Compliance & the managing organization's responsibility

The complete regulatory frame the AI-native platform is built around — rule by rule, with what the software enforces mapped against what remains, explicitly and unavoidably, the responsibility of the organization that runs it. Because the platform was designed for AI use of protected data from the first line, the compliance posture is architecture, not policy wrapped around an add-on: the gates, the segmentation and the audit chain are where these rules live. This page is documentation of design intent, written so counsel and compliance officers can do their own assessment. It is not legal advice.

1. The division of responsibility, in one table

2. HIPAA Privacy Rule (45 CFR Part 164, Subpart E)

2.1 Permitted uses and purpose of use (§164.506)

The Privacy Rule permits use and disclosure without authorization for treatment, payment and health care operations (TPO), and in defined further cases. The platform makes the claimed basis explicit: every session carries a purpose of use — Treatment, Payment, Health care operations, Patient request, or Legal — chosen in the header, and every audit event stores it. This does not make a use permissible; it makes the claimed basis reviewable, which is the precondition for enforcing anything.

2.2 Minimum necessary (§164.502(b), §164.514(d))

The rule requires reasonable efforts to limit PHI to the minimum necessary for the purpose. The platform implements this at four layers:

• Role design: permissions are enumerated per role, and the deliberate asymmetries carry the principle — the Administrator cannot open a chart, the Auditor holds only audit access, Psychotherapy grants nothing beyond the psychotherapy store.
• Tool design: the assistant's tools are scoped the same way, so the AI cannot read past the person it serves.
• Sweep confirmation: a question requiring a population-wide read pauses for an explicit confirmation stating what would be read.
• Withholding with counts: what a role may not see is withheld and the withholding stated — visible minimization, not silent truncation.

2.3 Psychotherapy notes (§164.501, §164.508(a)(2))

HIPAA defines psychotherapy notes as a clinician's notes documenting or analyzing conversation in a counseling session, kept separate from the rest of the record — and the definition explicitly excludes medication management, session start/stop times, modalities, results, diagnoses, functional status, treatment plan and progress to date (those live in the ordinary record). Almost every use or disclosure of psychotherapy notes requires the patient's specific authorization, separate from any general consent.

The platform mirrors the structure physically: psychotherapy notes are a separate store, reachable only by the Psychotherapy role, excluded from every export path, subject to a minimum-necessary confirmation even for that role, and every listing and read is an individually attributable audit event (psychotherapy.list / psychotherapy.disclosure).

2.4 Accounting of disclosures (§164.528)

An individual may request an accounting of certain disclosures for the six years prior. An accounting entry requires: date, recipient, a brief description of what was disclosed, and the purpose. The audit trail stores each of these — at, the destination or actor, resource_ref, purpose — for disclosure-family events (phi.disclosure, psychotherapy.disclosure, roi.fulfilled, export.completed), giving the organization the raw material; assembling and delivering the accounting remains the organization's workflow.

2.5 Individual rights (§164.524, §164.526)

Access and amendment are the organization's processes; the platform contributes the complete record inventory (the Data Store screen), the ROI queue for fulfilling access requests, and the disclosure log. Amendment of the platform's copies follows amendment of the legal record in the EMR.

3. HIPAA Security Rule (45 CFR Part 164, Subpart C)

The Security Rule's structure is administrative, physical and technical safeguards. The platform is a technical-safeguards instrument; the table states the mapping honestly:

3.1 Breach notification (45 CFR Part 164, Subpart D)

A breach is an impermissible use or disclosure that compromises PHI, presumed reportable unless a documented risk assessment shows a low probability of compromise — weighing (1) what was involved, (2) who received it, (3) whether it was actually acquired or viewed, and (4) mitigation. Deadlines: affected individuals within 60 days of discovery; HHS within 60 days (or annually for breaches under 500 individuals); prominent media for breaches over 500 in a jurisdiction. The platform's contribution is the forensic record — the trail answers "what did this account touch, when, under what claimed purpose" in one query, and access.denied patterns give detection a surface. The response program, the risk assessments and the notifications are the organization's.

4. 42 CFR Part 2 — substance use disorder records

4.1 Scope

Part 2 protects records of the identity, diagnosis, prognosis or treatment of any patient maintained in connection with a federally assisted SUD program. Whether a given unit is a "Part 2 program" is a legal determination the organization makes; the platform's posture is protective — records classified sud_part2 get Part 2 handling whether or not the determination is ultimately required.

4.2 Consent and redisclosure

A Part 2 consent must name the patient, the discloser, the recipient, the purpose, what is to be disclosed, an expiration, and be signed. Disclosures must be accompanied by the redisclosure notice — the federal rule's prohibition on redisclosing without consent. The platform enforces the mechanical half: SUD-classified content is withheld from ordinary retrieval, and any export whose record set contains Part 2 content refuses unless Part 2-specific consent is recorded for that disclosure. The refusal is a first-class outcome in the export history, where an auditor will find it recorded with its reason. Obtaining compliant consents and attaching the statutory notice to fulfilled disclosures is the organization's.

4.3 The 2024 alignment

The 2024 Part 2 final rule aligned enforcement with HIPAA (civil and criminal penalties, breach notification applied to Part 2 records) and permitted TPO redisclosure under a single consent in defined circumstances — while keeping SUD counseling notes under a psychotherapy-notes-like separate consent. The platform's separate-store and specific-consent design already matches the strictest reading; the organization's counsel decides where the newer flexibilities are taken.

5. Information blocking (21st Century Cures Act)

The info-blocking rule cuts the other way from everything above: a practice likely to interfere with access, exchange or use of electronic health information is prohibited unless an exception applies (privacy, security, infeasibility, among others). A governance platform must therefore be able to show that every withholding has a lawful basis. The platform's design supports exactly that showing: every withholding is rule-based (a named category, a named consent state, a named permission), recorded, and visible — never an unexplained absence. Evaluating each configured withholding against the exceptions, and not configuring withholdings without a basis, is the organization's obligation.

6. State law

Where state rules are stricter than HIPAA, the stricter rule wins. Concrete examples the design anticipates:

• Recording consent: ambient documentation records a conversation, and states split between one-party and all-party consent — with telehealth potentially crossing state lines. The ambient consent gate is keyed on jurisdiction and modality, denies by default, supports visit-level attestation, and audits every evaluation (consent.evaluated).
• Audio is PHI: a visit recording carries voice — one of the eighteen HIPAA identifiers — and its content. The ambient pipeline treats audio like any record: encrypted under the deployment's keys, role-gated, every transcription an audited model interaction, and raw-audio retention a governed setting (transcribe, then dispose on the configured schedule — indefinite voice archives are liability without clinical payoff). The speech model itself is self-hosted, so no third party touches the audio and no ASR vendor BAA exists; the drafting model runs under the organization's existing AI-provider agreement.
• Condition-specific statutes: many states require specific consent for HIV disclosure; several have reproductive-health shield laws restricting out-of-state disclosure. The hiv and reproductive sensitivity classes exist so those rules have something to attach to.
• Minors and retention: minor-consent services and state retention schedules vary; the retention screen enforces schedules the organization configures.

The platform ships the enforcement points; the organization configures the jurisdictions it operates in and keeps the rule table current with counsel.

7. FDA — clinical decision support

Under section 520(o)(1)(E) of the FD&C Act and the FDA's CDS guidance, software avoids device status when it (1) does not acquire or process signals from devices, (2) displays or analyzes medical information of the kind otherwise communicated between professionals, (3) provides recommendations — not directives — to a professional, and (4) enables that professional to independently review the basis, so they need not rely primarily on the software. The platform is engineered to the four criteria: it processes records, not device signals; its outputs are drafts and answers of the kind clinicians exchange; it issues no diagnostic or treatment directive; and the citation discipline exists precisely so the basis of every answer is independently reviewable. Deployment choices can forfeit this posture — e.g., surfacing outputs as alarms clinicians act on without review — and keeping inside it is an organizational duty.

8. AI-specific commitments

• the signature rule — human release. No generated content reaches the record or leaves the organization without a human signature. Structural: drafts have no write path except through signature.committed.
• counts, never lists. Population queries return counts, never named lists, on the de-identified plane. Structural: the cohort tool cannot enumerate.
• Grounding. Answers only from retrieved records; every claim cited; absence never asserted from unread documents; withholdings stated.
• Registry & preflight. Every AI capability is registered with a declared scope before use, and preflight checks the declaration against what a capability is about to do.
• Fairness. Predictive models carry subgroup performance on the Fairness screen. The screen is an instrument, not a conclusion — it shows where a model underperforms rather than averaging it away.
• No training on PHI. Excerpts sent to the model API are used to answer the question, not to train models, under the organization's agreement with the provider.

9. The managing organization's responsibilities, in full

The software is a tool. The organization that deploys it — the covered entity or business associate — holds the legal duties. PHI AI enforces the policy it is configured to enforce; it cannot discharge any of the following on the organization's behalf:

1. Security risk analysis and management (§164.308(a)(1)). A documented, periodic assessment of the entire environment the platform runs in — hosting, network, endpoints, workforce — with a managed remediation plan. Deploying this software is an input to that assessment, never a substitute for it.
2. Business associate agreements. Executed BAAs with the platform operator, the model provider (for PHI reaching the model API), the hosting provider, and every downstream recipient handling PHI on the organization's behalf — with the model provider's data-use and no-training terms verified against what the AI commitments above assume.
3. Access governance — in your identity source of truth, not here. This system is not the source of truth for users or roles: accounts, role assignments, recertification and deprovisioning are governed by the organization's identity provider and directory, and the platform enforces what that source asserts. Deciding who holds which role, reviewing grants on a schedule, and removing them the day they are no longer needed. The platform enforces the grants and records their use; the organization owns their correctness. The permission matrix in the system documentation is the review artifact.
4. Audit review. The hash chain preserves the record; someone must read it — on a defined cadence, with defined escalation for access.denied patterns, system.* overrides, and export refusals.
5. Consent capture. Obtaining and recording the consents the gates evaluate: Part 2 consents, psychotherapy authorizations, recording consent, data-sharing agreements. The platform evaluates consent state; it cannot create it.
6. Clinical accountability. Every signed note is the signer's. AI drafting does not move responsibility for clinical content off the licensed professional who signs it, and the organization's medical staff policies should say so explicitly.
7. Incident response and breach notification. Detection, containment, the four-factor risk assessment, and notification within the deadlines in §3.1. The audit trail is evidence; the program is the organization's.
8. Patient rights operations. Access, amendment, restriction and accounting requests, run on the organization's timelines, fed by the platform's record inventory and disclosure log.
9. Retention and disposition. Setting the schedules the Retention screen enforces — state and federal — and making the disposition decisions the Disposition role exists for.
10. AI oversight. Validating the assistant against the organization's own data before clinical use; monitoring for drift and subgroup degradation; maintaining an oversight structure with authority to pull the switches this platform deliberately provides (the Control panel's retrieval and model switches); and treating AI output incidents with the same discipline as any other patient-safety event.

10. The auditor's shortcut

Evidence for the common asks, and where it lives:

Nothing on this page is legal advice; it documents the platform's design intent so the organization's counsel and compliance officers can perform their own assessment against their own facts.