Using AI safely in clinical data capture: a GxP checklist

Clinical data is only usable if it is attributable, accurate and traceable, the ALCOA+ principles. AI puts pressure on each of them. A suggestion accepted without review has no clear author. A model update can change outputs between two sites or two months. A prompt can carry personal data to a place your data protection assessment never covered.

It helps to separate two places AI can sit. Build-time AI helps set up the study: drafting visit schedules, forms, edit checks or lab tables from a protocol. A person reviews the output, then it goes through your normal approval and testing. Data-time AI acts on participant data: auto-coding adverse events, flagging outliers, suggesting query text or deciding eligibility. Build-time errors are usually caught in review and UAT. Data-time errors can flow straight into the analysis, so they need far more evidence.

FDA's draft guidance gives a usable method even where it does not strictly apply. Define the question the AI helps answer and its context of use: what it does and what else informs the decision. Then rate model risk as a combination of model influence (how much the output drives the decision compared with other evidence) and decision consequence (how bad a wrong decision would be). The guidance's own clinical example, an AI model that alone decides which participants skip inpatient monitoring after dosing, is high influence and high consequence, so high risk.

Apply the same lens to data capture. An AI that drafts eCRF questions which a data manager reviews, edits and then tests in UAT has low influence, because the human review and testing decide what goes live. An AI that auto-codes adverse events into the safety tables has high influence over a consequential output. The depth of validation, monitoring and documentation should follow that rating. The draft guidance also excludes AI used for operational efficiencies that do not affect participant safety, product quality or the reliability of study results. Human review is often what keeps a build-time tool on the right side of that line, so document it.

Capture uses AI at build time only. The AI study builder reads a protocol (PDF, DOCX or DOC) and drafts the visit schedule and forms. On a single form, it can read a description or an uploaded file (CSV, XLSX, PDF or DOCX) and draft questions or a lab table. Nothing is saved without human review, and the AI only works on draft forms, so it cannot change an approved form that is live. Reviewed drafts then follow the normal lifecycle: forms move from draft to approved, approved forms are locked, and only approved forms appear in the casebook. Participant data entered later carries the usual field-level audit trail. See protocol to study setup for the full flow.

That covers checklist items 3 and 4 by design: the review step is part of the product and the output is testable in UAT. Items 1, 2 and 9 are yours to write. For items 6, 7 and 8, ask us during your supplier assessment, as you would ask any vendor with AI features. The computer system validation guide shows where these answers sit in your validation file.