Batch record search, explained.

Batch record search in plain words: a small question about a lot, and hundreds of pages to answer it

How one question about lot 0388 travels from the asking to a signed answer

1. A question arrivesAn investigator looking into a slow filtration emails: what was the hold before filtration on lot 0388?
2. The right record is foundProduct, plant, lot, and the master version in force when the lot ran.
3. The entry is read and placed in contextPage 58, step 7.5.1: two handwritten times, a temperature, two sets of initials, then the limit, the class of the step and the filing. This is the step this guide is about.
4. The answer goes out with its page5 hours 17 minutes against 8 hours, record, page and step cited.
5. A person checks the page and decidesThe reviewer reads the source, signs, and the answer can be used in the investigation.

The people, lot and figures in this example are fictional.

Three limits on one hold, and three more surprises in batch records

Master record, executed record, CPP and hold: four words for this work

The one-page summary in section 24 lists fourteen terms for this work.

Master batch record vs executed batch record: what sits on one page, and how long you must keep it

In shortThe master record is the instruction; the executed record is what happened on one lot. US rules require thirteen kinds of content in the executed record, and EU rules require something close. Every element on the page is there because a rule put it there, and the paper has to stay findable for years after the lot has been used up.

US rules require a master record "for each drug product, including each batch size thereof", prepared and signed by one person and independently checked and signed by a second.9 The batch record must begin with an accurate reproduction of the master, then document each significant step: dates, equipment, materials, weights, in-process results, yields and the people involved.10

The EU's good manufacturing practice rules, EU GMP, use other names for the same things: a Manufacturing Formula and Processing Instructions, then a Batch Processing Record for each batch.12 Plants making a drug substance, the active ingredient, follow ICH Q7, which also asks for "actual results recorded for critical process parameters".2

Medical device makers need the same habit under a different name. The Quality Management System Regulation of the US Food and Drug Administration (FDA) took effect on February 2, 2026 and incorporates ISO 13485, the international quality management standard for medical devices.13 FDA dropped the term device history record, and says many of its requirements are largely covered by the medical device or batch record of clause 7.5.1.14 A person searching device records is asking the same kinds of questions: which lot, which step, which limit, who checked.

One executed page, element by element, and the clause that puts it there

Records have to outlast the question. US rules keep a batch's records at least one year past its expiry date, and "readily available for authorized inspection" during that time.15 EU rules keep batch documentation for one year past expiry or five years after a Qualified Person certifies the batch, whichever is longer.12 A Qualified Person, or QP, is the named professional who certifies each batch for the EU market.

Why one batch record answer needs the record, the criticality analysis, the filing and the hold study

In shortThe executed record says what happened. Whether that was acceptable, how much the step matters and what the regulator approved sit in other documents, with other owners, and they change on separate timetables.

Take the sentence from our example: lot 0388 held for 5 hours 17 minutes. Four more questions decide what that sentence is worth.

• Is that inside the limit? The master record, in force on that day, says at most 8 hours at 15–25 °C.
• Does the step matter? A process parameter criticality analysis, the risk document that scores each setting against the quality of the product, says yes.
• Was the plant allowed to run it that way? Module 3 of the Common Technical Document (CTD), the quality part of the marketing application, registers the process in section 3.2.S.2.2 and the controls of critical steps in 3.2.S.2.4.5
• Is there room if it ran long? A hold time study gives the measured evidence.11

Different questions lean on different documents. The grid below is our reading of which sources a person usually has to open, drawn from what each document is required to hold. Your own procedures may change a square or two.

Who owns each number, and why a master range can drift from the filing

No system keeps these five in step on its own. A US biologics license holder must inform FDA about each change to a production process established in its license, and must assess the effect before distributing product made with the change.16 So a range in a plant's record that sits outside the registered range is a question for regulatory staff, whatever the batch itself looks like.

The plant reads its record and the regulatory team reads the dossier, its full filing. The two meet when something forces them together: an inspection, a transfer, or a question that happens to touch both. The new CTD quality text under discussion, ICH M4Q(R2), is still unfinished at Step 2, with Steps 3 and 4 planned for June 2027, so the section numbers above remain the ones in use.17

21 CFR 211, EU GMP Chapter 4 and the 2025 drafts: which batch record rules are in force in November 2026

In shortThe rules that govern batch records are old and stable. What moved in 2025 and 2026 is a set of EU drafts on documentation and AI, still unadopted on November 3, 2026, plus a US device rule that took effect in February 2026. Read every summary of "the new Annex 11", the EU annex on computerized systems, with that date in mind.

What a critical process parameter is, and why its class can change after a lot has run

In shortA critical process parameter is a setting that can move a critical quality attribute, so it must be monitored or controlled. The class comes from a risk assessment, with its own versions. When a reader asks whether a step is critical, the honest answer names the document, the version and whether that version applied when the lot ran.

ICH Q8(R2) gives the definition most firms use: "A process parameter whose variability has an impact on a critical quality attribute and therefore should be monitored or controlled to ensure the process produces the desired quality."3 A critical quality attribute is a property the product must have, such as purity. ICH Q9(R1) sets out the risk thinking behind the scoring.4

Some firms then add a middle tier, often called key, for settings that affect how the process performs without reaching quality; one published workflow for biopharmaceutical process validation labels factors that way.33 We found no definition of that tier in ICH Q7, Q8(R2), Q9(R1), Q10, Q11, Q12 or Q13,2,3,4,6,5,7,8 so its meaning comes from each firm's own procedure.

The class has practical weight, because ICH Q7 divides review by it.2 A step misread as routine can get lighter attention at review, in an investigation and at change control. A step misread as critical can trigger needless work.

Now imagine an investigator asks about a lot run on September 2, 2026, and the current analysis took effect on September 30. Two different questions hide inside "is this critical?" What did the plant's own documents say on the day the lot ran? And what do we believe today?

The first decides whether the lot was handled by its rules at the time. The second can trigger a look back at earlier lots. A good answer gives both, with the version of each. A reader who sees only "critical, version 3.0" has to ask.

How a CPP, a key parameter and an in-process control differ on a batch record page

How to verify a hold time on an executed batch record, with three lots worked through

In shortA hold is the gap between two written times, judged at a stated temperature against the printed limit. The arithmetic is easy and the traps are not: which event starts the clock, which limit applies, and how far the study's longer time counts.

US rules ask for time limits on each phase of production "when appropriate", and say a deviation "shall be justified and documented".27 ICH Q7 says limits written in the master instructions should be met, and deviations documented and evaluated.2 EU GMP requires "dates and times of commencement, of significant intermediate stages and of completion", which is what makes a hold possible to work out.12

The WHO's general guidance defines a hold as a period "under specified conditions", so a limit without a temperature is only half a limit. That guidance focuses on non-sterile solid dosage forms and says it does not cover active ingredients or biologicals, so read it here for the idea, and look to your own validated studies for a drug substance.11

Hold before depth filtration: lot 0388, a late lot, and one past the study

A real Dutch record would write 15:02. This guide shows US clock times, with the same values.

Now look across a campaign, the run of lots made in one stretch. Checking one lot tells you it passed. Six lots tell you how much room you have.

Lot 0402 passed, and a note on its page says why it ran long: filtration started late while filter housing F-3502 was replaced, recorded under deviation DV-2026-1162. On a single-lot review that is a page in order. Seen in a row of six, it is the lot with the least room, and a person might ask whether a spare housing should be kept closer.

Four ways the same hold gets miscalculated

Reading scanned, handwritten and translated batch records without changing a digit

In shortRules require clear handwriting and let a verified scan stand in for paper. Machines read modern handwriting well in test sets and far worse on hard material. No public benchmark exists for batch records in any language. One wrong digit changes the answer, so a person must be able to see the original.

Good documentation practice is the set of habits that keep a record readable and trustworthy. EU GMP says handwritten entries should be "clear, legible, indelible".12 PIC/S adds checks that an inspector might apply: is the handwriting consistent for the same person, are ditto marks avoided, are all pages present and in order?1 FDA says electronic copies can serve as true copies when they "preserve the content and meaning of the original record".19 PIC/S says a scanned copy should be checked against the original to confirm there is no tampering with the image.1

ALCOA+, the data integrity shorthand, stands for attributable, legible, contemporaneous, original and accurate, plus complete, consistent, enduring and available. The UK regulator says the extra letters carry "no difference in expectations".34

Machines are getting better at handwriting, with wide variation. An open benchmark of large language models (LLMs, the kind of AI behind chat tools) found character errors under 2% on modern English and French test sets and over 20% on old Italian manuscripts, and it reports a "weaker performance on languages other than English".35 A separate test on 18th and 19th century English letters gave 5.7% to 7%.36 These are academic data sets, far from a plant.

The nearest industrial study, on handwritten acceptance sampling records at a pharmaceutical company, reports efficiency gains and says limits in accuracy and the need for human validation "restrict full automation".37 It tested a hybrid recognition method rather than an LLM, and we read its abstract only.

Picture one wrong character on our page. If a machine read the start of filtration as 4:02 pm where the operator wrote 3:02 pm, lot 0388 would show 6 hours 17 minutes. It still passes, with an hour less room, and nothing on the screen would look wrong. On lot 0402, a misread of the same size would turn a pass into a deviation. So test for errors on the fields that matter, such as digits, times and units, on your own pages.

Language traps on Dutch and French batch records

What FDA inspectors write up on batch records and investigations: the Form 483 counts for 2024 and 2025

In shortA Form 483 lists what an FDA investigator observed at the end of an inspection. In FDA's file for fiscal 2025, the section on production record review and investigations is the second most cited in the drug manufacturing rules. Weak investigations are the largest part of it, and an investigation is where batch record questions land.

Inside 211.192 in that file, the row on investigations of discrepancies and failures carries 164 citations. Incomplete written investigation records add 32, and no written record 22. Quality unit review of records carries 9.38 So the typical finding is an investigation that went too shallow, which is when someone should have asked what other lots looked like. The rule itself says an investigation "shall extend to other batches" that may be associated with the failure.26

After an inspection, FDA gives the facility one of three classifications. Official action indicated means an unacceptable state of compliance.41 In fiscal 2025 that was the result for 18% of 1,309 drug quality inspections.42

Four FDA warning letters on batch records, holds and investigations

Each firm is named only for what FDA's letter says. These are examples, picked by us, and a warning letter reflects one inspection at one time.

The errors a person makes when answering from a batch record

How much paper plants still run on, how often batches fail, and which figures nobody publishes

In shortGood numbers are scarce. Vendor surveys suggest manufacturing execution systems, the software that runs a shop floor, reach only some of a company's sites. One plant's published description speaks of thousands of paper records a year. Batch failures, the moments when search matters most, came about a year apart at the average facility in a 2022 survey. We found no public figure at all for the time spent answering a batch record question.

Survey sizes this small only tell you the direction. A manufacturing execution system (MES) runs the shop floor, and an electronic batch record (eBR) is its digital record. Even where a site has one, older lots may still be paper, and they stay in play for the retention periods above.

A published description of one contract manufacturer's Spanish plant, in sponsored content from 2021, speaks of "thousands of paper batch records a year" and "hundreds of thousands of manual entries", with up to six people revising each document.50 A customer quoted on a vendor's page speaks of records of 400 to 500 pages with 2,000 to 3,000 manual entries.47

Each of those failures starts an investigation that reads batch records. BioPlan's 22nd annual report, summarized in December 2025, puts operator error at 3.8% of clinical-scale respondents and 3.5% of commercial ones as a cause of failure.52 Other published figures come from vendors or trade articles, and they give a rough size at best:

Same step, four plants: how master batch records differ across sites and versions

In shortMulti-site products run under several master records, in several languages, with different step numbers. All can sit inside one filed range and still differ. Comparing them takes hours by hand, and the differences that matter most are often in words, such as when the clock starts.

A single product can be made at several plants. Each plant writes its own master record to suit its equipment, and each record moves on its own version history. Quality teams and MSAT scientists compare them for tech transfers, investigations, audits and harmonization projects. Lining up the same step means matching by meaning, because the same operation may be step 7.4.3 in one record and 9.6.2 in another.

Fictional sample product. Numbers from the example used throughout this guide.

Any of these differences may be justified, and each still matters to a reader. A 70-minute hold passes at Plant 1 and at Plant 3, while the same minutes measure different things if the clock started at a different moment.

Versions add a second complication: a change control can narrow a range or add a limit that did not exist, and a lot made before the change is judged by the old text. So every comparison needs the effective date of each side, and any lot you ask about has to be matched to the version in force when it ran.

Batch record AI under draft Annex 22, EMA and FDA: what each text says about models, people and GMP decisions

In shortNo regulator has said a batch record search tool is banned or cleared. FDA's AI guidance is a draft. EMA's paper says to manage it through quality risk management. The draft EU annex says large language models (LLMs) should stay out of critical GMP applications, and EMA's own later report shows that point being reconsidered. The position that holds up under all of them: software finds and cites, a qualified person reads the source and decides.

Is a batch record search a critical GMP application? How to reason it through

Nobody has answered this for you, and the draft Annex 22 test turns on direct impact on patient safety, product quality or data integrity. A tool that finds pages for a person who then reads them looks different from a tool whose output clears a hold. EMA's workshop discussed an LLM supporting a GMP risk assessment for an electronic batch record system among its examples, some of them pilots or theoretical.24 So write your own risk assessment under ICH Q9(R1),4 for each use, and let your quality unit decide.

What can still go wrong when every answer has a citation

A citation lets a person check, and a wrong answer can still carry one. In a Stanford study of three legal research tools from two providers, all built to ground answers in retrieved sources, each gave a hallucinated answer, meaning one that was incorrect or cited a source that did not support it, on 17% to 33% of test queries.61 That study is in law, and the tools have changed since 2024.

The lesson carries over: measure on your own questions, and make the source one click away so the check is cheap enough that people do it. A reviewer who stops reading the page because the tool is usually right is the failure to design against.

How QA and MSAT teams answer batch record questions today, and where an estimated 42 minutes go

In shortA question goes to whoever knows the plant. They find the lot, pull the record, page to the step, read the entry, then chase the limit, the class and the filing through other systems. Nobody times it. An estimate we use below puts one answer at about 42 minutes, and you should replace it with your own.

QA reviewer, MSAT scientist, QP, regulatory CMC: who asks and who answers about a lot

Four ways plants answer batch record questions, and what each gives up

"What was the hold time before filtration on lot 0388?": the scanned page, the boxed step and the reviewer's signature

In shortThis is the example from the top of the page, inside the product. Dr. Maya Chen, a QA batch record reviewer, types her question. Agents, pieces of software that each do one defined job, find the lot, read page 58 of a 318-page Dutch record, and check the criticality analysis, the filing and the hold study. The answer comes back beside the scanned page with the step boxed, and Maya reads the page before she signs.

1

The question box beside a page

Search covers 3 products and Plants 1, 2, 3 and 5: 2,598 batch records, with their criticality analyses and filings. The screen already holds an earlier answer, on lot 0418's low-pH hold: 74 minutes, verified by Lena Ortiz. On the right sits the scanned page behind it, with the step boxed.

2

The steps show as they run

Maya asks about lot 0388. The Batch Record Agent first finds the lot: an executed batch record from Plant 3, 318 pages, in Dutch. It then reads page 58, working out that "before filtration" means the neutralized pool hold, and goes on to the criticality analysis, the filing and the hold study. Once the answer is in, the steps fold to "4 steps · 3.9 s".

3

5 h 17 min, beside page 58

The answer reads: lot VLT-L3-26-0388 held the neutralized pool for 5 hours 17 minutes before depth filtration, 9:45 am to 3:02 pm on September 2, at 21.8 °C. The Plant 3 batch record allows at most 8 hours at 15–25 °C, so it ended 2 hours 43 minutes inside the limit. Each number has a numbered citation, and each citation opens its page.

4

Whether the step is critical, on the same screen

Under the answer sits the criticality context. Neutralized pool hold time is a critical process parameter in analysis CA-VLT-DSP-004 v3.0, table 6-3: quality attributes aggregates (SEC, a test for clumped protein) and bioburden (the count of microorganisms present), severity 8 of 10. The filed limit is at most 8 hours at 15–25 °C; the hold study validated 12 hours at those temperatures. A bar shows the actual against both.

5

The step boxed, the initials circled

The boxed row is step 7.5.1 on the scanned page. J. de Vries entered the hold at 3:04 pm, and L. Ortiz made the second-person check at 3:20 pm. Both initials are circled and marked "signed", and a tag on the box marks it CPP. This is what Maya came for: she reads the entry herself. Pages 57 and 59 sit in a strip below, in case the context matters.

6

The same page in English, numbers untouched

One click turns the page into English. "Houdtijd vóór dieptefiltratie: maximaal 8 uur bij 15–25 °C" becomes "Hold before depth filtration: at most 8 hours at 15–25 °C". Step numbers, values and units stay as written. Decimal commas appear as points in the English only, and the original is one click back.

7

The full answer: step, translation and checks

"Full answer" opens the record of this one question. The source step sits in Dutch beside its English, with three checks under it: inside the batch record limit, inside the filed limit and the validated hold, and both initials present. Five sources sit on the right, each with its quoted line and a link that opens at the step.

8

Maya signs, and says what her signature means

Maya has read the page. The signature dialog asks what it means ("Verified: the cited steps match the source documents") and where the answer will be used. Her identity comes from single sign-on. Name, date and time, and meaning are the elements 21 CFR 11.50 lists for a signed electronic record.28

9

On record, with its history

The answer now reads Verified, and the History tab shows who asked, what each agent did and when Maya signed. If someone needs the same fact next week, they find a signed answer. If Maya signed in error, she can withdraw it, and that is recorded too.

Trending a hold across six lots, and catching a master record range outside the filed range

In shortTwo more questions show what a single lookup cannot. One lines up a step across a whole campaign and finds a lot with 19 minutes to spare. Another finds that Plant 3's AEX range runs 0.5 mS/cm above the filing, and sends that to people who can act on it. AEX is anion exchange, a purification step, and mS/cm measures conductivity.

1

"Show the hold for every lot this campaign"

Six lots, 0388 to 0418, step 7.5.1 on page 58 of each record. All held inside 8 hours, from 2 hours 55 minutes to 7 hours 41 minutes. Lot 0402 is marked: 19 minutes to spare. Clicking a lot opens its page. The answer took 5.6 seconds, against reading six records in turn.

2

The reason is written on lot 0402's page

The note on page 58 of lot 0402 reads "Filtratie later gestart: filterhuis F-3502 vervangen (zie DV-2026-1162)": filtration started late because the filter housing was replaced. Both versions sit side by side in the record viewer, and the initials are R. Visser's and M. Bakker's. Whether this deserves a closer look is for Maya to decide.

3

Every question across the plants

Questions lists everything anyone asked: who asked, which plant, the status, confidence, answer time. Filters narrow it to what waits for a person. This is how a team sees earlier answers before asking again.

4

Three answers wait for Maya

"Needs your check" holds Omar Haddad's Plant 5 questions on Protein A resin cycles and the neutralized pool hold, and Lena Ortiz's question on AEX load conductivity at Plant 3, flagged as differing from the filing.

5

0.5 mS/cm above the registered range

Plant 3's master record sets AEX load conductivity at 5.0–8.5 mS/cm in step 8.2.3. Module 3.2.S.2.4 registers 5.0–8.0 mS/cm, and Plant 1 runs 5.0–8.0. The answer says so plainly, and suggests raising it with Regulatory CMC before the next Plant 3 campaign. A difference from the filing is never cleared by software.

6

The table behind the flag

Plant 3 against Plant 1, parameter by parameter: 3 of 7 differ, one of them outside the filed range. Dana Okafor in Regulatory CMC has already been notified. What to do next is hers and Maya's to work out. They could correct the record or change the filing; a US license holder would look at 21 CFR 601.12 for how.16

Comparing plants and master record versions, and reading records where they already live

In shortCompare lines up one step across two plants or two versions of a document, and marks what differs after allowing for units and decimal commas. Records lists what the search reads, with type, plant, language, version and status. The records stay in your quality document system, read only.

Five batch record agents, what each may never do, and who signs which answer

In shortFive agents find, read, translate, compare and check. None of them edits a record, releases a lot or closes a deviation. A named person verifies at the cited page, and answers that differ from the filing need two people. These settings live with your quality team.

Batch Record, Language, Comparison, Criticality and Intake: the five agents

Who signs a batch record answer, by what it will be used for

The people on the Veltrimab screens

• Controls that support your compliance. Citations on every claim, effective versions by default, access by site, a signature with its meaning, and a full audit trail of questions, answers, pages opened and signatures. GMP, 21 CFR Part 11 and EU Annex 11 duties stay with you, and you assess the tool for your intended use under your own quality system.
• Credentials. Botminds is SOC 2 Type II audited, and its information security is managed to ISO 27001.
• Enterprise scale. Roles and maker-checker rules (one person acts, another checks) per site, audit export, read-in-place connection to your quality document system, and multi-plant, multi-language records on one search.

Batch record AI accuracy: what was measured in production, what the design targets are, what goes in and what comes out

In shortMeasured in production at a global medtech manufacturer: 0.95 accuracy against the client's own expected answers, up from 0.88 and 0.75 on earlier pipelines, in under 6 seconds per answer. That is one customer's result, on its own questions. The time saving is a design target, and the dashboard below shows sample data.

Accuracy here is scored against that client's own expected answers. It says nothing about your plants, languages or handwriting, which is why section 22 starts with your own expert answers. People still read every answer they use: at 0.95, some answers will be wrong, and the page is how you catch them.

From scanned lot records and Module 3 filings to a signed, reusable answer

What your document, MES and regulatory systems keep, and what search does with them

What slow answers and misread limits cost a quality operations team

In shortThe real cost is unpublished, so start from your own. A few published figures show the pressure: how often batches fail, what inspectors cite, how long records must stay findable. Every hour on this page comes from you.

How to time your own batch record questions in one week

What a late or wrong hold answer costs beyond the hours

What faster batch record answers return when you count hours, trends and filing checks

In shortPut in your own numbers. The calculator counts questions answered, campaign trends built and plant-against-filing checks done. The 42 and 14 minutes are an estimate and a design target; other defaults are assumptions.

Answer time, trend reports and filing checks: where batch record search saves

Questions a week, minutes per answer and your hourly cost

How ready your scans, versions and signers are for batch record search: an eight-question score

In shortScore each question 0, 1 or 2. Low scores tell you what to fix first, usually the state of your scans or the map of which version applied when. A ninth question has no score, and it may be the most useful one.

What your score means: build a baseline, pilot one campaign, or plan the plants

Twelve questions for any batch record AI vendor on handwriting, versions, the filing and signatures

In shortAsk us these too. Each comes with the reason, a good answer and a warning sign. Better still, bring your own worst pages and watch.

How to pilot batch record search on one campaign in twelve weeks, against your own expert answers

In shortPick one product, one campaign and one language you find hard. Write the expected answers before the tool sees anything. Then measure accuracy, time and the errors people catch, keeping a person on every answer. Your numbers replace ours.

What to measure on the expert set and the parallel run

Batch record search, executed batch record review and hold times: answers to what people ask

In shortThe questions QA and MSAT teams and buyers search for, each answered on its own.