Human-in-the-loop review

The person needs decision authority

Assign the review to someone whose competence, credential, licensure, supervision, payer status, and organizational role cover the decision. A biller cannot approve clinical reasoning. A clinician should not certify a security control. An owner does not gain clinical authority through ownership.

For covered BCBA and BCaBA certificants and applicants, the BACB Ethics Code keeps competence, documentation, assessment, intervention, consent and assent when applicable, and continual evaluation with the professional.

Review must occur before the consequence

Define the decision point. Review after a note is signed, claim is sent, family is contacted, plan is changed, or access is denied may be an audit rather than a control.

High-consequence outputs should remain in a hold state until the qualified reviewer acts. The system should show what will happen, which data will be used, and how to defer when evidence is missing.

Source access makes disagreement possible

Show the observation, measurement, report, policy, authorization, record, or other evidence behind the suggestion. Identify AI-originated content, model or service version, data time, uncertainty, and known limitations. Preserve conflicting evidence.

ASTP/ONC describes useful clinical decision support as timely, person-specific information that is clear, organized, and fit for workflow. A polished answer without traceable evidence weakens review.

Independent judgment needs practical support

When feasible, ask the reviewer to inspect key evidence or record an initial view before revealing a high-stakes recommendation. Avoid defaults that imply approval, countdowns that punish careful review, and production goals that reward agreement.

Provide enough time, manageable queue volume, accessible design, and a manual path. AHRQ's Human-AI Interaction brief links automation overreliance with workload and time pressure. Human factors belong in the control design.

Test whether the reviewer can act

Use representative fictional cases with correct suggestions, wrong suggestions, missing suggestions, conflicting sources, and incomplete evidence. Ask reviewers to locate the source, explain their initial judgment, change the output, defer the case, and escalate a concern. Observe time, errors, and interface barriers.

Include an unavailable supervisor, a busy shift, a vendor outage, and a changed model version. Verify that reviewers know the manual route and that using it does not erase evidence or create an unsafe delay. Training completion measures exposure to the material. A release test measures whether the control works under declared conditions. Production monitoring shows whether those conditions continue.

Retest after each material model, workflow, or staffing change. Log the result.

Make override and deferral usable

Reviewers should be able to edit, reject, request more information, escalate, and document a different decision. Capture a meaningful reason without forcing a false choice. Protect reviewers from retaliation for appropriate disagreement.

Repeated overrides can indicate bad thresholds, stale rules, missing data, poor interface design, or a changed population. Repeated agreement can indicate strong fit or automatic acceptance. Sample the source evidence before interpreting either pattern.

A fictional review cohort

Ren's practice locks 30 fictional clinical-AI decisions due for review. Twenty-seven have a qualified reviewer, complete source evidence, visible AI provenance, enough review time, and a recorded final decision: 27 of 30, or 90%.

Of those 27, reviewers agree with 18, change six, and defer three. Report 18 of 27, 6 of 27, and 3 of 27 separately. Three incomplete reviews stay in the original cohort. These figures describe workflow behavior, not accuracy or clinical benefit.

Preserve authorship and audit evidence

Record the input version, output, source evidence, reviewer, final decision, changes, reason, date, time, and downstream action. Preserve correction history. The final author should understand and stand behind the final content.

Audit logs need interpretation. A timestamp can prove that a button was clicked while revealing nothing about evidence access or reasoning. Pair system events with workflow testing and sampled records.

Monitor the control in production

Track decisions reviewed by target divided by decisions due, incomplete source packages, overrides by reason, deferrals, corrected agreements, missed errors, incidents, queue age, and review time. Include cases where the system stayed silent.

The CASP AI practice parameters address organizational selection, deployment, monitoring, change management, and auditing in ABA. The voluntary NIST AI Risk Management Framework organizes governance, context, measurement, and management. Neither transfers clinical authority to the system.

Pause or narrow use when workload prevents review, source data is unavailable, model behavior changes, errors cluster, or the manual path fails. Revalidate after material model, data, interface, law, population, or workflow changes.

Give one named owner authority to release, restrict, or stop the workflow after review. If clinical, privacy, security, and operational evidence point in different directions, keep the consequential action held and document who resolves the conflict.