Building an Audit Trail for AI-Assisted SDS and Dossier Work

As soon as AI touches SDS review or dossier preparation, someone will ask whether your records still stand up to scrutiny. The answer depends less on the model and more on whether you preserved a reconstructable history of human decisions.

What an inspector or auditor needs to reconstruct

For a given released document or review decision, you should be able to show:

• Input documents and versions used
• What the AI proposed (if anything)
• What the human changed
• Who approved the final result
• When it became effective
• How superseded versions were controlled

If any of those links are missing, you have a story, not an audit trail.

Minimum viable trail for SMEs

You do not need a pharmaceutical-grade Part 11 system on day one for every internal workflow, but you do need discipline proportional to use:

1. Case ID for every review
2. Immutable storage of source SDS files used
3. Stored draft output (or hash + retained draft)
4. Reviewer identity and timestamp
5. Final approved artifact stored in the controlled repository
6. Reason for changes when AI draft was materially altered

Dangerous shortcuts

• Chatting with a public model and pasting conclusions into a controlled SDS with no record of prompts/sources
• Overwriting drafts so the AI suggestion disappears
• Shared generic logins for approval steps
• “Approved in meeting” with no link to the document version

These shortcuts create findings even when the science was fine.

Designing AI features for evidence

Prefer products that:

• Keep assistance inside an authenticated workspace
• Attach sources to findings
• Require explicit approve actions
• Export an evidence package per case

If a tool cannot export history, assume you will build manual workarounds, and budget for that pain.

Policy one-liner worth adopting

“AI may assist preparation; only named qualified persons release controlled compliance artifacts.”

Put that in your SOP. Train to it. Configure systems to enforce it. That single rule prevents most audit-trail failures associated with emerging AI use.