AI-assisted documentation

Assistance can support several tasks

An approved tool might transcribe a recorded encounter, summarize supplied text, suggest a structure, improve readability, or flag an incomplete field. Each is a separate use case with different data, error modes, and review needs.

Define the task narrowly. “Write the note” hides who observed the service, which sources are allowed, which claims need verification, and who holds authorship.

Preserve the source of every statement

The reviewer should be able to trace proposed content to a direct observation, contemporaneous measurement, client or caregiver report, prior record, plan, authorization, or other authorized source. Label reported information and uncertainty.

AI cannot witness a session, infer that a target occurred from a template, invent a quote, fill a missing duration, or create consent. It should not convert an absent fact into fluent prose. Keep the source record and the AI-assisted draft distinct.

Set permitted and prohibited uses

A written use boundary can permit restructuring clinician-authored text, identifying missing required sections, or drafting a plain-language summary from approved sources. It can prohibit generating direct-observation claims, changing measurements, inventing client speech, deciding medical necessity, copying another client's content, or signing a record.

State whether audio capture, ambient listening, external retrieval, translation, or model feedback is allowed. Give each feature its own consent, privacy, security, accuracy, and accessibility analysis. A tool approved for de-identified training exercises may remain unsuitable for live notes. Train users with examples of acceptable edits, unsafe completion, uncertain attribution, and the correct escalation path. Keep a manual documentation route available.

For BCBA and BCaBA certificants and applicants, the current BACB Ethics Code addresses competence, confidentiality, documentation, consent and assent when applicable, accuracy, and accountability. Software does not acquire those responsibilities.

Give the clinician real review authority

Before signing, the qualified author needs time and access to compare the draft with source evidence. The interface should make AI involvement visible and let the author edit, reject, or start over. Preserve the final author, date, time, source, and correction history required by policy.

Review clinical facts, counts, denominators, dates, times, locations, participants, goals, prompts, client communication, risk, medical information, payer language, and internal consistency. A spell check is not a clinical review.

Control data and vendors

Start with entity status and the data path. Current HHS Security Rule guidance applies to covered entities and business associates handling ePHI. Map prompts, audio, notes, retrieved records, outputs, logs, feedback, support access, training use, subprocessors, retention, and deletion.

If a cloud vendor creates, receives, maintains, or transmits ePHI for a covered entity or business associate, HHS cloud guidance says business-associate status can apply even when the vendor lacks a decryption key. Confirm the applicable BAA, permitted uses, configuration, and each party's duties.

Use only the data required for the approved task. Avoid copying an entire record when a narrow excerpt will do. Separate fictional testing from live care and remove unsupported browser extensions or personal accounts.

A fictional draft review

Ana's practice reviews 20 fictional AI-assisted notes. Eighteen have complete source links, visible AI provenance, qualified authors, and required fields: 18 of 20, or 90%. Two remain incomplete and stay in the original cohort.

After source-by-source clinical review, 16 of the 18 complete drafts are usable with corrections: 16 of 18, or 88.9%. Original-cohort usable yield is 16 of 20, or 80%. These measures describe the review process, not note quality, time savings, safety, or clinical benefit.

Include the person affected

AHRQ's AI implementation brief suggests considering disclosure when AI generates documentation or influences decisions and notes that patient review can reveal errors. It is broad healthcare guidance rather than an ABA consent rule.

Use understandable communication and follow the applicable consent, assent, recording, access, and correction process. Keep augmentative and alternative communication available. A client or caregiver should have a route to question an attributed statement without having to understand the model.

Monitor errors in production

Sample a declared due cohort. Track unsupported statements, attribution errors, omitted facts, incorrect numbers, reviewer changes, rejected drafts, correction requests, incidents, and subgroup or workflow patterns. Keep incomplete and unreviewed drafts visible.

The CASP AI practice-parameters page addresses organizational selection, deployment, monitoring, change management, and auditing in ABA. Treat it as sector guidance. Pause or narrow the use when evidence, vendor terms, model behavior, integration, law, or workflow changes.

Before signature, require a source-by-source reconciliation that identifies AI-added language, reviewer changes, unresolved uncertainty, and the qualified author. Reject the draft when observation, time, quotation, consent, measurement, or clinical reasoning lacks attributable evidence. Sample signed notes after release and route any unsupported content through the practice's correction and incident processes.