To govern a new AI feature added by an ABA software vendor, keep it disabled until the practice identifies its purpose, users, data, model and subprocessor chain, output, affected decisions, human-review boundary, failure modes, contract terms, evaluation evidence, accessibility, monitoring, incident route, and rollback. Test the deployed version with representative and hard cases, preserve every failed result, and assign qualified approval for each allowed use.

How to govern new AI features added by an ABA software vendor

Farah treats a new model, prompt path, suggestion, summary, prediction, transcription, coding aid, or automation as a material product change. A feature already covered by the vendor's commercial agreement can still change data use, risk, clinical workflow, accessibility, billing, or professional responsibility.

Build the AI feature change and evidence register

The register captures feature ID; product, version and release date; default and disablement; stated purpose; allowed and prohibited uses; user and person affected; input, prompt, context and output data; model, host, subprocessor and location; retention, training and secondary use; affected decision; human reviewer and required actions; authority; evaluation cohort; error taxonomy; bias and subgroup analysis; accessibility; confidence and uncertainty display; monitoring; incident; terms change; rollback; owner; approval; and expiry. Structured fields support routing, comparison, evidence expiry, monitoring, alerts, and validation. Narrative preserves clinical reasoning, client and family experience, accessibility, uncertainty, disagreement, legal deferral, source limits, and why an accountable owner approved, restricted, repaired, deferred, or rejected the item.

Run Farah's workflow

Farah compares release notes, settings, contracts, BAAs, privacy terms, subprocessor lists, architecture, and actual product behavior. She locks representative fictional cases and high-risk edge cases, records raw outputs, and separates helpfulness from factual, clinical, privacy, accessibility, and operational acceptance. Material model or configuration changes reopen the review.

Protect the new vendor AI feature governance boundary

AI may draft, classify, surface, or summarize within an approved route. It cannot hold licensure, obtain consent, author a clinician's judgment, accept privacy or security risk, determine payer coverage, or approve its own output. The accountable human must have time, source information, competence, and actual authority to review or reject the result.

Keep authority and evidence attributable

Farah assigns every clinical, privacy, security, accessibility, technical, records, financial, workforce, and operational decision to the qualified owner. Software and vendors can surface evidence, automate an approved step, or propose an action. They cannot grant professional authority, accept the practice's risk, replace client involvement, or approve their own control effectiveness.

Keep unknowns, workarounds, and failures visible

Farah records each unknown, assumption, exception, dependency, workaround, failed or skipped test, owner, deadline, escalation, and retest. Conditional approval states the exact scope, safeguard, operating restriction, evidence, expiry, and result if remediation misses its date. Raw failures stay in the denominator.

Work through Farah's fictional example

Farah reviews 22 fictional AI use cases. Fourteen initially have approved purpose, data route, terms, human boundary, evaluation, accessibility, monitoring, and rollback. Two outputs invent clinical facts, one suppresses AAC context, one coding suggestion lacks its source, one setting sends prompts to a new subprocessor, one opt-out resets, and two uses have no decision owner. Five repair. Three stay disabled. The scenario is synthetic and tests workflow and denominator logic. It establishes no clinical, privacy, security, accessibility, contract, payer, employment, records, financial, or legal conclusion for a real person, practice, product, or vendor.

Calculate Farah's measures honestly

Initial use-case readiness is 14 of 22, or 63.6%. Nineteen reach approved limited use or documented prohibition, or 19 of 22, or 86.4%. Features, versions, use cases, prompts, outputs, people, errors, and review decisions remain separate units.

Address the main new vendor AI feature governance risk

A feature can arrive enabled by default and spread into clinical, billing, or family workflows before the practice notices changed data use, output quality, or human-review demands.

Test Farah's control against hard cases

Farah tests missing context, contradictory record, rare presentation, AAC input, different language, adversarial instruction, unsupported coding, stale payer rule, privacy-sensitive prompt, wrong user role, model change, disablement, and rollback. Every case retains product and version, configuration, data, user, starting state, expected safeguard, observed result, defect, owner, retest, and disposition. Test passage applies only to the named configuration and conditions.

Run Farah's independent acceptance test

Farah gives an independent reviewer the feature map, controlling terms, data flow, locked cases, raw outputs, error analysis, approvals, monitoring, and disablement proof. An undocumented model participant, unsafe default, missing reviewer authority, or unreproducible evaluation fails.

Maintain the AI feature change and evidence register

Farah assigns a review cadence and change triggers for requirement, product, version, configuration, workflow, integration, vendor, subprocessor, data use, law, contract, incident, staffing, access, and ownership changes. This new vendor AI feature governance page remains draft until every named external review finishes.

Use organizational guidance within its public scope

Farah uses the CASP Organizational Guidelines public overview only for high-level business, clinical-operations, and risk-management context. CASP sells the detailed guidelines. The AI feature change and evidence register is this article's editorial operating model; CASP has not approved the specific workflow or technology.

Map vendor and cloud roles from actual functions

Current HHS Business Associates guidance classifies roles by functions and data relationships, including subcontractors and exceptions. HHS cloud guidance explains that a cloud provider handling ePHI for a regulated customer can be a business associate even when it holds encrypted data without the key. Farah records the actual role and agreement chain for the deployed system.

Keep the current Security Rule boundary visible

HHS risk-analysis guidance covers all ePHI a regulated entity creates, receives, maintains, or transmits. The current Security Rule page still identifies the January 2025 cybersecurity update as proposed as of August 19, 2026, so current eCFR text governs. The HHS guidance index provides current risk, remote-use, mobile-device, and ransomware resources. Farah labels proposals and readiness ideas separately from operative requirements.

Apply current administrative, technical, and documentation safeguards

Current 45 CFR 164.308 supplies administrative-safeguard duties, 45 CFR 164.312 supplies technical-safeguard duties, and 45 CFR 164.316 supplies policy, procedure, documentation, and specified six-year retention rules. Farah evaluates each applicable standard and implementation specification without claiming HIPAA requires one product, architecture, or control label.

Separate medical records, devices, and documentation retention

HHS states in its medical-record retention FAQ that HIPAA sets no general medical-record retention period. State and other sources often control those records, while HIPAA retains specified rule documentation. HHS's personal mobile-device page also explains that many personal-device health-data activities fall outside HIPAA's covered-entity and business-associate scope. Farah maps entity, data, device, and record status instead of applying one rule everywhere.

Review consumer-health and AI promises separately

The FTC Health Breach Notification Rule guidance requires its own entity and qualifying PHR analysis. FTC staff also tells AI companies to uphold privacy and confidentiality commitments, including promises about model training and undisclosed uses. Farah treats that post as enforcement-oriented staff guidance and checks other law, contracts, and settings independently.

Use voluntary frameworks as organizing aids

The NIST Cybersecurity Framework 2.0 organizes outcomes across Govern, Identify, Protect, Detect, Respond, and Recover. The NIST AI RMF page says AI RMF 1.0 is voluntary and being revised. NIST SP 800-34 Rev. 1 is final federal information-system contingency guidance that private practices may adapt. The OIG General Compliance Program Guidance is voluntary and nonbinding. Farah uses them to structure the AI feature evidence register; none creates a legal safe harbor.

Test accessibility and communication in the real workflow

Farah checks the DOJ Title III overview and web-accessibility guidance within their scopes. The ASHA AAC Practice Portal says AAC users should always have access to their communication tools. Testing covers real tasks, alternative channels, privacy, support, and the person's ability to ask questions, correct information, assent, dissent, and report a problem.

Related resources

Sources