This is a ready-to-use checklist a qualified author or approver runs before AI-assisted content enters a regulatory submission. It is the last gate that confirms the content is grounded, every claim traces to a verified source, the high-risk specifics were checked, a named human authored and owns it, and the records and change controls are in place. Mark each item pass, fail, or N/A with a note, and do not release content with any open fail. Replace every <<FILL: ...>> placeholder with your own specifics. A filled specimen follows. Verify each cited regulation against the current source before you rely on it.
Checklist control
| Field | Entry |
|---|---|
| Checklist title | AI-Assisted Submission Content Review |
| Document number | <<FILL: CHK-ID, e.g. CHK-AIRA-002>> |
| Version | <<FILL: version>> |
| Content deliverable | <<FILL>> |
| Program / product | <<FILL>> |
| Reviewer | <<FILL>> |
| Date | <<FILL>> |
1. Classification and system
| # | Item | P/F/NA | Note | Reference |
|---|---|---|---|---|
| 1.1 | The AI use is classified (internal intelligence, assistive draft, or conformance check), not direct-to-record | SOP 5.1 | ||
| 1.2 | A named human author is recorded as accountable | 11.100 | ||
| 1.3 | The AI system and version that produced the draft are recorded | supplier assessment | ||
| 1.4 | The content was grounded in the correct source documents by retrieval | SOP 5.3 |
2. Claim traceability and accuracy
| # | Item | P/F/NA | Note | Reference |
|---|---|---|---|---|
| 2.1 | Every factual claim traces to a named source that actually supports it | ALCOA+ | ||
| 2.2 | Every numeric value (limits, results, doses, dates, identifiers) was verified against source individually | 5.3 | ||
| 2.3 | Every regulatory citation was checked against the primary reference | 5.3 | ||
| 2.4 | Any statement the model flagged uncertain was resolved from source, not left as the model’s guess | 5.3 | ||
| 2.5 | No fluent-but-unsupported statement remains (no plausible prose standing in for a checked fact) | 5.3 |
3. Human authorship
| # | Item | P/F/NA | Note | Reference |
|---|---|---|---|---|
| 3.1 | The author performed a substantive review, not a rubber stamp (editing extent recorded) | 5.4 | ||
| 3.2 | The author is trained on the AI’s known failure modes | 5.4 | ||
| 3.3 | The content is approved through the controlled document workflow with a Part 11 signature | 11.50, 11.70 |
4. Records and change control
| # | Item | P/F/NA | Note | Reference |
|---|---|---|---|---|
| 4.1 | A record of the AI involvement (system, version, sources, authoring, review) is retained, scaled to risk | 5.5 | ||
| 4.2 | Drafts and source links live in the controlled document system, not an AI tool’s chat history | 5.5 | ||
| 4.3 | If this content is a labeling change, it went through controlled change management, not AI output alone | 5.6 | ||
| 4.4 | If the vendor model changed since last use, a known-content re-check was completed | 5.6 |
5. Signoff
| Field | Entry |
|---|---|
| All items pass or are justified N/A | <<FILL: Yes/No>> |
| Open fails (must be zero to release) | <<FILL>> |
| Reviewer signature and date | <<FILL>> |
| Approver signature and date | <<FILL>> |
6. References
21 CFR Part 11, sections 11.50, 11.70, 11.100; ALCOA+ data-integrity attributes. FDA draft guidance on AI to support regulatory decision-making (January 2025, draft); EU draft Annex 22 on artificial intelligence (7 July 2025, not in force).
Confirm the current version and status of each reference before issue.
Filled specimen
The following shows section 2 completed for an example CMC quality summary section, so you can see the expected detail. The values are illustrative; replace them with your own.
| # | Item | P/F/NA | Note |
|---|---|---|---|
| 2.1 | Every factual claim traces to a supporting source | Pass | 41 claims, each mapped to the development report or specification |
| 2.2 | Every numeric value verified individually | Fail then Pass | Specification limit for an impurity read 0.20% in the draft; source said 0.15%; corrected and re-verified |
| 2.3 | Every regulatory citation checked | Pass | 2 citations, both confirmed against the primary reference |
| 2.4 | Uncertain statements resolved from source | Pass | 1 flagged value supplied from the batch record |
| 2.5 | No fluent-but-unsupported statement remains | Pass | Full read against source completed |
In this example item 2.2 caught the single most dangerous AI failure on regulatory content: a confident, well-written but wrong specification limit. The reviewer did not release until the value was corrected against source and re-verified, which is exactly why the individual numeric check is a hard gate rather than a skim.
Common inspection findings this checklist prevents
- A wrong specification limit or result in fluent text that no one checked against source.
- A fabricated or mis-numbered citation reaching a submission.
- Content approved with a nominal author who did not actually verify the substance.
- A labeling change driven by AI output without controlled change management.
How to adapt this checklist
- Point the reference column at your own SOP section numbers.
- Keep item 2.2 as a hard gate: no numeric value ships unverified.
- Add rows for content-type-specific checks (for example a redline completeness check for labeling).
- Require zero open fails before release, and record any N/A with a reason.
- Confirm every regulation in section 6 against the current published version and status before issue.