AI final review for GxP records: a reviewer's reading before approval, with what to fix and how

What the module does

• Runs from the Review and approval panel of any record with review and approval, while the record is a draft or heading for its next review cycle; a record in review, signed or cancelled is not reviewed.
• Reads the record against the standards a reviewer applies: complete (no placeholders, empty tables or unexplained not-applicables), consistent (dates, numbers, names, lots, references and conclusions agree), specific and attributable (who does what by when, evidence named rather than asserted), compliant with what the record type requires under GMP, including data integrity, and clear to a reader outside the team.
• Gives a verdict (Ready to submit, Ready after the changes below, or Not ready) with up to 25 findings as must fix, should fix or consider, each with the place in the record, the issue, the record's own words where there are any and a suggested fix, plus what is done well and should stay.
• Checks what each record type must contain: a controlled document's scope and steps, a change control's impact assessment and plan, a CAPA's root cause and effectiveness criterion, an investigation's evidence and conclusion.
• For an investigation, that the conclusion follows from the evidence, that ruled-out hypotheses are ruled out by data, that the root cause is not "human error" without the condition behind it, and that the scope of impact was assessed rather than assumed.
• For a complaint, that the fields agree with the complaint text, that any mention of harm or a patient is reflected in the adverse event assessment, and that the decision matches what was reported; for a batch record review, unanswered items and exceptions without a resolved reference.
• Works beside the system's readiness checks, which list what the cycle still needs and refuse a submission while anything is left: the AI is told what they found and looks for what lies behind it instead of repeating it.
• Judges only from the record: it does not invent facts about the product, site or process, and where the record needs information it does not hold, the finding says what is needed and where.
• Tied to the version it read: the panel says when the record has changed since the run, and each run is an audit trail entry with the verdict and the findings, while the record's content and version stay as they were.

How it fits the platform

The final review is a step in the review and approval workflow every QLogix record shares: a named author, reviewers and approvers, readiness, the AI final review, then electronic signatures in the order the author chooses. It reads complaints, investigations, CAPAs, change controls, controlled documents and batch record reviews, and the other records routed for approval across the platform, from validation documents and risk assessments to regulatory submissions and laboratory records.

When Q Agent or a scheduled agent creates a record, the record is filled from the source records, the product's AI drafting runs on it, and the AI final review reads it as a reviewer would before it is handed to its author. The author sees which agent drafted the record and, below it, the final review's list of what to fix, then reviews, edits and routes it.

Where people decide

The AI reads and reports; the author decides. Fixing a finding is the author's edit, recorded in the audit trail as theirs, and open must-fix findings do not stop a submission: the author may submit, and the submission records whether it followed a current review, an earlier version's review, no review, or must-fix findings left open by the author's decision. Reviewers and approvers then sign with their own credentials; the AI never reviews, approves or signs in their place.