Your batch passed AQL. Your documentation is complete. GMP is satisfied.
None of that tells you whether the 3,000 tablets produced while press station 6 had a chipped punch were individually inspected. They were not. They were produced, packaged, and released inside the statistical confidence interval of your sampling plan—unexamined, undocumented, and indistinguishable from the inspected units in your batch record.
Manual AQL sampling means you are statistically confident. It does not mean you are certain.
AQL is a compliance proxy. It was always a proxy.
FDA 21 CFR Part 211 and Singapore's HSA GMP guidelines require documented visual inspection with demonstrated detection capability at defined acceptance quality levels. AQL satisfies this. It satisfies it because regulators wrote the requirement around a physical constraint: in 1963, when FDA codified current GMP, inspecting every tablet at production speed with documented results was impossible.
AQL was the best available proxy for complete inspection. Regulators accepted it because the alternative was no documented inspection at all. Not because statistical sampling is equivalent to examining every unit. Not because a sample that passes guarantees the batch is defect-free. Because the technology to do better did not exist.
That constraint expired. The compliance artefact did not.
Pharma QA managers who defend AQL as "compliant therefore adequate" are answering a question written for an era when 100% inline inspection was physically impossible. The question has not been updated. The capability has.
Where defects actually escape: the interval between samples
Solid dose manufacturing fails in a characteristic pattern: localised, time-bounded events. A punch chips. A coating pan drifts on one sector. A print head loses registration. These events do not produce one defective tablet. They produce defective tablets continuously until detected.
AQL detects them at the sample point. The sample point is end-of-batch or mid-batch spot check. A chipping event at station 6 may run 20, 30, 60 minutes before the next sample is pulled.
Do the arithmetic. A tablet press running 200 tablets per minute with a 15-minute undetected chipping event produces 3,000 defective units before anyone knows. By the time the sample flags the problem, those units are mixed into the batch. The investigation cost exceeds the product value. The batch record shows: sample pulled, defect detected, corrective action initiated. It does not show: 3,000 units produced during the undetected interval, uninspected, shipped.
That gap is invisible in your documentation. It is not invisible in patient outcomes.
What regulators actually want to see
This is not an argument that AQL is non-compliant. It is compliant. It always will be. The argument is that compliance is a floor, and the floor was set at a height determined by 1963-era physical constraints.
For HSA audits in Singapore and Malaysian Ministry of Health GMP reviews, the question is shifting from "did you inspect to AQL?" toward "what is your inspection coverage?" A 100% inline inspection log—every unit, every decision, timestamped with image evidence—exceeds any AQL record on every dimension a regulator evaluates. Every unit accounted for. Every defect traceable to press station, timestamp, and batch position.
21 CFR Part 11 requires that electronic records be tamper-evident, attributable, and contemporaneous. A 100% inline AI vision system produces exactly this: structured inspection records per unit, model version documented, confidence scores logged, decision rationale stored. The output is not a statistical summary of a sample. It is a complete inspection record for every tablet that left the line.
When a regulator asks "how do you know this batch was clean?"—AQL answers with statistical probability. 100% inline answers with a record for every unit. Both are compliant. Only one is complete.
What changes on a solid dose line
The detection point moves. From end-of-batch sampling to real-time, per-unit. A chipping event is detected on the first defective tablet—not on the 3,000th, when the sample happens to catch it.
False positives drop. Tablet coating thickness shifts across a batch. Capsule colour varies between gelatin lots. Rule-based inspection systems either generate excessive false positives on acceptable variation or get tuned so loosely that real defects escape. AI models trained on representative datasets learn the full range of acceptable appearance—not hand-coded thresholds. Result: 60 to 80 percent reduction in false positives versus rule-based inspection.
The defect taxonomy expands. Chipping, capping, edge chips, surface discoloration, coating defects, print registration errors, embossing defects on tablets. Fill-level verification, cap-body alignment, surface contamination on capsules. Seal completeness, pocket fill, foil deformation on blister packs. All inspected at line speed on every unit.
Existing hardware stays. AI vision is an intelligence layer on existing cameras, conveyors, and triggers. The AI replaces the legacy rules engine. There is no requirement to rebuild the line or install new imaging infrastructure.
The reject rate tells the real story. One API manufacturer running a high-volume solid dose line was operating at a 1.8% tablet reject rate—meaning that share of defective tablets was reaching the packing stage before being caught downstream. After deploying HyperQ AI Vision on the existing line cameras, reject rate fell to 0.08%. A reduction exceeding 95%. The 1.8% figure was not a detection success. It was evidence of upstream inspection failure that 100% inline monitoring eliminated.
Three questions before your next GMP review
What is your inspection coverage rate? Not your batch acceptance rate. The percentage of units individually inspected. If that number is not approaching 100%, you are documenting statistical confidence—not inspection completeness. Know the difference before the auditor asks.
How many units were produced during your last defect event before detection? If the answer is hundreds or thousands—or if you cannot calculate it—your system detects at batch level, not event level. That gap is uninspected product. It has a number. Find it.
Is your traceability unit-level or batch-level? Batch-level satisfies GMP. Unit-level exceeds it. The difference matters when HSA or FDA asks what you can prove about a specific tablet from a specific press station at a specific time.
With 100% inline AI vision inspection, certainty is the baseline, not the goal. Talk to us about what changes on your line.