The annual product quality review, explained.

An annual product quality review asks what a year of batches says that no single batch can

One tablet's year, from the first batch to the signed review

1. Batches are made and released186 batches over twelve months. Each is tested and released on its own. A supplier changes in February and a sealing temperature in March, each under its own change control, the approved record of a planned change.
2. The review period closesThe twelve months end on July 31. The product's clock starts, because the plant's own procedure wants an approved review within 90 days.
3. The year is collected and trendedRecords come out of the batch, laboratory and quality systems. Each attribute is charted across the year, tested for a shift, and counted against a second system.
4. People decide what the trends meanA corrective action, a watch, or no action with a reason. Each decision is written down with the evidence behind it.
5. The review is signed and feeds forwardAuthor, Qualified Person, head of quality. Actions go to the quality system, and next year's review checks whether they worked.

Step 3 and the decisions that rest on it are the subject of this guide. The people, company and figures in the example are fictional.

Four surprises about the APR and the PQR, starting with the missing deadline

The texts behind the first and third cards. We read 21 CFR 211.180(e) and FDA's 1995 preamble, EU GMP Chapter 1 and its draft revision, Annexes 15 and 16, ICH Q7 and Q10, FDA's process validation and quality systems guidances and its inspection program 7356.002, Health Canada GUI-0001, WHO's main GMP principles and Part I of the PIC/S GMP Guide. None sets a day count for finishing the review or a Cpk of 1.33. We did not read every national rule.

APR, PQR, Cpk, out of trend: the words an annual review runs on

Fourteen terms are defined in the one-page summary.

What 21 CFR 211.180(e) and EU GMP Chapter 1 each say, word for word and with their dates

In shortThe US rule dates from 1978 and was amended in 1995. The EU text in force dates from January 31, 2013. The US asks whether specifications and procedures need to change. The EU also asks about consistency, trends and improvements, and lists twelve inputs. Both texts are in force. A revision of the EU chapter went to consultation in September 2025 and was not adopted as of November 3, 2026.

Names in this section. 21 CFR is title 21 of the US Code of Federal Regulations, and FR is the Federal Register, where US rules are published. EudraLex is the European Commission's collection of medicines rules, and the Annexes are parts of its GMP guide. ICH, the International Council for Harmonisation, writes shared guidelines, and its Step 4 is adoption. PIC/S, the Pharmaceutical Inspection Co-operation Scheme, links inspection authorities. EMA is the European Medicines Agency.

The other texts that shape the trending in a US or EU review

Who else asks. WHO's main GMP principles (2014) and the PIC/S GMP Guide (PE 009-18) each list twelve review items on the EU's lines.22,19 We did not read the current rules for Japan, China or India, so we say nothing about what they require. Check the rule for each market you supply. The annual report filed with FDA under an approved application is a separate document with its own purpose.

APR vs PQR: a 2008 count made the EU list three times longer, and four choices stay with you

In shortA 2008 comparison counted six required items for the US review, eleven for ICH Q7 and nineteen for the EU. A team supplying both markets can write one report to the longer list, as this guide's sample plant does. The rules leave you to decide the period, the grouping, the batches and the deadline. Beyond the lists, your review covers the inputs your quality system defines.

Period, grouping, batches and deadline: four things the APR and PQR rules leave to you

Where a product's year lives: ERP, MES, LIMS, the QMS and the rest

In shortNothing in a plant was built to produce this report. Batches sit in the planning system (ERP), process checks in the shop-floor system (MES), results in the laboratory system (LIMS), events in the quality system (QMS), filings in a regulatory system, agreements in a contract system. An author joins them by hand, and the join is where counts start to disagree.

The sum is 312. With a result for each of 24 attributes per batch where it applies, the numbers behind one review run into the thousands. Part B follows this year.

How to calculate Cpk for an annual review, and how far to trust the answer

In shortCpk compares a process's margin with its spread. The arithmetic takes a minute. The trap is treating the result as exact: 95 batches give a range of about ±0.17 around a Cpk of 1.08, and a split at the wrong date, or mixing both halves of a shift, can hide the story.

Dissolution at 30 minutes in the Dravenzo year, in three lines

Dissolution has only a lower limit here, so we use the lower form. The handbook from NIST, the US National Institute of Standards and Technology, gives it as the distance from the mean to the limit, divided by three standard deviations (SD, the usual spread of results around their mean).5

Both the mean and the spread moved. The margin fell by a third and the spread grew by two thirds. One batch before the change, CV25-1188, was rejected at 78.9%. The first SD leaves it out and says so, so one known failure does not blur the picture.

The range around a Cpk from one year's batches

A Cpk from a sample is an estimate. NIST gives approximate limits: Cpk ± 1.96 × √(1 ÷ 9n + Cpk² ÷ 2(n − 1)).5 For the 95 later batches that is 1.08 ± 0.17, or 0.91 to 1.25, by our arithmetic. The target of 1.33 lies outside the range, so this shortfall is real.

NIST adds that "large enough" is "generally thought to be about 50 independent data values".5 A product made a dozen times a year has a wide range: at 10 batches, a Cpk of 1.33 could be 0.68 or 1.98.

Welch, Mann-Kendall, run rules and c-charts: which trend test answers which question

In shortEach test answers one question: did the center move, is there a steady slope, did the spread change, is this run unlikely by chance, did events spike? Run 24 of them at the usual 5% and you should expect more than one false alarm. FDA asks you to watch closely and also to hold back.

Nine in a row or eight in a row: the run rule names people mix up

A control chart plots results in batch order around a center line, the average. A run rule flags a pattern on it that is unlikely by chance. These rules go by several names. In the manual for SAS, a statistics package, which follows the numbering of Nelson's 1984 and 1985 articles, test 2 is nine points in a row on one side of the center line, and test 3 is six points steadily rising or falling.25 We read that manual and did not read Nelson's articles. NIST's page on the Western Electric rules uses eight consecutive points on one side.26

The manual says such tests add sensitivity at a reasonable chance of a false alarm.25 No text we read makes either set a regulatory requirement. Pick a set, name it correctly in your procedure and apply it the same way to every product.

How to spot a stability result out of trend when no regulator prescribes a method

In shortEU 1.10(vii) asks about stability results "and any adverse trends". No FDA text says how to find one. Industry uses regression methods, which fit a line through earlier results. The common one compares a new stability result with what earlier results predicted, and the example below works it through.

Each product keeps samples on a shelf at controlled heat and humidity, and tests them at set times, called pulls. An impurity usually creeps up, and the question is whether it will stay under its limit until the end of the shelf life. A result can pass today and still surprise you.

FDA's guidance on failing results says, in a footnote, that much of it "may be useful for examining results that are out of trend".27 The journal of ISPE, the International Society for Pharmaceutical Engineering, notes that FDA "does not specify the process for identifying OOT results in stability data", and describes three methods from teams at PhRMA, the US industry group: a regression control chart, a by-time-point method and a slope control chart.28

For the shelf life itself, ICH Q1E compares the one-sided 95% confidence limit of the fitted line with the acceptance criterion.29 ICH is consolidating its stability guidelines. Its work plan of July 10, 2026 sets Step 3 sign-off and Assembly review for Step 4 in December 2026, so check before relying on old section numbers.17

Complaint clusters and two-system counts: the quiet errors in a product review

In shortA flat complaint total can hide five complaints of one kind from one packing line. And a count that differs between two systems means one record is wrong, which is worth knowing before an inspector counts. Both checks are simple and easy to skip when time runs out.

EU 1.10(viii) asks for "all quality-related returns, complaints and recalls and the investigations performed at the time".2 The US rule names complaints, recalls and returned or salvaged products.1 Read as one number, complaints rarely tell you much. In the fictional year there were 58, down from 61, and the month-by-month chart stays under its upper limit of 11.4. So the rate test says nothing moved.

Break the 58 down, though, and the five broken tablets all fall in April to June. All five come from batches packed on one blister line, after its sealing temperature was lowered on March 30. Another chart shows the depth of the blister pocket on that line dropping at the same date, still inside its limits. Neither chart looks serious alone. Side by side they point at one change on one line.

Why the same complaint count can come out 58 and 57

In the example, the complaint system says 58 and the QMS says 57. One discoloration complaint from May is linked in the QMS to the 40 mg strength, yet the batch on the customer's photo is a 20 mg batch. A wrong link in the 40 mg review would also be a wrong count there, so one error touches two reviews.

FDA noted in 1995 that "product complaints and recall information may not be included in the computer data".4 Check what your data feed leaves out, as well as what it holds.

FDA Form 483 citations of 21 CFR 211.180(e), counted year by year

In shortFDA publishes how often its investigators cite each rule on Form 483, the list of observations left at the end of an inspection. Citations of the annual review rule are back near their pre-pandemic level, at 55 in fiscal 2025 by our sum. It ranks mid-table, with the sections around quality units and investigations far ahead.

Missing APRs, thin trending and late PQRs: what warning letters and inspectors record

In shortWarning letters show the floor: firms with no annual review at all, or no proof of one. The UK regulator's examples show the bar: a review with no comment on trends, no stability conclusion and dozens of late reports. We found no warning letter that faults the statistics inside a review by name.

Seven FDA warning letters that cite 21 CFR 211.180(e), read from FDA's own copies

We chose these to show the range, so they do not measure how common each fault is. Most of the seven letters concern over-the-counter or homeopathic products. Each letter records what FDA found at one site at one time. The firm's reply and any later corrections may not appear. In the letters, QU means quality unit.

What MHRA wrote about real PQRs in its 2016 inspection data

These are examples MHRA published from 2016. A 2008 article's authors also report having "seen FDA 483s criticizing the failure to trend raw-material test results".7 That is their own experience.

What a shift that goes unseen can turn into

A review exists to catch problems while they are still cheap. The sample dissolution shift has not failed anything yet. If it keeps sliding, a batch fails. If an aged batch on the market fails, the result is a recall. FDA's enforcement reports show how often dissolution appears in the reason.

How much APQR work exists, and the figures we could not find published

In shortThe volume is simple arithmetic: one review per product, per site, per year. Published figures on time and cost are almost all from vendors. So treat them as a starting question, and measure your own. FDA's inspection activity is the one trend with firm numbers.

Your volume. Count products × sites where each is made or released × one a year. Contract-made products count too: Chapter 1 expects a technical agreement saying who produces the review when the authorization holder is not the manufacturer.2 A company with 30 products on two sites has about 30 to 60 reviews a year, by this arithmetic. Spread evenly, that is one or more every two weeks. Bunched on one date, it is a season.

AI in the annual product review: from FDA's 1995 sentence to draft EU GMP Annex 22

In shortSoftware can collect, count, chart and draft. People judge and sign, as FDA said in 1995. The new question is language models, which draft well and can be wrong in plausible sentences. Draft Annex 22 says they should not be used in critical GMP applications, and it is still a draft. Decide, in writing, where a review sits.

Three design rules that follow from the texts, for any annual review tool

How an annual product review gets written today: one author, seven steps, mostly exports

In shortOne author chases six departments for exports, pastes them into spreadsheets, draws charts, writes chapters and then goes through review rounds. In the fictional plant it takes 122 author hours, most of it before any judgment. There are three common ways to run it, and all three still need a person to decide.

Who touches a product review, from the author to the head of quality

Spreadsheets, a quality system add-on or outside help: three ways to run product reviews

PQR-2026-014: from 186 batches to three decisions and a locked review

In shortSoftware workers with one narrow job each, called agents, collect Dravenzo 20 mg's year from the source systems. They chart 24 attributes, test each one and draft 14 cited chapters. Priya Ranvelle, the product quality lead, decides three findings and clears one wrong count. The Qualified Person and head of quality then sign, 22 days before the plant's due date.

1

The review year shows every 90-day clock at once

Priya authors 17 of the plant group's 24 reviews. On October 7 the dial shows 8 in progress and 16 approved, with arcs for the clocks still running. Talvexa oral solution is 9 days late, waiting for the Qualified Person. Dravenzo 20 mg has 22 days left, 9 of 14 chapters ready, 3 findings to decide and 1 count to reconcile.

2

The whole year fits on one screen

The screens show the collector agents pulling the closed period on October 1 at 6:02 am: 186 batches from ERP, MES, LIMS and the QMS. By 6:24 am the analyst agent has charted 24 attributes in four rows: release testing, in-process controls, process and packaging, events and stability. Four need a decision, two are improving and eighteen are stable.

Drafts are set to build the morning after a period closes. This sample's period ended July 31 and its screens show October 1, so take that date as the sample's.

3

Dissolution's call comes with its test and its batches

Batch mean fell from 94.1% to 89.6%. A Welch t-test on 91 against 95 batches gives a difference of −4.5 points, t = −12.6, p < 0.001. A run of 23 batches sits below the earlier center line. Cpk is 1.08 against the 1.33 target, down from 2.61. The change point falls between CV26-1016 and CV26-1017, on change control CC-26-0412.

The screen's label says Nelson rule 2, while the settings write the rule as 8 in a row. Section 6 explains the two names. A run of 23 trips either.

4

The first batch after the change opens in one click

CV26-1017 was made February 2 with the new supplier's cellulose. It passes everything: dissolution 91.8% against a Q of 80%. So the record alone proves nothing. It shows where the change point sits, and it lets a reader check any value against the source systems. The sample also shows why that check matters: the record reads February 2, a week before the change control's effective date of February 9. In a real review, that gap would be a question of its own.

The analyst also compares the same tablet made under contract at Bracken Pharma Services, which uses only the original supplier. Across its 72 batches, before against after the same date: difference −0.2 points, p = 0.71, Cpk 1.41. Six in-process panels show "waiting for data", because 2 of 6 contract exports are outstanding.

5

Priya opens finding F1 and decides

The draft of chapter 4 carries the attribute table and the finding, with four citations. The analyst recommends a corrective action: add particle size (d50) to the incoming test for cellulose, test dissolution on every batch for six months, and ask MSAT to assess a supplier-specific granulation water amount. Her options are to raise it, watch only, or take no action and give a reason. She raises it.

6

Two systems disagree on complaints, and sign-off waits

The complaint system counts 58, the QMS link 57. The reconciliation checker found CMP-26-0311, a discoloration complaint from May, linked to the 40 mg strength, while the batch on the photo, CV26-1058, is a 20 mg batch. Priya relinks it. Both systems now read 58, and her decision is recorded with her name and the time.

7

Three signatures, in order, and the review is locked

With all three findings decided and the count cleared, signing opens. Priya signs that she prepared the review. Dr. Omar Haddad, the Qualified Person, signs that the review supports continued certification of batches. Dana Okafor, head of quality, approves the review and its actions. PQR-2026-014 is locked as version 1.0 on October 7, with the history drawn from the audit trail.

Forming depth and complaints are two panels behind one finding, F2, so four attributes that need a decision become three findings. The third finding, F3, stays on watch in the sample, on a projection that section 15 questions.

What else the review shows: a packaging shift, a stability pull, traced numbers and your trending rules

In shortBehind the walkthrough sit the other calls, the 14-chapter draft where every number opens its records, the portfolio's findings and actions, a dashboard for the head of quality and the rules your site sets.

Our arithmetic on that projection. A straight line through this batch's six pulls reaches about 0.52%, above the limit, with an upper one-sided 95% confidence limit of about 0.57%, the kind of limit ICH Q1E compares.29 The first five pulls alone give 0.46%. So in a real review, ask which model your stability protocol names, and whether batches may be pooled, before accepting a projection or a watch.

Seven review agents collect and count, and a product quality lead decides

In shortFour agents collect, one trends, one reconciles and one writes. None decides what a trend means. The author decides each finding, a mismatch holds sign-off, and three people sign in order. These controls support your compliance. Your validation and your procedures decide the rest.

Priya, Omar, Dana, Maya and Sam: who decides, checks and signs in the sample

Controls that support your compliance. Read-only source connections. Every number linked to its records. A mismatch that holds sign-off. A challenged call kept with its reason. Signatures in order, then a locked version. An audit trail of each agent run, decision and signature. Roles and maker-checker separation between author, Qualified Person and head of quality. The company's information security is SOC 2 Type II audited and managed to ISO 27001. GMP and 21 CFR Part 11 are not certificates a product can hold.

On draft EU GMP Annex 22. It is a draft, and it says generative AI and large language models should not be used in critical GMP applications.16 A writing step that drafts chapters needs your QA's view on where a PQR sits, and on whether you keep a person responsible, as the draft asks of non-critical uses. Ask us which steps use a language model, and test them on your own year.

From ERP, MES, LIMS and QMS records to a cited, signed APR and PQR

In shortIt reads the systems you already run, through read-only connections, and gives you the year's calls, a cited draft, findings and counted numbers. The default template combines the US and EU reviews, and your inputs, rules and signers are settings. The outcome numbers below are design targets.

It works with what you already run. ERP, MES, LIMS, QMS and complaint systems, regulatory and contract systems, engineering and document systems: the categories in the list above.

What a late or thin product quality review costs your plant

In shortTwo costs sit apart. Author hours are certain, and you can count them. A shift found a year late, or a citation, is rare and heavy, and you cannot predict it. We found no published figure for either, so the numbers that matter are the ones you can measure in a week.

Timing one APR and recounting one year: finding your own baseline in a week

The failed batch, the Qualified Person's doubt and other review costs a calculator misses

What automating APR data collection returns, counted in review hours

In shortPut in your own numbers. Every default is an assumption to replace. The calculator counts time alone and takes away what the software would cost you. It leaves out the value of an earlier catch, because nobody can promise that.

Five places an automated product review can pay, with the formula for each

Your yearly hours back across your product reviews, and what they are worth

Formulas. Author hours freed = reviews × (hours today − hours with automation). All hours freed adds other departments' hours. Value = hours × hourly cost. Test the sums with an example: 12 reviews, 80 hours falling to 40, no other hours, $90 an hour gives 480 hours and $43,200 before the yearly cost. Freed hours are worth their dollars only if people spend them on something you value, such as earlier trending. We left out avoided recalls and citations.

Is your data ready for an automated APQR? Eight scored questions and a reflection

In shortAnswer honestly, 0 to 2 on each. Your score suggests whether to build a baseline, pilot on a closed year, or plan a rollout. Nothing you choose leaves this page.

Cited numbers, stated tests, wrong counts: what to ask an APQR software vendor

In shortAsk every vendor the same questions, us included. Bring a closed year with a shift you already know about. The answers show whether a tool does the statistics in the open, and whether a person stays in charge.

Testing APR automation on a closed product year before you trust it with a live one

In shortPick one tablet or capsule product with a year you have already signed. Ten weeks: baseline, connect, replay against the signed review with planted problems, then one live review in parallel. Decide on measured results.

Five measures that decide an APQR pilot

The first three measures decide the pilot. The last two tell you whether an inspector could follow the result.

Annual product quality review questions: APR vs PQR, frequency, batches, Cpk and software

In shortEach answer stands alone. Regulatory status is as of November 2026.