AI Adoption · Pharma & life sciences · Written and maintained by Haink’s AI adoption team · Updated August 2026 · 12 min read

AI-assisted batch record review: what it can catch, and where it has to stop

A batch that has been manufactured but not released is inventory that cannot ship. That is the whole economic case, and it is why batch record review attracts automation budget faster than almost anything else in a GMP plant. Practitioners commonly cite around 48 hours of review effort for a single batch as an industry average, with complex products running far longer — and sites still on paper records reporting right-first-time rates in the 60–70% range against 85–95% for best-in-class operations. The gap is not manufacturing quality. It is documentation.

So the question is not whether to automate the review. It is which part of it is actually a language problem, which part is a rules problem, and which part is a decision that a human has to keep making. Getting those three apart is most of the work.

Where the review time actually goes

“Reviewing the batch record” is not one task. Decomposed, it is roughly seven, and they have very different automation profiles.

Note what falls out of that table: most of the volume is deterministic, and most of the pain is narrative. Teams that reach for a language model first often automate the part that a rules engine did better, then discover the deviations and annotations — the part that actually holds up release — still need reading.

Two automations that get confused with each other

Before scoping an AI project, establish which of these you are actually missing.

Review by exception is the MES/EBR capability: the system enforces the record as it is created, so at review time only non-normal events surface. No AI involved — it is deterministic rules over structured data, and where it exists it is extremely effective. A widely cited example from the BioPharm International literature: Valent BioSciences cut batch review from 20 days to 1 day and recovered around 2,700 person-hours a year, turning a 150-page record into a three-page exception report. If you have a modern MES and have not implemented review by exception, that project outranks any AI project on your list.

AI-assisted review is what you need when the record is not clean structured data in the first place: paper or scanned hybrid records, PDFs from a contract manufacturer, records spread across systems that were never integrated, or free-text annotations that no rule can anticipate. That is a document and language problem, and it is the same class of problem as the maintenance packages we handled in civil-aviation MRO — classification, signature and stamp detection, cross-package consistency.

What an assistant can actually catch

Framed as findings for a reviewer, not decisions. Each finding cites the page and section it came from, so verification takes seconds rather than a re-read.

The bottom row is not decoration. It is the difference between a system a quality unit will approve and one it will not.

Where it has to stop — and the letter that made that concrete

On 2 April 2026 the FDA issued what is widely described as its first warning letter identifying AI misuse as a cGMP issue. The firm had used AI agents to help create drug product specifications, procedures and master production records — and had not verified that the resulting documents were accurate and cGMP-compliant. The agency cited 21 CFR 211.22(c), the provision placing responsibility for approving procedures and specifications on the quality unit. A further finding was that required process validation had not been completed; personnel reportedly explained that the AI agent had never identified the requirement.

Two lessons transfer directly to review work.

The quality unit's responsibility does not move to a tool. Whatever the system produces — a draft, a finding, a summary — a person approves it and that approval is the record. Any design that quietly removes the human from the loop to save time is removing the only thing making the output admissible.

Absence of a finding is not evidence of compliance. The agent did not flag the missing validation, and nobody had specified that it should. An assistant that reviews records has a defined scope of checks; everything outside that scope remains the reviewer's job, and the scope has to be written down where the reviewer can see it. A system that appears to check everything is more dangerous than one that visibly checks eleven things.

Which regulatory lane this sits in

Under the draft EU GMP Annex 22, generative models are excluded from GMP-critical applications, while non-critical support — summarising, searching, drafting, with a qualified person reviewing the output and retaining documented responsibility — is explicitly permitted. Review assistance is designed into the permitted lane when, and only when, three things hold: the system produces findings rather than dispositions, a named person signs, and the record of that approval is retained.

The deterministic components — arithmetic, limit checks, sequence rules — sit in a different category again. Where they directly support a release decision they may fall in scope as static, deterministic models with the full validation burden that implies. This is why an honest scoping conversation separates the layers before anyone estimates effort. We cover the design consequences in non-functional requirements for AI systems.

What the system has to have to be auditable

• Version-aware retrieval. Checks run against the master record and SOP versions that were effective on the date of manufacture, not the current ones. A system that silently compares a 2024 batch to a 2026 master produces confident, useless findings.
• Citation on every finding. Page, section, and source document with version. Unverifiable findings get ignored by reviewers within a week — the fastest way to kill adoption.
• Confidence and abstention. A low-confidence read of a handwritten entry should be surfaced as unreadable, human check required, never as a confident value.
• Full audit trail. Input, model version, prompt, output, retrieved sources, reviewer, disposition, timestamp — retained as a GMP record under Annex 11 principles.
• Frozen versions under change control. Model, prompts and rules pinned; changes assessed for revalidation.
• Deployment inside the perimeter. Batch records are among the most sensitive documents a manufacturer holds. See private LLM for pharma for the deployment decision, private AI infrastructure for the hardware, air-gapped AI for pharmaceutical R&D for how far that constraint can go, and document extraction under validation for what an air gap requires as a set of procedures rather than a rack.

How to measure it — and what not to measure

The metric that matters commercially is time from end of manufacture to release, because that is the one finance recognises. But it is a lagging indicator with many contributors, so pair it with two others.

What not to measure: percentage of the review automated. It rewards expanding the model's scope, which is the opposite of what a quality unit wants, and it is unfalsifiable. Also avoid claiming a headline percentage from a vendor benchmark — the only number that will survive a management review is the one measured on your own records. On building that case properly, see how to measure AI ROI.

A realistic first increment

One product family. A retrospective set of closed batch records — one to two hundred is usually enough to be meaningful. A defined, written scope of checks. Acceptance measured against the site's own reviewers on that set, agreed before the work starts, not after the results are in.

What makes this project succeed or stall is almost never the model. It is whether the records, the master data, the specification limits and the QMS entries can be reached programmatically, and whether someone owns the answer when a finding is wrong. That is the same failure pattern as everywhere else in enterprise AI — see why AI pilots fail — with the difference that here the consequence of an unowned wrong answer is a regulatory one. How we take systems like this from design to production is described in AI solution implementation.

Frequently asked questions

Can AI approve or release a batch?

No. Release is a decision reserved for a qualified person, and under 21 CFR 211.22(c) responsibility for approving procedures and specifications sits with the quality unit. An assistant produces findings; a person decides. The draft EU GMP Annex 22 points the same way by excluding generative models from GMP-critical applications.

Do we need AI if we already have an MES with review by exception?

For completeness, arithmetic and limit checks — probably not. Review by exception handles structured, system-enforced records well; one documented implementation cut batch review from 20 days to 1. AI becomes relevant where records are paper, hybrid, supplied by a contract manufacturer, or where free-text narrative has to be read and cross-checked.

What did the FDA's April 2026 AI warning letter actually say?

It concerned a firm that used AI agents to help create specifications, procedures and master production records without verifying that the outputs were accurate and cGMP-compliant, cited under 21 CFR 211.22(c). The transferable lesson is that using AI to assist document work is not itself the violation — failing to review and verify what it produced is.

How much of the review can realistically be assisted?

Ask instead which findings the assistant is scoped to produce, and measure detection against your own reviewers on your own historical records. A percentage quoted from a vendor benchmark says nothing about your record formats, your product complexity or your data availability, and it will not survive scrutiny in a management review.

Does the assistant have to run on our own infrastructure?

Not by regulation, but batch records are among the most confidential documents a manufacturer holds, and Annex 22 expects model versions to be frozen under change control and model provenance to be documentable. Both are easier with self-hosted open-weight models than with a managed endpoint that can change beneath you.

What do we need before starting?

Programmatic access to closed batch records, master records and specification limits; a defined scope of checks; a retrospective set for acceptance testing; a baseline measurement of current review effort; and a named owner in the quality unit. The last two are the ones most often missing.