How to Review an AI-Suggested ABA Plan Change

An AI-suggested ABA plan change is unapproved decision-support output. Keep it outside the controlling plan until an appropriately qualified clinician verifies the source data, provenance, assumptions, omissions, risks, access, and fit with the client's priorities and context. Use authorized systems and minimum necessary data practices when applicable. Record human authorship, client involvement, the disposition, version, release controls, monitoring, and any correction to the tool or workflow.

Classify the tool and use

Record vendor, model or version, intended purpose, user role, input sources, data destination, retention, approved environment, human-review boundary, and prohibited actions.

Verify the organization's current approval, contract, privacy, security, and clinical-use boundaries before entering client information. State whether the tool summarizes, flags inconsistency, drafts language, or recommends action, and what it must never do. Preserve the model and configuration used because deployed behavior can change. Do not treat marketing claims or a generic capability label as evidence of suitability for this case.

Verify provenance

Trace every material input to its client, definition, date, setting, author, version, correction history, missingness, and access condition. Recompute derived values independently when feasible.

Check that the tool received the intended plan version and did not merge unlike definitions, clients, or settings. Inspect omitted records as carefully as included ones. A confident suggestion based on stale or incomplete data remains unreliable. Preserve the prompt or structured inputs and output under the approved audit route, while minimizing data and avoiding unnecessary sensitive content.

Test the suggestion

Check unsupported claims, omitted alternatives, contradictory records, health and safety factors, communication access, cultural context, burden, bias, uncertainty, and effects outside the optimized metric.

Compare the suggestion with raw evidence, current client priorities, actual exposure, integrity, and unwanted effects. Actively search for reasons it could fail and for groups or contexts absent from the data. A numerical improvement can coexist with greater burden or lost AAC access. Route medical, legal, payer, privacy, and other specialized questions to their responsible reviewers instead of asking the model to settle them.

Preserve direct client involvement

Explain the proposed outcome accessibly, obtain the person's priorities and response, and keep consent, assent when applicable, dissent, discomfort, and correction visible.

Tell Sofia that software produced an unapproved suggestion and identify the human who will decide what, if anything, to propose. Give her a private accessible route to ask questions, reject the framing, or offer alternatives. Do not use the system's wording as evidence that she agreed or that the change is understandable. Preserve direct input separately from clinician interpretation and proxy reports.

Require accountable human decisions

A qualified clinician authors and approves any clinical change within scope. Software may surface evidence or inconsistencies; it should never approve, release, or silently rewrite care.

The clinician should record a fresh rationale that cites verified evidence, alternatives, risk, and Sofia's priorities rather than merely accepting the output. Any material change follows normal version, consent, training, distribution, first-use, and monitoring controls. Reject or hold suggestions that cannot be reconstructed. Keep the human decision and the tool output as linked but distinct records.

Monitor tool and clinical outcomes

Track accepted, modified, rejected, and unevaluable suggestions; data defects; overrides; adverse effects; client experience; plan integrity; and whether the tool's deployed behavior changed.

Use a mature cohort and preserve all suggestions in the denominator, including those never released. Monitor clinical outcomes separately from tool accuracy and workflow efficiency. Investigate repeated source, omission, or bias defects and pause the relevant use when safeguards fail. After model or prompt changes, rerun representative tests and reapprove the use boundary rather than assuming prior validation still applies.

Build Sofia's AI suggestion review record

Sofia's record treats an "AI suggested ABA plan change" as unapproved input with its own provenance. It retains the tool and model version, prompt or task, input cohort, definitions, missing data, access conditions, generated output, cited evidence, Sofia's direct response, privacy basis, risk review, alternatives, qualified human reviewer, accepted or rejected reasoning, consent and assent when applicable, any human-authored plan version, implementation result, review date, and open limits. A reviewer should be able to reproduce the suggestion and distinguish it from the clinician's decision.

Work through Sofia's example

A pilot tool suggests increasing Sofia's prompt level after detecting slower responses. Four of ten sessions used an outdated latency definition. Of the remaining six, two occurred during an AAC device outage, leaving four sessions with both the intended definition and access condition. All ten sessions remain visible in the review. The clinician rejects the automated recommendation, repairs the data mapping, restores access, and asks Sofia about response time before reconsidering the question. This fictional decision-support example sets no universal clinical recommendation, legal authority, coverage rule, release threshold, or outcome promise.

Address Sofia's main request risk

A polished explanation can conceal stale definitions, missing context, biased training data, fabricated sources, or unauthorized data use. Sofia's process treats output confidence as metadata, never as clinical authority. Separate an observation or request from the authority to decide, the evidence needed for a clinical conclusion, and the operational work required to release a version. A prompt response can still preserve careful review.

Choose Sofia's next action

The practice records the rejected suggestion, source defects, affected outputs, correction, retest, user notification, and tool limitation. A later clinical decision begins from verified evidence and Sofia's current input. Record accept, adapt, assess, refer, hold, decline, pause, or close with the responsible role, rationale, effective scope, due date, evidence required for closure, accessible communication, and next review. Software may route requests and preserve state. Qualified professionals make case-specific decisions within scope.

Protect Sofia's access and choice

Keep Sofia's AAC, interpreters, mobility, food, water, bathroom use, prescribed care, pain communication, health support, movement, rest, relationships, and emergency help available. Offer a private and accessible way to accept, decline, pause, withdraw when applicable, report discomfort, ask a question, and correct the record. Proxy and professional input can inform review without replacing Sofia's direct experience.

Apply current sources to Sofia's request

The BACB ethics hub, current Ethics Code, CASP public summary, and BCBA Test Content Outline provide scoped professional, client-involvement, evaluation, documentation, and training context.

An evidence-based ABA framework supports integrating research, expertise, client values, and context.

A treatment-integrity practitioner guide, the Essig review, research on integrity's relationship to intervention effects, and a reporting review support explicit definitions, implementation evidence, and cautious conclusions.

ASHA supports continuous AAC access.

Rehearse Sofia's request workflow

Test the AI suggestion review record with a client request, caregiver disagreement, payer deadline, school proposal, medical concern, safety incident, staff report, AI suggestion, reassessment result, unavailable AAC, privacy limit, missing source, uncertain authority, urgent pause, system outage, and late follow-up. Confirm that immediate action, attribution, qualified review, version control, communication, and closure remain correct.

Close Sofia's request review

Review the AI suggestion review record with Sofia, the responsible clinician, affected requestors, and the specialists named by the manifest. Preserve original requests, evidence, direct client input, authority, decisions, implementation, open findings, and limitations. Keep the page draft and noindex until required clinical, treatment-integrity, client or family, accessibility, interdisciplinary, payer, privacy, AI, and legal reviews are complete.