Work Instruction: Human Review of AI-Assisted Deviation and CAPA Drafts

This is a ready-to-use work instruction for the reviewer who takes a generative-AI draft and turns it into a record they own. It is the operational detail behind the human-review control that the parent SOP relies on. Replace every <<FILL: ...>> placeholder with your own specifics. A filled specimen follows the template. This content is educational and general; adapt it and verify it before use.

Control header

Purpose

This work instruction defines the meaningful review a person performs on a generative-AI draft before adopting it into a GxP record. Meaningful review means reading the draft against the source facts, correcting what is wrong, and taking ownership, as opposed to approving a polished draft because it reads well. The whole approach in the parent SOP depends on this step actually being done.

Before you start

You need: the AI draft, the verified source facts or records the draft was supposed to be built from, and access to the disclosure and review log (<<FILL: log ID>>). If you do not have the source facts, stop; you cannot verify a draft against nothing.

The review steps

Step 1: Confirm grounding and scope

• Confirm the draft is a deviation description, root cause narrative, CAPA plan, or complaint/trend summary, and not a quality decision the tool is trying to make for you.
• Confirm the draft was produced from supplied verified facts, not from an open prompt that let the model invent context.

Per-step acceptance: the content is within the permitted drafting/summarizing boundary, and the source facts it should rest on are available to you.

Step 2: Verify every fact against the source

• Read the draft sentence by sentence. For each factual claim (a batch number, a time, a measurement, a sequence of events), find it in the source and confirm it matches.
• Mark any fact you cannot trace to the source. A traceless fact is presumed a confabulation until proven otherwise; delete it or correct it against the record.

Per-step acceptance: no fact remains in the draft that you have not confirmed against a source record.

Step 3: Verify every number and count

• For any count, percentage, or trend claim (14 occurrences, up 30 percent), do not trust the model’s arithmetic. Confirm the number against the source or a deterministic query.
• Where an exact number matters and you cannot verify it, replace it with the verified figure or remove the claim.

Per-step acceptance: every quantitative claim in the draft is verified or removed; none is left on the model’s word.

Step 4: Check for premature cause and generic actions

• In a deviation description, confirm the text states observations only and does not prejudge cause.
• In a root cause narrative, confirm the stated cause is the team’s evidenced conclusion, not the model’s first suggestion, and that alternatives the team ruled out are reflected.
• In a CAPA, challenge any generic reflex action (blanket retraining, “update the SOP”) that does not trace to the confirmed root cause. Reject it unless there is a genuine, cause-linked reason for it.

Per-step acceptance: no premature cause in a description, no unevidenced cause in an RCA, and no generic action in a CAPA that fails to trace to the root cause.

Step 5: Check for completeness and silent omission

• Confirm the draft covers what it claims to cover. For a summary, confirm no records were silently dropped because the input exceeded the tool’s context window.
• Add anything material the model could not know that belongs in the record.

Per-step acceptance: the record is complete for its purpose, with nothing material missing or silently truncated.

Step 6: Adopt, disclose, and own

• Make your edits directly, so the final text is yours.
• Record the review on the disclosure and review log: what the model produced, what you changed, that facts and numbers were verified, and that you are the accountable author.
• Submit the record under your own name, with the AI-assistance flag set.

Per-step acceptance: the record is flagged AI-assisted, the log captures your edits and verification, and you are recorded as the author and owner.

Overall acceptance criteria

The review is complete and defensible when:

• Every fact and number is verified against the source.
• No quality decision was made by the tool; the cause, the CAPA adequacy, and any classification are your judgment.
• No generic action survives that fails to trace to the confirmed root cause.
• The record is flagged AI-assisted, your edits are captured, and you are the recorded author.

Common inspection findings this work instruction prevents

• A defined review step that reviewers sign but do not perform, revealed by near-total acceptance of drafts with no edits.
• A model-generated number that reached a record without verification.
• A confidently written but factually wrong draft adopted because it read well.
• A generic “retrain the operator” CAPA that did not address the mechanism.
• No record of what the reviewer actually changed, so the review cannot be shown to have happened.

How to adapt this work instruction

1. Set the document number and point the parent-SOP field at your governing procedure.
2. Align the log fields in step 6 with what your audit trail and disclosure log can capture.
3. If you use a deterministic query for counts (step 3), name the validated report or query.
4. Add any house-specific checks your quality unit requires for a given record type.

Filled specimen

The following shows the review record completed for one CAPA draft, so you can see the expected level of detail. The content is illustrative; replace it with your own.

In this example the reviewer caught the single most common weak-CAPA pattern, a reflex retraining action that did not address the mechanism, rejected it, and rewrote the preventive action to target the real cause. The model produced the structure; the reviewer produced the record and owns it.