AI-native, without giving up Part 11 discipline
AI is limited to a preparatory task: drafting study configuration. Captured data, clinical judgment, approval and signature remain outside its boundary.
The defensible design is a small, testable surface. The assistant receives an explicit design brief and selected historical form configurations only: field names, types, choices, units, sections and visit structure.
1 — Reuse configuration, never results
Each historical study can become an immutable configuration template after independent review and approval. The extractor uses an allowlist and has no query path to CRF values, subjects, queries, safety events, attachments, signatures or audit records.
2 — Generate a bounded draft
The provider returns only a bounded schema of forms, fields and visits. The target study must still be in setup with zero subjects; output is stored as a hashed generation record and is never applied automatically.
3 — Human review and normal validation
A data manager edits and applies the draft. The resulting configuration follows the same versioning, independent approval, UAT/PQ and activation gates as a fully human-designed study.
The line that never moves
AI cannot sign, activate a study, make eligibility or safety decisions, or receive captured data. Provider, model, prompt, schema or allowlist changes are controlled software changes and require impact assessment and regression evidence.