signs a release so you can coordinate with his primary care physician, who is not a Part 2 program. If you are a Part 2 program, the disclosure to the PCP must ride on Part 2-compliant consent, and the records you send carry the redisclosure notice, because the PCP's office cannot be allowed to treat your SUD documentation as ordinary medical records that flow onward freely. One more rule that AI tooling makes urgent: if Part 2 applies, your AI scribe vendor is processing Part 2 records, and your agreements must cover that, not just HIPAA. A vendor whose paperwork has never heard of Part 2 is a vendor you do not feed an SUD intake.

The BPS Skeleton: A Section-by-Section Build

Here is the worked skeleton, the structure your AI draft will be poured into. 1. Identifying information and referral source: age, pronouns, living situation, who sent them and why now. For D.M.: 41, lives with partner of nine years, self-referred after partner's ultimatum regarding drinking. 2. Presenting problem in the client's own words: quoted, not paraphrased; "I can't turn my head off, and the only thing that turns it off is a drink." 3. History of presenting problem: onset, course, severity, prior episodes, what has been tried. 4. Psychiatric history: prior diagnoses, treatment episodes, hospitalizations, medication trials and responses. 5. Substance use history: substances, quantity, frequency, pattern over time, longest abstinence, prior treatment, withdrawal history, consequences; anchored to the AUDIT-C of 7 and your follow-up assessment. This is the section the Part 2 determination governs. 6. Medical history: conditions, medications, allergies, last physical, relevant labs; for a daily drinker, liver and withdrawal-risk questions are clinical necessities, not formalities. 7. Family history: psychiatric and substance use history in the family system. 8. Social history: relationships, employment, housing, legal involvement, finances, culture, faith, supports. 9. Developmental and trauma history: screened explicitly, never assumed silent. 10. Risk assessment: suicide and violence risk, clinician-assessed and clinician-written, every word yours; AI never scores risk, never assigns a level, and this section is composed by you in your own words after your own assessment. 11. Mental status examination: your in-session observations (the next lesson covers why AI cannot do this section alone). 12. Diagnostic formulation: DSM-5-TR reasoning to F41.1 and F10.20, including what you ruled out and why. 13. Initial treatment recommendations and disposition: level of care, modality, referrals (PCP coordination, possible medication evaluation), and the measurement cadence you will run.

Two structural observations a supervisor would underline. First, the skeleton is also your interview completeness check: run it before the client leaves, because AI can only structure what you actually asked, and the most common intake defect is not bad writing but unasked questions. Second, sections 10, 11, and 12, risk, MSE, and diagnosis, are clinician-authored zones. The transcript-driven sections (history, social context) are where AI support earns its keep; the judgment sections are where it is excluded by design.

An intake's history sections can be assembled from the transcript; its judgment sections can only be assembled from you. The craft of AI-supported intake writing is knowing exactly where that line runs and never letting the draft blur it.

From Transcript to Draft: The Extraction Prompt

With the skeleton fixed and the Part 2 question answered, the drafting step is a structured extraction, not a creative writing exercise. Inside your BAA-covered platform (and Part 2-covered, if applicable), the working prompt: "From the attached intake interview transcript and the completed AUDIT-C (score 7), populate the following biopsychosocial intake sections: identifying information, presenting problem (quote the client's own words), history of presenting problem, psychiatric history, substance use history, medical history, family history, social history, developmental and trauma history. For each statement, stay strictly within what the transcript contains. Where the transcript does not address a section, write exactly: [NOT ASSESSED IN INTERVIEW]. Do not write the risk assessment, mental status exam, diagnostic formulation, or treatment recommendations; leave those sections as empty headers for the clinician. Do not infer diagnoses, quantities, dates, or family history not explicitly stated."

The [NOT ASSESSED IN INTERVIEW] instruction is the single highest-value line in the prompt. A language model's default behavior when a section is empty is to fill it plausibly: a "normal developmental history," an "unremarkable medical history," two phrases that look complete and assert facts no one gathered. Forcing an explicit gap marker converts hallucination risk into a visible to-do list: every bracket in the draft is either a question for your next session with D.M. or an honest "not assessed" you document deliberately. An intake that admits its gaps is defensible; an intake that papers over them with invented normality is a falsified record wearing good formatting.

Expect, and hunt for, the known intake-specific failure modes in the verification pass. Quantity drift: the client said "four or five drinks most nights" and the draft says "six drinks nightly." Timeline smoothing: scattered references to "a rough patch after the layoff" become a confident "symptoms began in March 2024." Family history invention: one mention of "my dad drank" becomes "significant paternal history of alcohol use disorder and depression." Quote laundering: the client's exact words get paraphrased into clinical language and then presented inside quotation marks. Each of these survives a casual read and dies under your line-by-line check against the transcript, which is why the check is non-negotiable.

Writing the Clinician-Authored Zones

Now the sections the prompt deliberately left empty. The risk assessment is first and absolute: you assessed D.M.'s suicide risk directly in the interview (a daily drinker with relationship rupture and sleep loss is not a low-acuity profile by default), and you write that assessment in your own words: ideation, plan, intent, means, protective factors, your assigned risk level, and your safety actions. AI's only role here, if any, is formatting what you have already written or dictated, never composing it. The same applies to any mandated-report consideration and any duty-to-protect analysis: clinician determinations, clinician prose, AI strictly afterward if at all.

The diagnostic formulation is where the intake becomes a clinical document instead of a transcript summary. For D.M., your written reasoning connects criteria to evidence: excessive anxiety and worry more days than not across work, health, and relationship domains, difficult to control, with sleep disturbance, muscle tension, and irritability, duration exceeding six months, supporting F41.1; alcohol consumed in larger amounts than intended, unsuccessful efforts to cut down, craving, continued use despite relational consequences, and tolerance, four-plus criteria supporting F10.20 at moderate severity. Then the differential you considered: is the anxiety substance-induced, given daily alcohol use and possible withdrawal-mediated rebound anxiety? You document why you currently judge GAD to be independent (anxiety history predating the escalation of drinking, per the client's account) and you flag that the formulation will be revisited as drinking changes, because honest provisional reasoning is stronger under review than false certainty. AI did not and cannot do this paragraph; it is the clinical core of the 90791.

Disposition closes the document: weekly individual therapy with an anxiety-and-SUD-informed approach, coordination with the PCP under the appropriate consent (Part 2-compliant if applicable), referral discussion for medication evaluation including possible withdrawal-risk consultation given daily use, and the measurement cadence: GAD-7 and AUDIT-C re-administration on schedule, PHQ-9 at intake baseline. Every element of disposition is a decision you made; the draft merely records it cleanly.

The Verification Pass and the Signature

The verification pass for an intake is longer than for a progress note because the document is longer and the stakes compound. Work with the transcript open in one pane and the draft in the other. Pass one, provenance: every factual claim in the history sections must trace to a transcript moment or the AUDIT-C; anything you cannot locate gets struck or bracketed. Pass two, quotes: every quoted phrase checked verbatim against the transcript; paraphrase presented as quotation is corrected on sight. Pass three, gaps: every [NOT ASSESSED IN INTERVIEW] marker is resolved into either a documented plan to assess ("developmental history to be completed at session two") or a deliberate clinical note that it was deferred and why. Pass four, the judgment zones: confirm risk, MSE, formulation, and disposition are entirely your prose, and that nothing in the AI-drafted sections quietly asserts a diagnosis, a risk conclusion, or a treatment decision.

Then the Part 2 pass, if applicable: the substance use history is complete and accurate; the record is flagged per your practice's Part 2 procedures; the redisclosure notice block is attached to the template that will accompany any disclosure; and the PCP coordination is queued behind a Part 2-compliant consent rather than a general release. This is also the moment to confirm the AI vendor question was answered before the transcript ever reached the tool, because verification cannot retroactively fix a disclosure that already happened.

Finally the signature, and the cardinal rule of every documentation lesson in this program: the clinician signs the note, and the signature is a legal attestation, not a formatting step. For an intake, the attestation is heavier than usual, because you are attesting a diagnostic evaluation, a risk assessment, and a multi-domain history in one document. Read every word before signing. The twenty minutes that takes is the cheapest malpractice premium you will ever pay.

The Applied Problem: The BPS Intake Template with Part 2 Block

Your deliverable is the BPS Intake Template with Part 2 Block, built once and reused for every intake. Step one: write the thirteen-section skeleton from this lesson into your EHR or document system as a template, with the three clinician-authored zones (risk assessment, MSE, diagnostic formulation) visually marked, a header note that reads "CLINICIAN-AUTHORED: AI drafting prohibited in this section" works, and disposition flagged as decisions-only.

Step two: make the Part 2 determination for your practice in writing: do you hold yourself out as providing SUD diagnosis, treatment, or referral for treatment? Document the answer, the reasoning, and the date. If yes, or if the answer is truly uncertain, add the redisclosure notice block from this lesson to the template as a standing component of any SUD-record disclosure, verify the wording against the current 42 CFR Part 2 text, and add a template line for Part 2-compliant consent before any coordination disclosure, including the PCP scenario. If no, document that determination too; an auditor finding a written, dated, reasoned determination treats you very differently than one finding silence.

Step three: embed the extraction prompt in the template's instructions, including the [NOT ASSESSED IN INTERVIEW] line and the explicit exclusion of the judgment zones. Step four: run the pilot on your next intake (or on D.M.'s scenario as a dry run): record with documented consent in a BAA-covered (and Part 2-covered, if applicable) tool, complete the AUDIT-C, run the extraction, then execute all five verification passes from this lesson with the transcript open. "Done" looks like: a reusable thirteen-section template; a written, dated Part 2 determination; the redisclosure block ready to travel with any SUD disclosure; an extraction prompt that marks gaps instead of inventing normality; and a first completed intake in which every history claim traces to the transcript, every judgment section is in your voice, and your signature went on last, after a full read.

Key Takeaways

• The BPS intake does three jobs at once: clinical raw material for formulation, the documentation that justifies the 90791, and the legal baseline the whole chart sits on. It is the foundation pour of the record, and AI-invented history is worse than a gap because the defect stays invisible until weight lands on it.
• Answer the 42 CFR Part 2 question before drafting any substance use history: does the practice hold itself out as providing SUD diagnosis, treatment, or referral for treatment? Make the determination in writing, dated and reasoned, and if Part 2 applies, the record carries the prohibition-on-redisclosure notice and discloses only under Part 2-compliant consent, including to a PCP who is not a Part 2 program.
• The thirteen-section skeleton is both the document structure and the interview completeness check: identifying information through disposition, with risk assessment, MSE, and diagnostic formulation marked as clinician-authored zones where AI drafting is excluded by design. AI structures the transcript-driven history sections; judgment sections come only from you.
• The extraction prompt's highest-value line forces the model to write [NOT ASSESSED IN INTERVIEW] wherever the transcript is silent, converting hallucination risk into a visible to-do list. An intake that admits its gaps is defensible; one that papers over them with invented "unremarkable" histories is a falsified record in good formatting.
• Hunt the intake-specific failure modes in verification: quantity drift in substance use amounts, timeline smoothing into false precise onsets, family history invention from single mentions, and quote laundering where paraphrase appears inside quotation marks. Each survives a casual read and dies under a line-by-line check against the transcript.
• The diagnostic formulation for comorbid F41.1 and F10.20 is the clinical core AI cannot write: criteria connected to evidence, the substance-induced-anxiety differential considered, and provisional reasoning documented candidly, with a plan to revisit as drinking changes. Risk assessment is composed entirely in the clinician's own words after the clinician's own assessment; AI formats afterward at most.
• The signature on an intake attests a diagnostic evaluation, a risk assessment, and a multi-domain history at once; it is a legal attestation, not a formatting step. Read every word against the transcript before signing, and treat the time it takes as the cheapest malpractice premium you will ever pay.