Can AI be used for automated visual inspection of injectable products, and how do you qualify it?

1What the rules allow.

EU, PIC/S and FDA texts set the same bar for a trained model as for any other automated inspection. Where automated methods are used, Annex 1 says the process “should be validated to detect known defects (which may impact product quality or safety) and be equal to, or better than, manual inspection methods”, and the equipment should be challenged “using representative defects prior to start up and at regular intervals throughout the batch”.¹ FDA’s draft guidance says automated inspection machines “should be validated to meet or surpass human inspection capabilities and can be qualified using training standards or artificial intelligence technology”.²

FDA’s AI draft rates AI checking every vial’s fill level as medium risk, because release testing still samples fill volume; its steps are on FDA’s AI guidance.¹² Draft Annex 22 covers only static, deterministic models, though EMA’s June 2026 workshop report says the drafting group has discussed allowing dynamic and probabilistic models backed by “a fully documented, robust, risk-based control strategy”.¹⁵¹⁶ And FDA told Baxter that automated particle inspection “does not supplant the need for 100% manual visual inspection for various other attributes (e.g., cracks)”, so write down what the model judges and what stays with other stations or people.¹⁷

1.1Where a trained model fits.

FDA’s particulates draft leaves out defects inspected alongside particles, such as “container integrity flaws, fill volume, appearance of lyophilized cake/suspension solids”, though Annex 1 still wants every container inspected for “extraneous contamination or other defects”.²¹ It names opaque products, “lyophilized powders, suspension products, tinted vials”, as obvious challenges. Supplemental destructive testing after 100% inspection “may not be warranted” if the technology “is validated to meet or surpass human inspection capabilities”, shown by a feasibility study for the specific product.² Where a rule-based recipe already passes the comparison below, a trained model adds draft Annex 22’s controls without lowering the bar, so test a candidate model the same way before you buy.

2Qualification protocol template.

One protocol can meet Annex 1, which compares the machine with manual inspection, and draft Annex 22, which compares it with the process replaced and wants the plan approved before testing.¹¹⁵

2.1Knapp test procedure and acceptance criteria.

Neither Annex 1 nor FDA’s draft prescribes how to compare a machine with people. The usual method is Knapp’s, described in a 2011 paper FDA’s draft cites.¹⁹

1. Build a challenge set of good units and units with one particle each, rejects never more than 20% of it.
2. Qualified inspectors inspect every unit many times, giving each its own probability of rejection.
3. Sort units into zones: rejected under 30% of the time is acceptable, 30 to 70% grey, above 70% a reject.
4. Reject zone efficiency (RZE) is the probability of detecting reject-zone units; false rejection (RAG) is the rejection of acceptable and grey-zone units.
5. Run the machine on the same set many times and calculate the same figures.

The paper says a replacement must show “a detection probability that is greater than or equal to manual baseline”, and that RAG “is used to set the maximum acceptable level of False Rejection”.¹⁹ So the usual acceptance rule is machine RZE at least the baseline’s, overall and per class, and machine RAG under a ceiling the protocol sets, such as the baseline’s own. FDA has also set a floor for one firm, people and machines alike: “At a minimum, your visual inspection systems should reliably be capable of detecting particulates of 150 microns or greater at a 70% probability of detection.”²⁰

2.2Who does what.

• Process expert: owns the intended use and acceptance criteria, approved before testing, and helps write the test plan (Annex 22 3.1, 4.2, 7.2).¹⁵
• Quality assurance: prepares, approves and controls the test set, reviews the defect library and approves the protocol (FDA draft V.C).²
• Inspectors: confirm each defective unit is detectable under normal conditions before it joins the set (FDA draft V.C),² then form the manual baseline.¹⁹
• Model developers: use only training and validation data; anyone who has seen test data stays out, or pairs with a colleague who has not (Annex 22 6.1, 6.5).¹⁵

For the rest of the draft, see EU GMP Annex 22.

3Example: qualifying a deep-learning inspection model.

This example is invented for illustration. Exemplomab 50 mg/mL injection in 2 mL clear glass vials is filled at the Larchford site, which replaces its semi-automated particle inspection with a machine running a deep-learning model.

Every reject-zone vial was rejected in at least 8 of 10 machine runs. With all 21 fibre vials detected, the one-sided 95% lower bound for that class is 86.7%; for all 90, 96.7%.

On the Annex 22 metrics, sensitivity on reject-zone vials was 97.1% (874 of 900) and specificity on good vials 98.4% (6,691 of 6,800). Every criterion was met, and heat maps for every rejection were centred on the particle, not on the vial shoulder or a reflection. QA approved the report on 21 April 2026 and the model was released under change control.

Worth copying: scope written down, so nobody assumes the model finds cracks; grey-zone results reported, so heavy grey-zone rejection, the buffer the paper describes, shows beside lower overall false rejection; and distinct vials counted per class, because ten runs of one vial are not ten vials.

4What an inspector asks to see.

Expect these records at a pre-approval or routine inspection.

1. Protocol and report, with Knapp data per vial and baseline: Annex 1 8.32, FDA draft V.C, draft Annex 22 7.2 to 7.4.¹²¹⁵
2. Test-set and challenge-kit records: vial inventory, particle size and material, label checks, QA approval, access log: FDA draft V.C; draft Annex 22 5.3 and 6.2.²¹⁵
3. Routine challenge records for each batch: Annex 1 8.32 and the written records 211.68(a) requires.¹⁷
4. Investigations of failed challenges and AQL failures, extended to every unit and batch in doubt: FDA draft V.E, citing 211.192.²
5. Model version, checksum and change history, with each retest decision: draft Annex 22 10.1 and 10.2; FDA’s life-cycle maintenance plan.¹⁵¹²
6. Reject trends by defect type and monitoring metrics, adverse trends assessed for marketed product: Annex 1 8.33; draft Annex 22 10.3 and 10.4.¹¹⁵
7. The role list for changing recipes, thresholds, the cut-off and model versions, with the audit trail. FDA’s data integrity Q&A names “automated visual inspection records” among records only authorised people should change under 211.68(b), and recommends restricting “the ability to alter specifications, process parameters, data, or manufacturing or testing methods by technical means where possible”.¹⁸⁷

5Retraining and model updates.

New training data, labels, pre-processing and the confidence cut-off are all changes. Draft Annex 22 asks for every change, “including any change to physical objects the model is using as input”, to be evaluated for retesting, and a decision not to retest should be “fully justified”.¹⁵ The model in use does not learn: production rejects are labelled, trained on offline and released as a new version.

6What inspectors cite.

“We encourage the use of automated visual inspection for particulates to augment the 100% manual visual inspection program. Automated methods should be rigorously studied for their capability and robustness under various conditions, machine settings, container-closure sizes, defect types, and other variables.” Baxter Healthcare Corporation, warning letter 320-23-18, 25 July 2023¹⁷

The Form 483 behind that letter recorded that “The existing machine recipes were found to be inadequate for further qualification work”, yet they stayed in use until 13 October 2022.²¹ In 2026 FDA found Fresenius Kabi’s systems “often exhibited low detectability and high variability regardless of particle type”.²⁰

“Particulates in challenge kits were excessively large (e.g., greater than (b)(4) microns), approximately (b)(4) times larger than what operators should be capable of reliably detecting.” Baxter International Inc., warning letter 320-26-132, 29 September 2026²²

A model qualified on that kit inherits its blind spot.

“Specifically, we found that your SCADA system had user accounts with extensive administrative privileges, including the ability to modify/edit critical process parameters and recipes without adequate oversight.” Immacule Lifesciences, Form FDA 483, 18 September 2025²³

That was a filling line; inspection recipes and model versions need the same lock. For inspector qualification and AQL sampling, see visual inspection of injectables.

7How it runs in Seal.

In Seal, the visual inspection blueprint gives automated inspection a recipe and evidence state: cameras, lighting, algorithms, thresholds, software and recipe versions, setup checks and challenges. Each production population references the released recipe actually used, and a failed challenge identifies the time window and population inspected under an unacceptable state before reinspection or investigation is routed. The defect library keeps provenance, approval and effective dates, first-pass findings are kept through reinspection, and the inspection machine can stay the source for its images and classifications. It is part of Seal’s records for sterile injectables.

Seal’s computer system assurance records link a regulated system’s intended use, function risk, supplier evidence, configuration baseline, testing, anomalies and release, and confirm that production matches the released baseline before regulated use. An inspection system’s software and model versions can be released the same way, each change assessed from what moved.