This is a ready-to-use log for capturing that a generative-AI assistant helped produce a GxP quality record, who reviewed it, and what they verified. It makes AI use transparent and reconstructable, which is what keeps the practice defensible in an inspection. Where your validated quality system can capture these fields in its own audit trail, use that instead of a standalone log; this template defines the minimum field set either way. Replace every <<FILL: ...>> placeholder. A filled sample row follows. This content is educational and general; adapt it and verify it before use.
Purpose
To provide a traceable record that generative AI assisted a specific quality record, under what conditions it was generated, what a named human reviewed and changed, and that facts and numbers were verified against the source. The log is the evidence that AI use was disclosed, reviewed, and owned, rather than hidden.
Field definitions
| Field | Format | Required | Who enters | When |
|---|---|---|---|---|
| Entry ID | <<FILL: prefix>>-NNNN | Yes | Reviewer | At review |
| Date/time | ISO 8601 | Yes | Reviewer | At review |
| Linked record ID | Deviation / CAPA / investigation / complaint ID | Yes | Reviewer | At review |
| Workflow | Deviation description / RCA narrative / CAPA draft / complaint-trend summary | Yes | Reviewer | At review |
| Assistant name / instance | Text | Yes | Reviewer | At review |
| Model version | Vendor build or version string; “not exposed” if unavailable | Yes | Reviewer | At review |
| Prompt / template version | Internal version ID | Yes where feasible | Reviewer | At review |
| Grounding | Facts supplied / retrieval corpus + version / open (open is not permitted for GxP content) | Yes | Reviewer | At review |
| What the model produced | Short description | Yes | Reviewer | At review |
| Facts verified against source | Yes / No / N/A | Yes | Reviewer | At review |
| Numbers verified or computed deterministically | Yes / No / N/A | Yes | Reviewer | At review |
| Edits made by reviewer | Short description or “none” | Yes | Reviewer | At review |
| Quality decision made by tool | Must be “None” | Yes | Reviewer | At review |
| Accountable author | Name | Yes | Reviewer | At review |
| QA oversight (where required) | Name / date | Per procedure | QA | At oversight |
Instructions
- Create one entry per AI-assisted record. If the same record was assisted more than once (for example a description then a CAPA draft), create an entry for each use.
- “Grounding” must never read “open” for GxP content; an open-prompt draft is not permitted by the parent SOP.
- “Quality decision made by tool” must always read “None.” If it does not, the use violated the permitted boundary; do not finalize the record and raise it to QA.
- Record the model version where the vendor exposes it. Where it does not, enter “not exposed” and note that a vendor model change is managed through change control instead.
- The accountable author is the human, never the tool.
Retention
Retain each entry for not less than <<FILL: retention period>>, aligned to the retention of the linked GxP record, per <<FILL: records retention schedule ID>>. The disclosure entry and the linked record together are the story of how the content was produced.
Acceptance criteria for a complete entry
- Every required field is populated and the linked record ID resolves.
- Facts and numbers are marked verified (or N/A with a reason), never left blank.
- “Quality decision made by tool” reads “None.”
- An accountable human author is named.
Filled sample row
The following shows one completed entry. The content is illustrative; replace it with your own.
| Field | Entry |
|---|---|
| Entry ID | AID-2026-0311 |
| Date/time | 2026-06-08T14:50-04:00 |
| Linked record ID | DEV-2026-0417 |
| Workflow | Deviation description |
| Assistant name / instance | Contained quality assistant, prod instance |
| Model version | build 2026-06 (pinned) |
| Prompt / template version | dev-desc-tmpl v3 |
| Grounding | Facts supplied via structured intake; open-web knowledge disabled |
| What the model produced | Draft description of a tablet-press stop with hopper-level alarm |
| Facts verified against source | Yes (equipment log, batch record) |
| Numbers verified or computed deterministically | Yes (stop duration corrected 20 to 22 min against equipment log) |
| Edits made by reviewer | Corrected duration; confirmed segregation occurred; no invented facts |
| Quality decision made by tool | None |
| Accountable author | A. Patel |
| QA oversight | R. Gomez, 2026-06-09 |
In this entry the model drafted the description, the reviewer corrected one number and confirmed the physical facts, and the log shows exactly what the model produced, what the human changed, and who owns the record. If an investigator asks how AI was used in this deviation, this single row answers it.
Common inspection findings this log prevents
- AI in use in quality records with no disclosure, discovered rather than declared.
- No way to show a human verified the facts and numbers in an AI-assisted record.
- No record of which model version or prompt produced a draft, so behavior cannot be reconstructed after a vendor change.
- A record where the tool, in effect, made a quality call that no field or reviewer flagged.
How to adapt this log
- Map these fields onto your validated quality system’s audit trail where it can hold them, and use a standalone log only for what the system cannot capture.
- Set your retention period to match the linked record’s retention.
- Add a QA-oversight cadence appropriate to the risk of the workflows you support.
- If you use retrieval-augmented generation, record the corpus name and version in the grounding field.