Batch record review by exception, explained.

A batch record is the plant's signed diary of one batch, and the quality unit must approve it before the medicine ships

One tablet batch, from issued copy to the shipping dock

1. QA issues a copy of the master recordBatch DR26-1109 gets a numbered copy of master record v7.2. Page counts are recorded so a missing page can be noticed later.
2. The floor makes the batch and fills in the pagesOctober 2 to 6, dispensing through packaging. Values, times, initials, printouts.
3. The record is reviewedEvery entry against the master record and the documentation procedure. This is where a 12-minute blend is caught, or carried forward.
4. Problems are workedCorrections go back to production. Deviations, the formal records of a departure from the instructions, are opened and investigated, including for other batches they may touch.
5. The batch gets its decisionPlants call this disposition. The quality unit approves or rejects it. In the EU a Qualified Person certifies it. Only then can it ship.

Step 3 is the subject of this guide. The people, companies and figures in this example are fictional.

Four findings that surprise people new to review by exception

The two records, the exception and the critical parameter: four terms this guide uses

Fourteen terms are defined in the one-page summary.

Three different things people call review by exception, and why the difference decides your risk

In shortThe phrase, often shortened to RBE or BRBE, covers three practices. An electronic batch system can report its own exceptions. A tool can screen an audit trail, a system's log of who changed what, for unusual changes to data. Newer software reads paper and exports, then checks what it read. They share a name and carry very different risks, because the first one prevented many errors at the point of entry and the third one starts by reading handwriting.

The nearest thing to a shared definition comes from the industry. A 2019 trade article quotes ISPE, the engineering society for the drug industry, as defining review by exception as "an approach in which manufacturing and quality data are screened to present or report only critical process exceptions".9 We could not find that sentence on ISPE's own site, so treat it as a quotation of a quotation.

ISPE's 2010 good practice guide on manufacturing execution systems lists an appendix on review by exception with a part headed "Definition and Justification", which may be where it comes from. That guide is sold, and we read only its contents page.11

MHRA, the UK medicines regulator, and PIC/S, a co-operative scheme among GMP regulators, give a tighter definition for the tool itself, in the context of audit trails: an exception report is "a validated search tool that identifies and documents predetermined 'abnormal' data or actions".12,13 Two words carry the weight. "Validated" means the plant has proven the tool does what it claims. "Predetermined" means someone decided in advance what counts as abnormal.

Exception reports from systems, audit trails and paper readers

What stays with named people in all three forms

A named person in the quality unit decides. Rules are written and approved before they are used. Someone checks that the entries the system passed were fine. And the full record stays available: EU GMP Chapter 4, on documentation, says "All records should be available to the Qualified Person".15 Review by exception changes what a person reads first and how much of the rest they read. It leaves the responsibility where it was.

What FDA, EMA, MHRA and PIC/S have put in writing about reviewing by exception, with dates

In shortEMA says review by exception of electronic data is permitted when scientifically justified, with a tested and validated system. EU Chapter 4 lets automatically generated reports be limited to summaries and exceptions, for a validated, continuously monitored process. FDA's rule says "All" records and we found no FDA guidance on the method. No regulator text we read covers software reading paper records. Chapter 4 and Annex 11 are in draft revision, and Annex 22 on AI is still a draft.

The texts, in their own words

Two more texts come up in this argument and settle less than people hope. ICH Q9(R1), an international guideline adopted January 18, 2023, is the general basis for risk-based decisions in quality work, and it says nothing specific about batch record review.24 FDA's January 2025 draft guidance on 21 CFR 211.110 covers in-process controls, advanced manufacturing and process models; FDA marks it as a draft, not for implementation, and it does not address the review of records.25

How we looked for FDA text and for text on paper records, and what a search cannot prove

FDA. On November 3, 2026 we searched fda.gov, federalregister.gov and regulations.gov for "review by exception" and "exception report". We word-searched the 2018 data integrity Q&A, the 2023 paper on AI in manufacturing, the two January 2025 drafts and the 2026 software assurance guidance,17,26,27,25,28 six other FDA texts,29 and 515 warning letters.6 None uses either phrase.

The search did find one use: a November 2023 paper on distributed manufacturing from CDER, FDA's drug center, reports stakeholders proposing automated batch decisions "by using review by exception algorithms". That paper says it "is not a draft or final guidance".7

Paper read by software. We searched those FDA texts, EU Chapter 4 and Annex 11 with their drafts, draft Annex 22, Annex 16, EMA's data integrity answers, MHRA 2018, PIC/S PI 041-1, WHO TRS 1033 Annex 4 and ICH Q9(R1) for handwriting, optical character recognition and scanning. None addresses software reading the entries on a paper record so that a person can skip pages.

The nearest line is MHRA's definition of a "validated scanning process", which uses controls "such as location identifiers and OCR" so that a duplicated page "does not have to be further checked by a human". That is about making a copy.12 FDA's Q&A on electronic systems in clinical investigations likewise speaks of validated scanning to make certified copies.29 EMA says its data integrity set applies to paper and electronic systems, and its answer on review by exception speaks of electronic data.16,5

A search can miss a text, and inspectors may hold views no one has published. Treat each "we found none" as true of what we read on that date, and ask your own inspectorate before relying on it.

What ISPE, the industry body, has published

ISPE is an engineering society with no regulatory power, and GAMP is its work on computerized systems. We found no ISPE guide dedicated to batch record review by exception. The timeline lists what we did find, and how much of each we read.

The quality unit in the US and the Qualified Person in the EU keep the batch decision, whatever reads the pages

In shortUS rules give the quality unit the authority to approve or reject and the duty to review records. EU rules add a Qualified Person, who certifies each batch and must be sure all records are complete and endorsed. Neither role can pass to software. A second-person check can sometimes be met by automated equipment, but only with a person checking the equipment.

21 CFR 211.22(a) gives the quality control unit "the responsibility and authority to approve or reject" drug products, and "the authority to review production records to assure that no errors have occurred or, if errors have occurred, that they have been fully investigated".35 ICH Q7, written for active ingredients, adds that the main responsibilities of the independent quality unit "should not be delegated".21

In the EU, GMP Chapter 1 requires that no batch be released "prior to certification by a Qualified Person".2 Annex 16, on that certification, makes that person responsible for ensuring that "All records are complete and endorsed by appropriate personnel".3 The draft EU Chapter 4 says that accountability for records processed by AI or other automatic means "rests with the regulated user".22

Does a Qualified Person read every page?

EU rules do not say so in those words. Chapter 4 says all records "should be available" to the Qualified Person and that "a system should be in place to indicate special observations and any changes to critical data".15 ECA Academy, a training body, reads this as no express duty for the Qualified Person to review each record personally, with production reviewing first and a competent quality function second.37 That is one body's reading of the text, and many companies expect more of their Qualified Persons than the text does.

Second-person checks: where a machine can stand in, and where it cannot

What 21 CFR 211.188 and EU Chapter 4 require on the page, and which errors each part invites

In shortThe rules list what an executed record must show: dates, equipment, lots used, weights, in-process results, yields, labeling, sampling, who did and checked each step, and any investigation. Some errors show on one line. Others only appear when two pages are compared, and a few can only be found outside the record, in a training file or a balance log.

21 CFR 211.188 requires "complete information relating to the production and control of each batch". It lists an accurate copy of the master record and thirteen more items, including dates, major equipment, lots of components, weights and measures, in-process and laboratory results, actual and percentage yields, labeling records, sampling, and the people "performing and directly supervising or checking each significant step".36 EU Chapter 4 paragraph 4.20 asks for much the same, in addition to "dates and times" of the start, significant stages and completion, and "signed authorisation for any deviation".15

Where the common batch record errors can be seen

The documentation rules, known as good documentation practice, are usually a site's own procedure built on those texts. The "ALCOA" test (attributable, legible, contemporaneous, original, accurate) comes from data integrity guidance, including MHRA's.12

Hold times, reconciliations and a spot-check's real strength, worked through with numbers

In shortThe sums are simple, which is why they get missed: nobody expects to find a wrong sum on page 291. Two pages can hide a 26-hour hold behind a 24-hour limit. A written 99.6% can be a true 99.1%. And finding zero misses in a small spot-check proves far less than it sounds.

A hold time no single page shows

A hold time is how long material may wait between two steps. Batch DR26-1108's master record says wet granule may wait no longer than 24 hours between the end of granulation and the start of drying, and calls that a critical process parameter. Page 58 records granulation ending October 1 at 9:40 pm. Page 81 records drying starting October 2 at 11:55 pm. Each page is correct and signed by different people.

21 CFR 211.111 says a deviation from a time limit "may be acceptable if such deviation does not compromise the quality of the drug product", and "shall be justified and documented".43 So going over a limit has to be justified in writing. On this record the drying page shows "N/A" beside deviation reference, which is a second finding. Both records are fictional.

A reconciliation written wrong, inside its limit

A reconciliation compares what came out with what should have, as a percentage. On batch DR26-1109, page 291, the packaging team records 590,140 tablets packed, 1,200 QC samples and 3,260 rejects, against 600,000 theoretical, and writes 99.6%.

Both figures sit inside the 98.0% to 100.5% limit, so no alarm rings. It is still an error to correct: 21 CFR 211.103 requires yields to be "independently verified",38 and a second person signed this one. A check that only asks "is the written number in range?" passes it. A check that recomputes it does not.

What zero misses in a spot-check can and cannot prove

Any exception method needs a sample of the pages it passed, opened by a person, to measure how often "clean" is wrong. A small sample proves little. If you check n pages and find no miss, a standard rule gives the most you can say at 95% confidence: the true miss rate is below about 3 in n.44 That ceiling is called the upper bound.

Six pages with no miss says the rate is below 39%, by the exact formula. It takes about 300 pages with no miss to say below 1%, and a 1% rate on a 295-page clean set would still mean about three wrong pages per record.

Two practices follow. Pool the spot-checks across records to get a meaningful number over a month or a quarter, and report the bound along with the count. And add planted errors: a challenge set of records with known errors tests a system directly, while a spot-check only samples what it happened to pass.

What FDA's inspection data shows under 21 CFR 211.192, and why most of it concerns investigations

In short21 CFR 211.192 is the second most cited section of Part 211, FDA's drug manufacturing rules, on FDA's FY2025 drug Form 483s, the observations investigators give a firm's management at the end of an inspection, with 236 citations by our count. But 227 of those concern investigating a discrepancy, and 9 concern the quality unit's review of records. Faster reading can surface a discrepancy sooner. The finding usually comes from what a plant did next.

How to read the "+171%" you may see quoted

An article in Pharmaceutical Online, January 30, 2025, says that for sterile drug manufacturers 211.192 was the second most frequently observed section in 2024, with "a 171% increase" on 2023, the largest of any section. It uses FDA's Data Dashboard and prints no counts in its text.45

That is a statement about one kind of plant and one pair of years. Across all drugs, FDA's own files show 175 citations in FY2023, 173 in FY2024 and 236 in FY2025.8 ECA Academy was already calling 211.192 the most cited paragraph of FDA warning letters in November 2011, so this is a long-running feature of inspections.46

How often FDA's warning letters name production record review

A warning letter is FDA's notice to a firm it believes has committed significant violations. The letters follow the same pattern as the Form 483 counts.

For context, FDA reports 1,248 drug quality assurance inspections in FY2025.47 These counts tell you what investigators wrote down. They cannot tell you how many review errors were never noticed.

Released before anyone read it: what FDA's letters show about production record review failures

In shortWhen FDA does write about the review itself, the story is rarely a subtle missed entry. Batches ship before the review happens, records are never received, dates and initials are filled in ahead of the work, or a firm says it lacks the people to do the review. An exception method fixes none of those by itself.

Each firm is named only for what its letter says, as of that letter's date. Firms may since have responded or corrected. Quotations are FDA's.

Why reviewers miss the entries that matter

A long, repetitive check on mostly correct material is hard work for any person. Evidence specific to batch records is thin. A 2019 trade article, written without a stated method, gives manual entry errors of about 1 in 100 and right-first-time batches "as low as 47%".9 We use those only as a sign of what people report. Your own record of corrections per batch is better evidence, and most plants already hold it.

Three patterns recur in practice and in the letters: an error that crosses pages, an entry that looks complete and is arithmetically wrong, and a value signed by someone not trained on the step. The first two are good candidates for a rule. The third needs the training system beside the record.

How long a batch record review takes, and why nobody can give you a trustworthy benchmark

In shortThe figures in print range from about 3 hours to 500 for one record, and 10 to 40 days to release on paper. None gives a dataset, a page count or a method. A record's length, whether it is paper, and how many hands it passes through all change the answer. Time five of your own.

The MIT result shows that a large share of the delay can sit in hand-offs and waiting, and that process work alone can cut it. Before pricing any software, walk one record through the building with a stopwatch and see how much of the elapsed time is someone reading.

What late release and missed entries cost a plant

We found no public figure for the cost of a day of release delay or for a missed entry. The cost lines are easy to name: finished goods held while a record waits, shelf life used up before shipping, supply gaps, rework when a problem is found after the line has moved on, and investigations that have to reach back to other batches.

21 CFR 211.192 requires an investigation to "extend to other batches of the same drug product and other drug products" that may be associated with the failure,1 so a discrepancy found late can touch far more than the batch in hand. Part C shows how to put your own numbers on each.

What AI can add to reading batch records, what draft Annex 22 says, and the limits you should plan for

In shortAI can read scans and handwriting, which is what brings exception review to paper records. It also brings risk: a misread value can pass as clean, and the EU's draft AI annex says generative models "should not be used in critical GMP applications". EMA has since reported that comments favored allowing them under controls, and no revised text exists. Whether a step is critical is a classification your quality unit has to make and defend.

Exact rules and reading are two different jobs

Checking a typed value against a limit, recomputing a yield or subtracting two times is exact arithmetic: it gives one answer each time and can be tested like any other software. Reading a smudged handwritten "7" is a judgment by a model, and it may vary. Most of the regulatory worry attaches to the second job, and any reader of a paper record does it.

The evidence on reading executed batch records is thin

We found no independent, peer-reviewed study that measures AI accuracy on executed handwritten batch records against verified truth. The closest paper is a preprint submitted November 26, 2025 by three authors at a company in the field. It ran three records of 15, 44 and 66 pages through a conversion to structured data, and it did not test catching errors against a master record.

It reports composite confidence of 82.08 to 89.00, a score its own workflow produces, and states handwriting accuracy of "approximately 85%" without describing how that was measured, calling it an upper bound on the system's performance.10 That is a warning: on a page with many handwritten entries, an accuracy in the 80s would be far too low to trust without a person checking what matters.

Five limits to plan for when AI reads batch records

• A false clean, a wrong page the system passed, is the failure that matters. A false alarm costs a reviewer a little time. A misread value that lands inside its limit costs a release decision. Test for it with planted errors, on your own handwriting.
• An illegible entry must stop the read. Any tool should flag what it cannot read and never guess.
• The electronic record is the original. EMA's question 16 warns against review based only on printouts.5 If your source is an export, know what it leaves out, audit trails above all.
• Forms and models change. A new form version or a model update can move accuracy. Treat either as a change under your quality system.
• A software tool is never compliant by itself. 21 CFR Part 11, FDA's rule on electronic records, Annex 11 and GMP apply to you, in your use. A vendor can offer controls that support your compliance and you do the proving.

Who reads a batch record at a typical plant, where the days go, and the three ways plants organize the review

In shortProduction reviews first, then QA, then the quality decision, with corrections bouncing back between them. Reading is only part of the elapsed time; queues and correction loops take much of the rest. Plants choose among line-by-line paper review, exception reports from an electronic batch system, and software that reads paper or exports, often using more than one across their products.

The five people who carry one record

Where a reviewer's 31 hours might go on a 300-page paper record

Line-by-line paper, a batch system's reports, or a reader on paper and exports

Market size: we found no audited figure for review by exception, and no credible survey of how many plants still run paper. Market reports we read gave adoption shares with no method, so we leave them out.

Dana Okafor's Wednesday: batch DR26-1109 goes from 304 unread pages to one critical deviation

In shortDana is a QA batch record reviewer. Her home screen draws each executed record as a strip of its pages. A new paper record arrives at 7:40 am. Six agents read it against the master record, and nine pages come back marked. The first she opens is a blend that ran 12 minutes against 15 ± 1, and she raises the deviation herself.

The review bench draws each batch record as a strip of its pages

The bench opens straight on the work. Each line is a record, and each stripe in the strip is one page. Colored ticks mark exceptions by severity, and the list on the right holds the 15 still open for Dana, worst first. The new arrival sits on top, not yet read.

Six agents read every page in place

She presses Read the record. The strip fills from left to right. The page reader confirms 304 of 304 pages against the issued copy and reports how many are handwritten and how many boxes could not be read. The entry extractor turns the pages into 4,270 entries, with signatures, initials, dates and times. The comparator, the documentation checker and the cross-page reconciler then each do their job, and any page that was unreadable would go to a person. The scan itself is never altered.

Nine pages marked out of 304, and six clean pages drawn for her

The strip settles. Nine exceptions on nine pages: one critical, three major, one check and four documentation errors. Of the 295 clean pages, six are drawn at random for her spot-check. Eight of the nine need a decision. The ninth, a check, shows a re-dry that followed the route the master record allows, marked so she can confirm the route was followed.

The final blend on page 101, with the rule quoted beside it

She clicks the red tick and lands on the scanned page with the entry highlighted. Step 7.12 runs from 1:05 pm to 1:17 pm. A band works out the 12 minutes. On the right, the sample master record's own words: "Blend 15 ± 1 minutes at 8 rpm after adding magnesium stearate. Blend time is a critical process parameter." The master record comparator suggests a deviation, and shows its confidence. Three choices wait: raise a deviation, send a correction, or accept with a reason.

A deviation drafted from the record, raised by Dana and signed twice

She confirms. The deviation is already drafted: DEV-26-0631, the batch, the evidence (page 101, master record step 7.12), classified critical, owned by Sam Patel in production with Dana for QA, and the batch on hold until the deviation is closed or assessed. She checks it and raises it into the quality system. Because it is critical, the QA operations lead, Omar Haddad, signs second. An agent has drafted the words. People have made every decision.

A hold time across pages, a missing printout, a gap in the hardness series and a sum written wrong

In shortExceptions are of several kinds, each cited to its page and rule. Some are one entry; some exist only between two pages. Each comes with a suggested decision that the reviewer can overturn, and corrections collect into one list for production.

A second-person check missing on the binder weighing

Next on the list, page 49: 41.4 kg of purified water weighed by J.M., inside 41.5 ± 0.5 kg, with the checker's box empty. The suggestion here is a correction to production. A value inside its limit, with no second signature, is what a rule catches every time and a tired reader can pass.

26 hours 15 minutes that exist only between pages 58 and 81

On the batch before, DR26-1108, the cross-page reconciler paired the end of granulation with the start of drying and showed both pages with the sum between them: October 1 at 9:40 pm to October 2 at 11:55 pm, against a 24-hour limit in step 6.4. The drying page's "N/A" beside the deviation reference tells the rest. No single page shows this.

An active ingredient weight with no balance printout behind it

Page 24 records the active ingredient at 12.006 kg, written by hand. The master record requires the balance printout on the page, and it is missing. The suggestion is to ask production to attach the BAL-07 printout from the weighing log and, if it cannot be found, to raise a deviation.

A hardness check every 30 minutes, and one empty row

Page 158. The press ran from 10:40 am to 2:15 pm without a stop, and the 11:30 am check is empty between two good results. The value that is missing cannot be out of range, so a range check alone would pass the page. It takes a check of the series to find it.

A re-dry that followed the allowed route

Page 86 shows loss on drying of 2.7% against a 2.0% limit, then a 10-minute re-dry and 1.9%. The master record allows one re-dry, so the comparator marks this as a check, with "accept" suggested. Following allowed routes before raising an exception is what keeps false alarms down, and the page stays visible so a person can agree.

99.6% written, 99.1% by the entries on page 291

Both are inside 98.0% to 100.5%, so this is a calculation error with a documentation fix, and the screen says so. The calculation sits under the page: (590,140 + 1,200 + 3,260) ÷ 600,000.

A late entry that nobody marked late

Coating ended at 5:05 pm on October 5. The entry is signed October 6 at 8:20 am, with no "late entry" note. The sample documentation procedure's section 5.8 is quoted, and the suggestion is a correction.

One correction list per batch for Sam Patel

Corrections collect into one list, with page, rule and what to do. Production corrects and rescans only those pages, and the agents re-read only those pages. The original scan stays as it was.

How the 295 clean pages get checked, what a miss does, and what Dana's signature says

In shortClean pages are sampled at random on every record, before it can be signed. A problem found there is logged as a miss against the plant's validation report and the validation lead is told. Critical parameter steps are shown even on a record with nothing to decide. Only a person signs, and the signature shows its meaning, name and time.

Six clean pages, opened one by one

The sample runs at 2% of clean pages by default. "Clean" here comes with its evidence: each page shows the entries the agents checked. If everything agrees, she moves to the next page. If not, I found something starts an exception.

What happens when a spot-check finds a miss on a clean page

A problem found on a clean page becomes an exception of its own, logged as a miss against the validation report, and Lena Ortiz, the validation lead, is notified. The rule that should have caught it goes into the next rule set. Part A's section 6 explains why those misses need pooling across records before they mean much.

A 478-page electronic record with nothing raised

VL26-0428, a batch of Veltrimab drug substance, came through with no exceptions. The screen says "Nothing needs you" and asks only for the spot-check. Critical parameter steps are still listed for her to see. Paper, hybrid and electronic records share one bench.

The signature box and its meaning

Agents cannot sign. Dana's name, the time and the meaning are bound to the record: "Production record review complete — 21 CFR 211.192". That wording is a label the sample company chose in its settings, and it records what her signature means.

The signed review goes to disposition with any open deviations listed beside it. The sample plants supply both US and EU markets, so Dr. Maya Chen, the Qualified Person, receives it for certification. A US-only site would route it to its quality unit decision, and the settings also offer "Batch record reviewed — EU GMP Chapter 4".

Six agents read and check each page; a QA reviewer decides, and a second person signs anything critical

In shortEach agent has one job and a limit it cannot cross. None of them decides. A named QA reviewer decides each exception and signs. The QA operations lead second-signs anything critical, production owns its corrections, the validation lead owns the rule set, and the quality unit, or the Qualified Person in the EU, decides the batch.

The QA reviewer signs each review; anything critical is signed twice

Each agent finding, decision, correction list, deviation, spot-check and signature goes on the audit trail with who, what and when. These controls support your compliance with 21 CFR Part 11 and EU Annex 11, and with your own procedures; the proof in your use remains yours to build. Which steps count as critical under your risk assessment, including any step that uses a language model, is your quality unit's classification. See section 10.

The records list, the documentation error map, the dashboard and the rule settings a plant controls

In shortAround the review sit the screens a QA lead lives in: every record by stage, a heat map of documentation errors, review time and reviewer agreement across plants, and the settings that decide what reaches a person. Each change to those settings is a new, approved rule set version.

Every batch record across three plants

Paper, hybrid and electronic records by plant and stage, with open exceptions by severity, reviewer and release date.

The documentation error map that points training at one shift

All documentation errors by type and area. In the sample, the largest cluster is corrections without initials, date or reason at compression, 17 in the last 30 days, mostly on one shift. One click sends the pattern to training with a date to re-measure. This is what turns a stream of one-at-a-time corrections into a fix.

Review hours, pages opened and how often reviewers agree

The dashboard shows review time per record, what the agents raise, exceptions by area, and agreement with suggestions by agent. The numbers below are design targets, shown on sample data. The live solution measures your own.

What reaches a person, who signs, and what the signature means

Critical parameter steps are always shown by default. The near-limit check, late entry rule, spot-check share, record types, second signature, deviation owner, correction recipients and signature meaning are all settings. Each change is a new rule set version, approved by the validation lead. Check-only rules can add exceptions and never hide one. The "Validated" mark on this screen refers to the sample plant's own report, VR-RBE-0412, on its own rule set, with a challenge set of 40 seeded records: a customer's validation is its own work.

What the agents read, and what the review produces

Systems a batch record review connects to

What a slow or thin batch record review costs your plant, and where to find those numbers in a week

In shortHours are the cost you can see. A release held up by a queue ties up finished goods and shelf life. A correction found on day ten sends a record back through production. A discrepancy found late can reach other batches. Nobody publishes these for your plant, so here is how to measure them.

Where review costs hide on a batch record, and how to size each one

Batch record costs that never reach a spreadsheet

What exception review could return per batch record, with the formula behind every figure

In shortThe return comes from four places: reviewer hours, fewer correction loops, less inventory held, and documentation errors fixed at the source. The calculator covers the first three, with every default marked illustrative. Set your own numbers and subtract the cost of proving the tool works, which a regulated plant must carry. You also gain a measured miss rate to show an inspector, which no calculator prices.

Formulas: time saved = records × (hours today − hours by exception) × hourly cost. Loops = records × round trips avoided × hours per round × hourly cost. Carrying = records × days saved × batch value × carrying % ÷ 365. Net = the three together, less the yearly cost. Every default is an assumption of ours, for you to replace: no public benchmark exists for any of them. Enter a saving you have measured, or you are pricing a hope.

Is your batch record operation ready for exception review? Eight questions, scored as you go

In shortScore yourself 0 to 2 on eight questions about your records, rules, people and inspection position. Nothing you choose leaves this page. A low score usually points to work worth doing anyway: master records that can be checked by a rule, and a documentation procedure written clearly enough to enforce.

What to ask any batch record review vendor about reading, misses and who answers to an inspector

In shortAsk us too. A demonstration on clean sample data tells you little. Ask how the tool does on your handwriting, what it does with an unreadable field, how misses are measured, which steps use a generative model, and what happens when an inspector wants to follow one decision back to its page.

A twelve-week batch record pilot that runs beside your normal review and ends with three numbers

In shortPick one product and one record type. Keep your normal review running, so no release decision depends on the pilot. Plant known errors in copies. At the end, decide on three numbers: errors caught, misses on pages called clean, and reviewer minutes per record.

Catches, misses and reviewer minutes per record

Three of these decide the pilot: planted errors caught, misses on clean pages and reviewer minutes. The others explain the result.

Questions people ask about reviewing batch records by exception

In shortThe questions people search about review by exception, each answered on its own, with the regulators' position stated as we read it on November 3, 2026.