← Guides

Will AI-assisted batch record review hold up in an FDA inspection?

By Savi C., Co-Founder · Published Nov 16, 2026

What the regulations actually require

21 CFR 211.192 requires the quality control unit to review and approve production and control records for compliance with written procedures before a batch is released, and to thoroughly investigate any unexplained discrepancy. It does not say how the review is done. 21 CFR 211.68 allows computer equipment in CGMP work if it performs its function satisfactorily, with inputs and outputs checked for accuracy. Where the tool keeps electronic records, Part 11 adds validation under 11.10(a) and secure, time-stamped audit trails under 11.10(e).

So the question an investigator asks is not whether AI is allowed. It is whether your review was adequate and whether you can prove it. The broader regulator positions on AI, including the January 2026 EMA and FDA principles and draft EU Annex 22, are covered in our guide on whether AI can be used for GMP batch record review.

Questions to expect, and the evidence that answers them

Likely question

Who approved this batch record?

Evidence to have ready

The QA reviewer’s signature or Part 11 electronic signature, applied after the review was complete

Likely question

What did the tool check, and what did it not check?

Evidence to have ready

A documented intended use and a list of the checks configured for this product and master record

Likely question

How do you know it works?

Evidence to have ready

The validation report: test records, acceptance criteria, results compared with human reviewers

Likely question

What did the reviewer see for this batch?

Evidence to have ready

The list of findings, each linked to its source page, with the reviewer’s decision on each

Likely question

Why was this finding dismissed?

Evidence to have ready

A recorded justification by a named reviewer, not a silent close

Likely question

Has the tool changed since validation?

Evidence to have ready

Change control records and regression test results for every model, rule or configuration change

Likely question

What happens when it cannot read a page?

Evidence to have ready

Evidence that unreadable or missing content is reported to the reviewer as unverified, not passed

Where AI review tends to be weak in an inspection

The risks are the familiar data integrity ones, in a new place. A tool whose findings can be dismissed with no reason recorded. A reviewer who signs within minutes of the findings appearing, which invites the question of whether the review was real. Model or prompt changes made by a supplier without the site knowing. Checks that only cover part of the record while the approval implies all of it was reviewed.

Each of these has a straightforward control, and FDA’s own reasoning points the same way. Its computer software assurance guidance, written for device production and quality software, treats software that determines acceptability with limited or no additional human review as high process risk. Keeping a qualified person reviewing every finding is not only good practice, it is part of the risk argument for your validation.

An inspection readiness checklist

Procedures

  • The batch record review SOP describes the AI-assisted step, what it checks and what the reviewer must still do
  • Training records show reviewers were trained on the tool and on how to document decisions
  • The SOP says what to do when the tool is unavailable

Records

  • Validation report and current validation status for the configured version
  • Per-batch findings, reviewer decisions and justifications, retained with the batch record
  • Audit trail showing who did what and when, reviewable in an intelligible form as Annex 11 section 9 expects for EU sites
  • Change control history for the tool

Performance

  • Routine monitoring of missed findings and overturned findings
  • Deviations raised against the tool, if any, and their investigations

How Litewave fits

Litewave reads paper or electronic records, checks every entry against your master record and specifications, and shows QA only the exceptions, each linked to the page it came from. Litewave proposes, qualified people approve, and it never dispositions a batch on its own. It supports 21 CFR Part 11 electronic records, signatures and audit trails, so the trail from finding to reviewer decision is there when an investigator asks for it.

Questions we get asked

Do we have to tell FDA we use AI for batch record review?

Treat it as part of your quality system, not a secret. It should be visible in your procedures, validation records and batch records, and you should be able to explain it clearly when an investigator asks how records are reviewed. Check with regulatory affairs whether any filing describes your review process.

Can the AI sign off a batch record?

It should not. 21 CFR 211.192 places the review and approval with the quality control unit, and in the EU, Annex 11 section 15 expects only Qualified Persons to certify batch release. The tool prepares the review, a person approves it.

Is a manual review safer in an inspection?

Not automatically. A manual review is only as defensible as its documentation. An AI-assisted review that records every check, every finding and every decision can be easier to reconstruct than a signature at the end of a 150 page record.

Sources

  1. 21 CFR 211.192, Production record review (eCFR)
  2. 21 CFR 211.68, Automatic, mechanical and electronic equipment (eCFR)
  3. 21 CFR 11.10, Controls for closed systems (eCFR)
  4. EudraLex Volume 4, Annex 11, Computerised Systems
  5. FDA, Data Integrity and Compliance With Drug CGMP: Questions and Answers (December 2018)
  6. FDA, Computer Software Assurance for Production and Quality Management System Software (February 2026)