Five Places to Prioritize When Reviewing an AI-Written Clinical Protocol

An AI-generated protocol may have every section and read fluently while still needing review of content that affects participants, execution and interpretation. When time is limited, begin with five areas and expand according to the study’s risks. This is a prioritization method, not a substitute for a complete protocol review.

First: the connection between the research question and the intended conclusion.

Can the objective be stated clearly? Does the primary endpoint match it? Do the population, treatments and outcome answer that question? If the objective promises a long-term benefit while the study measures a shorter-term indicator, the team needs to explain the relationship and limits. AI should not promote a measurable indicator into an unestablished clinical benefit.

In the analysis section, check how treatment discontinuation, subsequent therapies and missing assessments are handled. ICH E9(R1) provides a framework for defining the treatment effect. Clinical and statistical expertise are needed; repeating an endpoint name across sections does not establish alignment.

Second: population definitions and their logic.

Cross-check inclusion, exclusion, screening and group-specific requirements. Verify age, thresholds, treatment lines, methods, time windows, AND/OR conditions and exceptions against the original source. Simplifying a criterion can change whom the study concerns. Do not fill evidence gaps with model-generated “usual” criteria.

Ask whether the required records and tests can actually be obtained. A grammatically correct definition may still be difficult to confirm consistently in local practice. Operations staff and study sites can help test feasibility.

Third: treatment and comparator arrangements.

Review groups, doses, frequency, combinations, permitted and prohibited medicines, dose adjustments and stopping rules against the clinical and pharmaceutical basis. AI may import familiar arrangements from another indication or version. Medically consequential doses and procedures require supporting material and responsible professional confirmation.

An editorial working document can identify an item for clinical and pharmaceutical confirmation. A protocol ready for execution cannot contain unresolved gaps without ownership. Maintain a draft-assumption list showing what is not yet eligible for the final text.

Fourth: endpoints and the assessment schedule.

Compare definitions, starting points, assessment frequency, methods, visit windows and data collection after early treatment discontinuation across the text and schedule. If an outcome depends on an assessment absent from the schedule, the planned result lacks an operational foundation.

Use a traceability table linking each endpoint to required information, collection times, owners and sections. It can reveal omissions and unnecessary assessment burdens. AI may locate content and differences, but professionals need to judge whether additional procedures are appropriate.

Fifth: safety actions and their connection to analysis data.

When adverse events, unscheduled care, discontinuation or withdrawal occur, specify who acts, what is recorded, when it is reported and how relevant information enters analysis. Safety responsibilities should connect to actual study procedures rather than remain generic language.

The ICH M11 guideline, template and specifications provide a common protocol structure. Using the template does not establish that the design is sound. Cross-section review is still required after AI fills the document.

Record the location, impact, responsible function and evidence needed to close each issue. Closure means affected content has been checked, not merely rewritten more smoothly. Protect these five connections, then complete the remaining reviews to produce an executable and interpretable protocol.