Jordan's practice never set out to run a substance use disorder program. But the website lists "addiction counseling" under services, two clinicians hold SUD certifications featured in their bios, and a referral relationship with a local MAT prescriber sends a steady stream of clients whose intake AUDIT-C scores tell the real story. Last month the practice signed a single AI scribe vendor for all twenty-five clinicians, with one BAA covering everything. If any part of the practice "holds itself out" as providing SUD treatment, that architecture is a 42 CFR Part 2 violation waiting for its first complaint, because the federal confidentiality regime for SUD records does not bend to the convenience of a single vendor contract. By the end of this lesson you will be able to run the holds-itself-out analysis for your own practice, build the consent, segmentation, and redisclosure architecture the 2024 final rule requires, and implement vendor segmentation at the tool level so the AI that drafts general therapy notes never touches a Part 2 record it is not contractually cleared to hold.

The Second Vault Inside the Practice

Think of 42 CFR Part 2 as a vault inside your vault. HIPAA already locks the building: every record in a behavioral health practice is PHI, every vendor that touches it needs a BAA, every disclosure follows the privacy rule. Part 2 builds a second, smaller vault inside that building for one category of records, those created by a federally assisted program that holds itself out as providing substance use disorder diagnosis, treatment, or referral for treatment. Records inside the second vault carry stricter rules: consent requirements with specific content, a redisclosure prohibition that travels with the record, and penalties that, since the 2024 final rule aligned enforcement with HIPAA, now run through the same civil and criminal enforcement machinery that OCR uses for HIPAA violations.

The 2024 final rule changed the texture of Part 2 without removing the vault. It aligned Part 2 more closely with HIPAA, introduced a single-consent option under which a patient can sign one consent covering all future uses and disclosures for treatment, payment, and health care operations, and revised the redisclosure rules so that records disclosed under that consent can be redisclosed by HIPAA-covered recipients largely as HIPAA permits, with critical exceptions, most prominently the enduring restrictions on use in legal proceedings against the patient. What the final rule did not do is dissolve the boundary. Part 2 records remain a legally distinct class. The consent must still meet Part 2's content requirements. The notice accompanying disclosures still matters. And a vendor that holds Part 2 records needs contractual language that addresses Part 2 specifically, not a generic behavioral health BAA that never mentions it.

For the practice owner, the second vault creates one threshold question and two architecture questions. Threshold: does any part of my practice hold itself out as providing SUD treatment, making it a Part 2 program? Architecture: if yes, which records live in the second vault, and which vendors are cleared to enter it? This lesson takes them in order.

The Holds-Itself-Out Analysis: What Your Website Says You Are

"Holds itself out" is the phrase that decides whether the second vault exists in your practice at all, and it is broader than most owners assume. A program holds itself out as providing SUD treatment when it presents itself to the public as a place to get SUD diagnosis, treatment, or referral, through any channel: the services page that lists "addiction counseling" or "substance abuse treatment," the clinician bios advertising SUD certifications, the insurance directory entries listing SUD specialties, the referral relationships marketed to MAT prescribers, the H0015 intensive outpatient billing or the SUD-specific program track with its own intake. The analysis is about the practice's public posture, not its internal intent. Jordan never decided to run a Part 2 program; Jordan's website decided for him.

The inverse matters just as much, because over-applying Part 2 is its own failure mode. A general psychotherapy practice where clinicians sometimes treat clients who happen to have co-occurring alcohol use disorder, documented as F10.20 alongside the presenting F41.1, is not automatically a Part 2 program. Treating SUD incidentally, without holding the practice out as an SUD treatment destination, does not create the second vault. Maria, the solo LCSW in Oakland, sees clients whose drinking comes up in session every week; her chart is HIPAA, not Part 2, because nothing about her public posture says "come here for addiction treatment." The line is posture, not diagnosis codes. A practice that wrongly assumes Part 2 applies everywhere will wrap every record in consent machinery clients do not need, slow every coordination letter, and burn goodwill with every PCP who waits for paperwork the law never required.

The analysis also runs at the unit level, not only the practice level. A group practice can contain a Part 2 program without being one entirely: the SUD track with its own marketing, intake, and program structure is the Part 2 program; the general psychotherapy caseload is not. This unit-level reading is what makes segmentation possible and necessary. The records of the SUD track go in the second vault. The general records do not. The architecture question becomes: can your EHR, your workflows, and your AI vendors actually keep them apart?

Part 2 status is decided by your public posture, not your intentions: the services page, the bios, and the directory listings answer the holds-itself-out question before you do.

The Consent Architecture Under the 2024 Final Rule

If the analysis says yes, the first build is consent. The 2024 final rule's single-consent option is the workhorse: the patient signs one consent at intake that covers all future uses and disclosures for treatment, payment, and health care operations, eliminating the old regime's per-disclosure consent burden for routine care coordination. But the consent must still be a Part 2 consent, with the content Part 2 requires, naming or describing the recipients, stating the purpose, explaining the right to revoke, and making clear what the patient is agreeing to. A HIPAA notice of privacy practices does not substitute. The intake packet for the SUD track carries its own consent instrument, and your front-desk workflow has to route clients to the right packet based on which program they are entering, which is the first place segmentation becomes operational rather than theoretical.

Redisclosure is the second consent-architecture component, and it is where AI-drafted correspondence creates daily exposure. When a Part 2 record leaves the program under consent, the disclosure carries notice obligations, and the classic redisclosure prohibition language tells the recipient that the record is protected by federal confidentiality rules and may not be redisclosed except as permitted. Under the 2024 final rule, a HIPAA-covered recipient who receives records under the single TPO consent may generally redisclose under HIPAA's rules, but the protections against use in proceedings against the patient persist, and disclosures outside the single-consent pathway still carry the stricter regime. The practical pattern this program has taught since Level 1: every coordination letter, every ROI response, every continuity-of-care letter that contains Part 2 content gets the redisclosure block, and the AI that drafts those letters must be configured with a template that includes it by default, because the model will not know which records came from the second vault unless your architecture tells it.

The third component is the scenario the playbook calls out as the one clinicians actually face: the Part 2 client who also has a PHI release to a primary care doctor who is not a Part 2 program. The MAT coordination letter to that PCP is a Part 2 disclosure. It needs the consent on file, the minimum-necessary scoping, and the redisclosure language. The general-therapy progress note about the same client's anxiety work may sit outside the second vault if the practice's segmentation truly separates the SUD program record from the general record, which is exactly why the segmentation has to be real and not aspirational.

Vendor Segmentation at the Tool Level

Here is the rule this lesson exists to install: the AI vendor that handles general therapy notes is not the AI vendor that handles Part 2 SUD notes unless the BAA explicitly covers Part 2. Read that sentence as an architecture requirement, not a slogan. Most behavioral health AI scribes sign a generic BAA written for HIPAA. A generic BAA makes the vendor a business associate; it does not make the vendor a party that has acknowledged the second vault's rules, the redisclosure obligations, the restrictions on use of Part 2 records, or the consequences of treating Part 2 records as ordinary PHI in its subprocessor chain. When your SUD-track clinician runs a session through a scribe whose contract never mentions Part 2, you have moved a record out of the second vault into a vendor environment that has never agreed the vault exists.

The diligence question set is short and decisive. One: does the vendor's BAA or a Part 2 addendum explicitly address 42 CFR Part 2 records, by name, with redisclosure-restriction language? Two: does the vendor's subprocessor map, the model providers and cloud infrastructure the data actually flows through, hold at a tier where those commitments are real, or does PHI route to a model provider that signs no BAA at the vendor's pricing tier? Three: can the vendor technically segment, meaning the practice can designate which clinicians, programs, or note types the tool serves, so the SUD track can be excluded or included deliberately? A vendor that answers question one with "we're HIPAA compliant" has answered no. This is the same diligence posture the program's vendor RFI lesson taught; Part 2 adds the named-regulation requirement on top.

The resulting architecture takes one of three shapes. Shape one, single vendor, full coverage: one scribe whose BAA explicitly covers Part 2, used practice-wide; simplest, if such a vendor passes diligence. Shape two, two vendors, hard split: a Part 2-cleared tool for the SUD track, a general tool for everything else, with the EHR and workflow routing each session to the right tool; more operational friction, cleaner legal lines. Shape three, one vendor, partial deployment: the general tool is simply not used for the SUD track at all, and SUD-track clinicians document manually or with EHR-native tooling covered by a Part 2-aware agreement; the honest fallback when no cleared vendor fits. What the architecture may never be is the default Jordan stumbled into: one generic-BAA vendor, all clinicians, no routing, with Part 2 records flowing through it by inertia.

Making Segmentation Real: Chart, Workflow, and Training

Vendor segmentation fails if chart segmentation does not exist underneath it. The second vault has to be visible in the EHR: a program flag, a separate chart section, or a record-type designation that marks Part 2 content so every downstream workflow can behave differently when it encounters it. Without that marker, the front desk cannot route the right intake packet, the clinician cannot know which scribe to launch, the ROI workflow cannot attach the redisclosure block, and the release-processing staff cannot distinguish a request that needs Part 2 consent review from one that needs ordinary HIPAA processing. The marker is the keystone; everything else hangs from it.

Workflow follows the marker. At intake, the program assignment decision, general track or SUD track, triggers the right consent packet, the single-consent instrument for Part 2 clients, and the right chart configuration. At the session, the clinician's tooling follows the track: the SUD-track session uses the Part 2-cleared documentation path; the general session uses the general path. At disclosure, every outbound document generated from a Part 2 chart carries the redisclosure language by template, and the AI prompt for coordination letters includes a standing instruction: if the source chart is flagged Part 2, include the federal confidentiality notice block and scope content to minimum necessary. At audit, the practice can demonstrate the boundary: which records are in the vault, which vendors may enter, which consents are on file.

Training closes the loop, because the most common Part 2 breach in an AI-equipped practice will be a clinician convenience decision: the SUD-track clinician who likes the general scribe better and uses it "just this once." The training message is the cardinal rule of this chapter applied to Part 2: the tool boundary is a legal boundary, not a preference. The supervision agreements of any associates working the SUD track state the boundary explicitly, because the supervisor's license sits behind the associate's tooling choices, and a Part 2 violation by an unsupervised tooling shortcut lands on both of them.

What AI Does Inside the Vault, and What It Never Does

Inside a properly segmented architecture, AI earns its keep in the SUD program the same way it does everywhere else in this curriculum. The Part 2-cleared scribe drafts the weekly individual 90834 note and the group 90853 note from sessions the clinician conducted. AI drafts the MAT coordination letter to the PCP with the redisclosure block and minimum-necessary scoping baked into the template. AI drafts the discharge summary and the continuity-of-care letter with Part 2 awareness. AI helps the compliance officer draft the practice's own Part 2 policies and the consent instrument for counsel's review. Every one of those outputs is verified and signed by the responsible human, because the signature remains a legal attestation in the second vault exactly as it is outside it.

What AI never does: AI does not make the holds-itself-out determination. That is a legal judgment about the practice's public posture with regulatory consequences, made by the owner with counsel, informed by the inventory this lesson's artifact assembles. AI also never decides which records are Part 2 records; the chart marker, set by humans through the program-assignment workflow, decides that, and the AI behaves according to the flag, never according to its own inference from content. A model that reads an AUDIT-C score in a general-track chart and starts appending redisclosure blocks is misclassifying records, and a model that fails to append the block on a Part 2 chart because the session focused on anxiety is breaching. Classification by inference is precisely the failure mode segmentation exists to prevent. And, as everywhere in this program, AI never makes clinical determinations: not the SUD diagnosis, not the level-of-care decision, not the risk call when the relapse disclosure comes with suicidal ideation attached.

One more boundary deserves naming because the 2024 final rule preserved it with emphasis: the protections against using Part 2 records in legal proceedings against the patient endure. The record in the second vault is shielded in ways an ordinary HIPAA record is not, and any AI-assisted records-production workflow, the subpoena-response patterns taught in the court-and-forensic chapter, must treat Part 2 records as a category that does not leave the vault for litigation purposes without the specific authorizations that pathway requires.

The Applied Problem: The Holds-Itself-Out Decision Memo and Segmentation Architecture

Your artifact has two parts: a Holds-Itself-Out Decision Memo and the Segmentation Architecture that follows from it. Start the memo with the public-posture inventory. Pull the actual evidence: the services page text, every clinician bio mentioning SUD credentials or specialties, every insurance-directory listing, the referral relationships and how they are described, the billing history for SUD-specific codes such as H0015, and any program track with SUD-specific intake or marketing. AI accelerates this inventory legitimately: prompt a capable model with "Review the following website copy, directory listings, and bio text. Flag every statement that could support a finding that this practice holds itself out as providing substance use disorder diagnosis, treatment, or referral for treatment under 42 CFR Part 2. Quote each flagged statement verbatim and explain why it is relevant. Do not conclude whether the practice is a Part 2 program."

That final instruction is the hard limit in prompt form. The model assembles the evidence; the owner and counsel make the determination, and the memo records it in three parts: the evidence inventory, the determination (whole practice, specific program unit, or not a Part 2 program), and the rationale. If the determination is "not a Part 2 program," the memo still matters: it documents the analysis was done, and it triggers a review obligation whenever the public posture changes, a new service line, a new bio, a new directory listing.

If the determination is yes for any unit, build the segmentation architecture as a one-page diagram plus a vendor table. The diagram shows the two vaults: which programs and record types are Part 2, the EHR marker that designates them, the consent instrument each track's intake uses, and the disclosure workflow with the redisclosure block. The vendor table lists every AI tool in the stack with four columns: tool, BAA status, explicit Part 2 coverage yes or no with the contract section cited, and deployment scope (general track only, SUD track cleared, or excluded). Any tool whose Part 2 column says no and whose deployment scope says anything other than "general track only" is a finding to fix this week, not this quarter.

The verification pass: counsel reviews the determination and the consent instrument; the compliance officer tests the workflow by running a fictional SUD-track intake end to end, checking that the right packet, the right chart flag, the right scribe, and the right letter template all fire; and one real coordination letter draft is inspected for the redisclosure block and minimum-necessary scoping before the pattern goes live. Done looks like a memo counsel has initialed, a diagram on the compliance wall next to the multi-state matrix from the last lesson, and a vendor table with zero rows in violation.

Key Takeaways

• 42 CFR Part 2 is a second vault inside HIPAA: a stricter federal confidentiality regime for records created by a federally assisted program that holds itself out as providing SUD diagnosis, treatment, or referral. The 2024 final rule aligned Part 2 with HIPAA, added a single-consent option for treatment, payment, and operations, and revised redisclosure, but it did not dissolve the boundary or the record class.
• The holds-itself-out analysis turns on public posture, not intent or diagnosis codes: services pages, clinician bios, directory listings, referral marketing, and SUD-specific billing like H0015 decide it. A general practice treating co-occurring SUD incidentally is not automatically a Part 2 program, and over-applying Part 2 burdens every record and coordination letter with machinery the law never required.
• The analysis runs at the unit level: a group practice can contain a Part 2 program, the SUD track, without being one entirely. That unit-level reading is what makes segmentation both possible and mandatory, with the second vault's records marked in the EHR so every downstream workflow can behave differently.
• Vendor segmentation is the tool-level rule: the AI vendor handling general therapy notes is not the vendor handling Part 2 SUD notes unless the BAA explicitly covers Part 2 by name, with redisclosure-restriction language and a subprocessor chain that holds. "We're HIPAA compliant" is a no.
• Every disclosure from the Part 2 chart, the MAT coordination letter to a non-Part 2 PCP above all, carries the consent on file, minimum-necessary scoping, and the redisclosure notice block, and the AI templates that draft those letters include the block by default based on the chart flag, never on the model's inference from content.
• The hard limits: AI does not make the holds-itself-out determination, does not classify which records are Part 2 by inference, and does not make any clinical determination inside the vault. Humans set the flag, counsel signs the memo, and the clinician's signature remains a legal attestation. The 2024 rule's enduring protections against use of Part 2 records in proceedings against the patient also constrain every AI-assisted records-production workflow.
• The deliverable is the two-part artifact: a Holds-Itself-Out Decision Memo with verbatim evidence, a human determination, and a change-trigger review obligation; plus a Segmentation Architecture diagram and vendor table where any tool lacking explicit Part 2 coverage but touching SUD-track records is a fix-this-week finding.