Manufacturing batch analytics, explained.

Manufacturing batch analytics answers three questions: which batches, where did they go, and why did this one fall

From a lot arriving at the dock to the decision on a suspect batch

1. Lots arrive and get a codeNorthgate's cryomedium CRYO-5512 arrives August 4 with a certificate that says it passed.
2. Batches consume themFourteen bottles go into seven finished batches between August 18 and September 8. The record notes each one.
3. Finished batches shipThree markets, four treatment centers, five doses infused.
4. A question arrives and someone walks the linksOctober 6, 8:42 am: the supplier recalls the lot.
5. Quality decidesHold, quarantine, notify. A named person signs, using what the walk found.

Step 4 is the subject of this guide. The people, companies, lots and figures in this example are fictional.

What surprises people about batch data: one-way reports, self-set clocks, passing tests, FDA's citations

The terms a batch answer is written in: lot and batch, genealogy, where-used, control limit

Fourteen terms are defined in the one-page summary.

What manufacturing batch analytics covers, and where it stops: traces, trends, cited procedures

In shortBatch analytics answers questions that cross batches and systems. It sits beside the reviews that clear one batch for release, and it feeds them. Three kinds of answer cover nearly every question: a trace through lot links, a chart across batches, and the sentence in a procedure that says what must happen next.

Plants already review batches. Someone reads each production record before release, and US rules require the quality unit, the group with the authority to approve or reject,4 to do that for all records "before a batch is released or distributed".5 In the EU a Qualified Person, a named individual, certifies each batch.6 Once a year the plant reviews the product as a whole.

Batch analytics differs in its starting point. It starts with a question, at any time, about any set of batches, often long after release. The same data may serve a release review on Monday and a recall on Tuesday.

Three answers to almost any batch question

Each batch question needs a different set of plant systems

How batch analytics differs from batch record review, the annual review and CPV

A "golden batch" is a related idea: the profile of the best past runs, used as a benchmark for new ones. It is a way of comparing batches, one tool among several, and it needs clean data first.11

Lot genealogy, backward and forward: why a recall needs both walks in one answer

In shortA forward walk starts at a suspect lot and finds the batches it entered and the people or markets they reached. A backward walk starts at a batch and finds what went in. A recall needs both, plus a third move: back to the failed input, then forward again to every other lot made from it, called a sibling lot.

Four US rules that, joined, make a genealogy

A component, in these rules, is any ingredient used to make the drug.12

Join those four and you have the graph, though no US rule calls it one.

Graphs break at the points where a person wrote down a material name and left off its code, where a supplier's lot number was never captured, or where a material changed containers and got a new code. If the code is not on the consumption record, the link has to be rebuilt from weighing printouts and memory.

Three walks over the same links

What a where-used report tells you, and what it leaves out

This is a fictional case. A full genealogy is only as good as its weakest link: if supplier certificates stay as PDFs nobody has read into data, the backward walk stops at the plant's door.

US tissue rules ask for tracking from donor to consignee and from consignee to donor,15 and licensed cell products answer to those rules in addition to Part 211.16

The GMP rules behind a lot trace: 21 CFR 211, EU Chapter 8 and Part IV for advanced therapies, as of November 2026

In shortNo rule says "batch analytics". Many rules assume its results: that any lot can be traced to its batches, any batch to its customers, that an investigation reaches other batches, and that process trends get watched. Most of these texts are years old and still in force. The newer activity is around cell therapies and AI, and much of that is still draft.

The dates that shape batch traceability and AI in GMP

What each text asks of a batch answer, and where it stands

Supplier recall impact assessments: which clocks come from a regulator and which from your own SOP

In shortField alerts run 3 working days, biological product deviations up to 45 calendar days, and EU notifications "in a timely manner". We found no US or EU text that sets a deadline for the impact assessment itself, so a 24-hour clock like the one in this guide's example is a firm's own procedure. Each clock has its own trigger, so a plant should know which ones it is on before the first notice arrives.

A supplier's recall reaches a medicine maker as a letter. An impact assessment is the firm's written answer: what we used it in, where those batches went, what state they are in, what our data says, what else shares the input, and what we propose. EU rules name the work as "the determination of the extent of the quality defect", and say a review of the batch production record, certification record and distribution records may be part of it.1

Check which row applies with your regulatory team. The 24-hour Rapid Alert row binds authorities.

What we read for the impact assessment row, on November 3, 2026: 21 CFR Part 7 on recalls, 211.80, .84, .150, .184, .188 and .192, 314.81, 600.14 and 1271.290; EU GMP Chapter 8, Annex 16 and Part IV; and PIC/S PI 010-5. Of these, only PI 010-5 uses a 24-hour period. We did not read every country's rules, and clocks for a recall itself differ by country: Singapore's regulator asks for notice "no later than 24 hours" before a recall starts.41 We also found no survey of how long firms' own procedures allow.

How FDA classes drug recalls, and how many fell in each class in 2025

FDA sorts recalls by hazard. Class I means a reasonable probability of serious harm or death. Class II means temporary or reversible harm, or a remote chance of serious harm. Class III means harm is unlikely.37 A recall strategy also sets how deep a recall goes (wholesale, retail, or user level) and how effectiveness gets checked.42 FDA's guidance to firms asks for coding "to make possible positive lot identification" and distribution records kept beyond shelf life.43

When a supplier's recall spreads across dozens of product entries

Scoping a supplier recall with numbers: bottles, doses, markets and what a passed sterility test proves

In shortScope comes from arithmetic that must tie: what was received equals what was used plus what is left, per lot, and every batch has a state. Worked through here on the fictional recall. Then a second calculation shows why passing release tests is weak evidence against a container failure.

Does the recalled lot reconcile?

Fictional example. Compare the accounted-for total with the receipt quantity on the inventory record. A gap means bottles that are unaccounted for, which is its own finding. EU rules ask for "the reconciliation between the delivered and the recovered quantities" in a recall.14

Where the seven finished batches went

Why, on our reading, seven passed sterility tests cannot clear a seal failure

All seven batches passed sterility (no growth) and endotoxin (0.12 to 0.31 EU/mL against the sample limit of 1.0). Endotoxin is a substance from the outer membrane of certain bacteria.45 Both passes are real evidence, and it is natural to lean on them. They are also weaker than they look. A contaminant may affect only some containers, and a test only sees the sample it takes. The chart below is basic probability: the chance that a test of 20 units catches contamination when a given share of all units is contaminated.

So regulators treat the passing result as one fact among several. EU sterile rules say the test "cannot be used to assure sterility of a product that does not meet its design, procedural or validation parameters".2 FDA told one firm in September 2026 that it "inappropriately concluded there was no product impact based on overreliance on passing finished drug product test results".46

Two limits on that reading. Annex 1 covers sterile medicinal products and never mentions seal failures or advanced therapies; applying it to a supplier's failed bottle is our reading. The EU's advanced therapy rules stand apart in Part IV,19 which accepts that a sterility result may not be available at release, with mitigation that includes telling the treating physician.14

A good assessment states the result and goes on to the failure mechanism: which containers, what contact with product, what the supplier's retained samples showed, and what the medium is exposed to at the plant.

Batch outliers worked through: control limits, action limits and the raw material lot two low batches share

In shortA control limit comes from the process's own history; an action limit and a release limit come from procedures and specifications. They answer different questions. When a result drops, compare every material lot the low batches used, and check where each sat inside its allowed range. Two batches give you a lead to test, and the proof takes more.

Viability at harvest against a control limit, and the limits a procedure sets

The sample records give the result and the rule, without the mean or the spread. As an illustration, a mean of 89.4 and a standard deviation of 2.0 give 89.4 − (3 × 2.0) = 83.4. The 20-batch average, 87.6%, includes the two low results; without them the other 18 average 89.5%. FDA's process validation guidance asks for trending by trained people and says procedures should "guard against overreaction to individual events as well as against failure to detect unintended process variability".10 Limits here are fictional. Yours come from your own data and filings.

Raw material variability: when a lot inside its specification still moves viability

A specification is a pass mark. A certificate says the lot passed it. Your process, though, responds to the actual value. So the useful question is where a lot sat inside its range, and how that compares with the lots you normally receive. In the example, culture medium MED-24-0917 arrived at 318 mOsm/kg (a measure of how concentrated the medium is), inside its 280 to 320 range and at the very top of it. The two other lots in use arrived at 298 and 301. Every certificate read "pass".

Published work supports the idea in general. A 2025 study of commercial antibody production at one company reports that analysis "identified a single raw material" as the major contributor to variability in productivity, and built models from spectroscopy of that material to predict its effect.47 FDA's process validation guidance lists "quality of incoming materials" among the data to trend.10

In a cell therapy the starting material is also the patient's own, so a low result may have nothing to do with the plant: EU rules acknowledge that out of specification products "are not always attributable to failures in the manufacturing process".14

Yield: the limit written into your own master record

US master records, the approved instructions for making a product,23 must state theoretical yield and "the maximum and minimum percentages of theoretical yield beyond which investigation according to § 211.192 is required".23 The batch record must state "the actual yield" and the percentage of theoretical.8 EU rules ask that "any significant deviation from the expected yield" be recorded and investigated.48

Take a fictional step with a theoretical yield of 40 billion cells and an actual yield of 30 billion: 30 ÷ 40 = 75%. If the master record's range is 60% to 110%, that batch needs no yield investigation, though a slow slide from 90% to 75% over a dozen batches may deserve one, and only a trend will show it.

Why CAR-T manufacturing raises the stakes: one batch per patient, published failure rates and reasons

In shortIn autologous cell therapy, a batch is one patient's cells, so a failure can cost a treatment window. A UK national review found failure in 3.87% of 981 patients, and other reports range from 1% to 13%. The most common reasons are numbers plants measure on every batch.

Autologous means made from the patient's own cells. Cells are collected, sent to a plant, altered, grown, frozen and shipped back. A consulting review puts this vein-to-vein time at three to five weeks for commercial products, with "batch failure rates that can exceed 10%".49

The UK's national panel gives firmer numbers for one disease and one country: of 981 large B-cell lymphoma patients approved for CAR-T over four years, 38 (3.87%) had a manufacturing failure. The same paper cites reports from 1% to 13% of all cases, and up to 25% in non-Hodgkin lymphoma.50 Rates differ partly because people count differently: per patient or per attempt, and with or without products that missed a specification and were still given.

What became of 59 manufacturing attempts for 38 patients

Each of those five is measured routinely. That makes them good subjects for analytics, and it also limits what analytics can do. Starting cells vary with the patient's health and prior treatment, and no query fixes that. A plant can separate what comes from the patient from what comes from its materials and process, and act on the second part.

Rules written for product that is one of a kind

Cost per batch: we found no figure from a dataset we could read. One analyst note says lowering failure rates can save "millions of dollars per batch", with no study named.51 Use your own cost of goods.

Where FDA's Form 483 citations land: investigations and supplier lots, far ahead of lot records

In shortFDA's inspection data point at investigations and supplier controls far more than at lot records. Investigations that stop at the one batch, supplier certificates accepted without proof they can be relied on, and data nobody can reproduce are the common failings. Each of them is a batch analytics question that was answered badly or too late.

A Form 483 is the list an FDA investigator gives a firm's management at the end of an inspection, naming conditions the investigator judges may be violations.52 The counts below are citations from those forms, grouped by the section of the rules each one names.3

The low counts for 211.184 and 211.150 say little about whether traceability works, because they count what inspectors wrote down. Our reading: most plants hold the records, and the trouble starts when someone has to use them under pressure. FDA's file does not test that.

The errors behind late or wrong batch answers

How many drug recalls cite lack of sterility assurance

What batch failures and slow investigations cost: thin public numbers, and the ones to measure yourself

In shortPublic numbers here are thin. A trade article gives $20,000 to $1 million per batch deviation, with no dataset. An automation vendor's paper says 2 to 8 weeks of root cause work. FDA's shortage analysis is solid and shows manufacturing and quality problems behind most shortages. The best number is the one you measure yourself.

That analysis, by FDA's own task force, covers older drugs, many of them generic injectables. FDA concluded the deeper causes are economic, which goes beyond this guide. The chart does show how quality problems travel from a plant to a pharmacy shelf.57 FDA ran 1,309 drug quality inspections in fiscal 2025 with its partners, and rated 18% official action indicated, the most serious of its three outcomes. The other two are no action and voluntary action.58

Where nobody has published a number, measure these yourself

Large language models on batch data: draft Annex 22, FDA's warning letter, text-to-SQL scores and what stays with people

In shortSoftware can join systems, walk recorded links and draft from cited records. Large language models, the AI behind chat tools, also make mistakes, so every answer needs its query or page beside it, and a person who signs. Draft EU Annex 22 says these models "should not be used in critical GMP applications", and defines non-critical uses by no direct impact on patient safety, product quality or data integrity. EMA reports its drafters have discussed a wider scope. Decide which of your uses are critical before you buy.

What regulators have said about AI in manufacturing

How often a plain-language question becomes the right database query

Turning a sentence into a database query, called text-to-SQL, is the step most batch question tools depend on. Researchers measure it. The BIRD benchmark, 12,751 question and query pairs across 95 databases, reports that its most effective model executed correctly 40.08% of the time, against 92.96% for people.66

The benchmark's leaderboard, read November 3, 2026, shows 83.40 on the test set for its top entry with hints supplied, dated October 22, 2026, and 78.49 on the development set. It prints the human score in a row that does not say which set, so do not subtract one from the other.67

Move to messy, enterprise-style work and scores fall hard. For the BIRD paper and the Spider 2.0 paper charted below, we read the abstracts, not the full text.

Two lessons follow for a plant, on our reading. Scores fall as databases get larger and messier, so agreed metric definitions are worth writing first. And even a good score leaves wrong answers, so the query must travel with the answer. A wrong column is easy to spot in a query and invisible in a chart.

Where each AI task sits against draft Annex 22

From a flagged result to a QA decision: how a plant answers a batch question today, and where the weeks go

In shortA process scientist usually does the digging, a data team may run queries, QA decides, and manufacturing has to act before anyone finishes writing. Most of the time goes into pulling and joining data and agreeing what it means. Teams work with spreadsheets, dashboards, a central data group or outside help, and each way has a cost.

The six roles a batch question passes through

Eight hand-offs in one low-viability investigation

Where the days of a three-month investigation go

Spreadsheets, dashboards, data teams, resource planning, consultants or query software: how plants answer, and what each gives up

Cryomedium lot CRYO-5512 recalled at 8:42 am: seven batches traced, a second lot found, one impact assessment signed

In shortThis is the example from the top of the page, inside the product. Lena Ortiz, the process scientist, asks one question. The answer is a genealogy graph, backward and forward, with its sources. An agent drafts the impact assessment from the trace. Lena signs the facts, and Omar Haddad, the Qualified Person, decides every action.

1

One question, walked both ways

Tuesday, October 6. Lena types: "Cryomedium lot CRYO-5512 is recalled: which finished batches used it, and where did they go?" In 2.5 seconds on this sample data, the genealogy specialist, one of the solution's agents, walks back to three supplier inputs on the certificate and forward to seven finished batches in three markets.

Each box carries its own facts: viability at harvest, where the dose is now, any deviation. Amber marks what needs a look. Red marks what failed.

2

Every sentence carries a numbered source

Beside the graph sits the written answer. Five numbered sources back it: the supplier notice, the material movements, shipments and chain of custody, release results, and the certificate for the second lot. Each opens at the passage it came from.

The next step cites the sample firm's own procedure, SOP-QA-0112 section 5.2, which asks for an impact assessment within 24 hours. That clock is this firm's rule. Section 5 lists the regulators' clocks and the texts we read.

3

Two doses can still be held

Lena opens batch CRV-26-0426. Harvested September 8, 88.0% viable, sterility no growth, endotoxin 0.19 EU/mL. Shipped to St Aldwyn's Hospital in Manchester, in storage, infusion due October 8. That date is two days away, so this box matters most. CRV-26-0425 at Klinikum Rheinau in Cologne is due October 14.

The Table view lists all seven with market, center and state, for anyone who prefers rows.

4

The lot the notice did not name

The walk then looked for the failed bottle lot elsewhere. NB-BT-0390 also filled CRYO-5530, received September 22: 12 bottles in the plant's store, none used. A forward report on CRYO-5512 alone would have stopped at the seven batches.

5

The impact assessment, drafted from the trace

The impact drafter reads the trace and the seven batch records, then fills IA-2026-0031: exposure, status, four proposed actions, a proposed risk and a due time of 8:42 am on October 7. The drafter proposes. It names Omar Haddad as the person who decides.

Read the proposed risk with care. "Low, release sterility passed" is the draft's reasoning on this sample. EU Annex 1 says the sterility test "cannot be used to assure sterility of a product that does not meet its design, procedural or validation parameters".2 Annex 1 does not name this case or advanced therapies, so applying it here is our reading. A Qualified Person would weigh that, which is why a person holds this decision.

6

Lena signs the facts; the graph updates

Lena signs with a recorded meaning: "I confirm the facts in this impact assessment are correct". Her signature covers the facts. The decisions on two doses, two lots and three authorities go to Omar. On the graph, both lots now read "To quarantine" and the two stored doses "hold requested".

The 2.5 seconds is what the sample shows on sample data. Time it on your own lots.

The read-only walk behind the CRYO-5512 answer

Two low harvests, one culture medium lot and a 2:00 pm feed: tracing a viability outlier on Line C2

In shortLena's other open question is a trend. The batch data specialist draws viability for 20 batches and shows its query. The genealogy specialist compares every raw material of the two low batches and finds one shared lot, also in a batch now in expansion. The procedure specialist cites what must happen. A drafter prepares the deviation, and Dana Okafor decides its class.

1

A chart that shows its query

Asked for viability at harvest over the last 20 batches, the specialist finds the metric among 312 columns (lab test VIAB-HARV), runs one read-only query, and draws the line with its limits. Average 87.6%. CRV-26-0431 at 68.9% falls under the 70% release limit. CRV-26-0418 at 72.4% falls under the 75% action limit.

2

The one lot both low batches share

On the batch page, the genealogy runs from the patient's starting cells to the frozen drug product, with every raw material lot on it. The specialist compares them with the other outlier. One is shared: culture medium MED-24-0917, from Halvern, at 318 mOsm/kg, the top of its 280 to 320 range.

Below the graph, the viability by medium lot: 89.5%, 89.4%, 70.7%. Two batches make a lead for the investigation, and the screen presents it as a finding to review.

3

Forward, to the batch still in the bioreactor

The same walk turns forward. CRV-26-0436 used the same lot. It is on day 6 of expansion, the stage where cells grow, viability 81.2%, down from 88.0% on day 4, with its next feed at 2:00 pm. This is the part of the answer that can change today's work.

4

What the procedure asks, from the effective version only

From the effective version, the procedure specialist answers: tell MSAT and QA within 4 hours, open a deviation the same working day, keep two 1 mL samples at or below −150 °C. Version 8.0 is in review, so it is not quoted. These limits are the sample firm's own procedure.

5

A deviation drafted; the class is QA's

The drafter prepares DEV-2026-0187 from the genealogy and the cited procedures, with a class of major marked as a proposal. It proposes quarantining the remaining 14 bottles and feeding CRV-26-0436 from another lot at 2:00 pm, once QA approves. Dana Okafor, the QA deviation lead, decides the class. Lena signs as author.

The rest of the batch analytics screens: where-used, seeding density, the batch list, procedures, dashboard, settings

In shortBeyond the two cases, the same screens answer everyday questions: a plain where-used with nothing to act on, a process question about seeding density, the list of batches on the line, the procedure library, a dashboard and the settings that set each specialist's boundaries.

Select any screen to open it larger. The company, plant, people, lots, hospitals and figures are fictional. Limits and procedure text are sample data, written for this example.

What the orchestrator, three specialists and a drafter do, and what the author, deviation lead and Qualified Person sign

In shortOne agent routes each question. Three specialists answer it, from lot links, batch data or procedures. One drafts. None can change a plant system, and none decides a hold, a class or a notification. Every step, read and signature goes on the audit trail.

Who signs on Line C2: author, deviation lead, Qualified Person

Maker and checker stay separate people. Approvers are settings, set by the head of MSAT. Every question, query, walk, answer, correction, signature and decision is recorded with who and when, including which specialist answered and what it read.

The Line C2 team, before and after

A draft that informs a hold on a dose may be hard to place under that test, so whether each use is non-critical in your plant is your quality unit's call, and the final text may change it: EMA reports the drafters have discussed a wider scope.59 A person decides on any batch, whatever the final text says.

Controls that support your GMP compliance on a batch question

What goes into batch analytics, what comes out, and the result seen in user acceptance testing

In shortYour plant systems go in, read-only, along with your procedures and supplier documents. A genealogy graph, a chart with its query, a cited procedure answer and a signed draft come out. The solution page reports one result from user acceptance testing and one design target. Measure both on your own batches.

The plant systems a batch analytics rollout reads from

For each, check whether it offers an API, a file feed or a read replica, and who owns it. Enterprise scale matters here: the sample line holds 1,284 batches, 9,412 lots and 642 procedures.

The result the solution page states, and how to read it

The solution page also gives about $25,000 saved per batch problem caught early, labeled as an estimate. Ask how accuracy is defined on any vendor's figure, ours included, and score your own golden set of questions. See section 22.

Lost batches, late holds, scientists joining tables: the price of a slow batch answer

In shortPublic cost figures are thin, so most of this section helps you find your own. The biggest costs are rare and large: a dose given that should have been held, a sibling lot left in use, a failed batch that a trend could have warned about. The steady cost is senior scientists doing data work.

Timing one recall drill and five investigations to find your own costs

Doses, treatment windows and scientists' attention: costs a batch calculator leaves out

Your return on batch analytics: investigations a year, hours spent and batches caught early

In shortSix levers carry the value. Time released is the easiest to count, and batches caught early are often worth more. Every default below is a placeholder: replace it with what you measured in section 18. The formulas sit under the results.

Six ways faster batch answers pay back

Your hours back on batch investigations and recall drills

A site with 40 investigations a year clears the first two bars and falls short of the third on time alone. A site with five a year should weigh this mainly on recall readiness and early catches, or wait.

Lot links, definitions and sign-off: score your plant's batch data in eight questions

In shortPick the answer closest to your plant today. Eight questions score 0 to 2; the last is for reflection. The meter adds up your score as you go, and each question says why it matters and what a strong answer sounds like.

A low score means fixing lot codes before buying anything

Batch analytics and traceability software: vendor questions on lot walks, queries and who decides

In shortAsk every vendor the same questions, us included. Better still, bring two closed batches and last year's supplier notice, and watch the tool answer something you already know. These questions show whether it walks both ways, shows its work, stays out of your records and leaves decisions with your people.

Piloting batch analytics on one line: a golden set and a recall drill over 12 weeks

In shortPick one line. Agree your test before you start: a golden set of questions with known answers, closed investigations, and a mock recall on a shared input. Run read-only beside your current process, with your people checking everything, then decide on numbers you wrote down in week 2.

Scoring a batch analytics pilot: accuracy, sibling lots, hours and overrides

Batch analytics FAQ: lot genealogy, where-used reports, sterility and supplier recalls

In shortShort answers, each of which stands alone. The sections above carry the sources and the detail.