During Production Inspection in China: 50 Units or 200?

A 4,800-unit order. The inspector draws 50 pieces, finds sink marks on 3 of them, and the report says pass.

If the batch actually carries an 8 percent defect rate, a 50-piece sample passes it 42.5 percent of the time.

Not a small chance. Close to a coin flip. Run the same batch against the 200 pieces the standard calls for and the pass rate drops to 6.9 percent.

Both numbers come from the same table, the one in the third section. What they say is simple: a during production inspection in China is not worth much because somebody showed up. It is worth something because of how many units they pulled.

Three Windows, Three Different Things You Can See

Mid-production checks are usually set at three points: 30 percent, 50 percent, 80 percent of the run. Each point sits at a different stage of the process, and each shows you something the others cannot.

At 30 percent, material is in, the first article is done, tooling is on the machine. What you can see is direction — right material, right process sequence, colour and silkscreen facing the right way.

At 50 percent, the line has settled and the first finished units are off. What you can see is stability — dimensions drifting or holding, assembly fit, consistency of surface finish.

At 80 percent, most of the order is done and only the tail remains. What you can see is the batch result — whether defects cluster in one shift, whether packaging and labelling are right, whether the carton count matches.

Three windows are not a repeat of the same job. They are a division of labour. Run only one and you have checked a third of the process.

How Many Units to Check Is a Table Lookup

Sample size is set by two variables: lot size and inspection level.

The default level is General Inspection Level II. It is the default in ISO 2859-1:1999 and the level used on the overwhelming majority of consumer goods inspections. Special levels S-1 to S-4 exist for destructive testing and very low-value lots, and are rarely used.

With the level fixed, lot size gives you a code letter from Table I, and the code letter gives you the sample size and the accept/reject numbers from Table II-A.

AQL 2.5 is the most common setting. It maps to major defects: those that affect function, or visibly affect appearance enough to drive a return. Minor defects usually run at AQL 4.0, stricter categories at AQL 1.0, and critical defects — safety, regulatory non-compliance — at AQL 0, which means not one is acceptable. If you want the sampling logic behind those numbers rather than the numbers themselves, AQL sampling in China covers it separately.

The inspection level is the other dial, and it is the one buyers touch least. Level I draws a smaller sample than Level II and is used where the process is known and the loss from a bad batch is small. Level III draws a larger sample and is used on high-value goods, on safety-critical categories, or on a supplier whose history you do not trust yet. Most buyers never move off Level II, which is fine — but it is a choice, not a rule.

Four lot sizes, looked up:

Reading it: a 2,000-unit lot is code K, sample 125; up to 7 defectives passes, 8 or more rejects.

One thing in that table runs against intuition. Sample size does not scale with the lot. Going from 500 to 5,000 units — ten times the order — moves the sample from 50 to 200, four times. Going from 5,000 to 50,000 — another ten times — moves it from 200 to 500, two and a half times.

That is where the economics of sampling come from. The cost is that any single unit is less likely to be pulled: 200 pieces standing in for 10,000 means each unit has a 2 percent chance of being selected.

281–500 units: sample 50, accept 3

The small-lot case. Three defectives out of 50 is 6 percent, far looser than AQL 2.5.

That is the table working as designed, not an error in execution. Small samples carry limited statistical power, so the acceptance number has to loosen or good batches get killed. The price is a high miss rate, and the next section puts a number on it.

If a small order is quality-sensitive, the fix is not a tighter AQL. It is a higher inspection level, or a full check. At a few hundred units, the labour cost of checking everything is often below the cost of one wrong call.

1,201–3,200 units: sample 125, accept 7

The most common bracket. Seven out of 125 is 5.6 percent.

At this size sampling starts to mean something. A batch with an 8 percent defect rate passes a 125-piece sample 20.9 percent of the time, roughly half the 42.5 percent it gets from 50 pieces.

The condition is that the sample is actually random. An inspector who takes units from the head of the line is drawing from one shift and one material lot. A hundred and twenty-five pieces from one spot is the same as one piece.

3,201–10,000 units: sample 200, accept 10

Two hundred pieces for 10,000, ten defectives allowed, 5 percent.

There is a trap at this size: 10,000 units rarely come off one line. It may be three lines, two shifts, or two buildings. The 200 pieces have to be drawn by line and by shift, or every piece comes from line A on days and the night shift’s problems stay invisible.

Stratified sampling: 67 pieces per line across three lines; half from each shift. Same sample size, coverage fixed.

35,001–150,000 units: sample 500, accept 21

At 50,000 units, a single mid-production visit stops being enough.

The reason is not sample size. It is time. An order that size runs for weeks, and one visit sees one slice of it. The better structure is repeated line walks, or three visits at 30, 50 and 80 percent, with the three slices read as a trend.

The Same 8% Defect Rate, 50 Units vs 200 Units

One batch, an 8 percent defect rate, four sampling plans. Here is the chance each plan passes it:

The table is built from the ISO 2859-1 acceptance numbers. Read across for the same batch under different sample sizes; read down for one plan and how bad a batch can get before it stops slipping through.

Look at the first column. At a 2.5 percent true rate, all four plans pass the batch more than 96 percent of the time. That is the design logic of the standard: when a supplier’s long-run average sits exactly at the AQL, the batch is almost always accepted.

The problem is in the third and fourth columns. At an 8 percent true rate, the 50-piece plan still passes 42.5 percent of batches. The 200-piece plan passes 6.9 percent.

Same bad batch, different sample size: from “nearly half get through” to “almost none do”. The inspector is the same, the factory is the same, the product is the same. The only thing that changed is the number on the paper.

The middle of the table matters too. At a 5 percent true rate, 50 pieces passes 76 percent and 200 pieces passes 58.3 percent. A 17.7 point gap looks modest on one order and adds up to several missed batches across a year.

The table also explains the 0.0 in the bottom right. It is not zero. It is under 0.05 percent, rounded.

What 30% Catches and What It Cannot

30 percent cannot catch packaging problems, and the reason is not that the inspector skipped it. At 30 percent the packaging material may not have arrived, or arrived and not gone on the line. Ask about packaging at 30 percent and the factory says they have not started packing. Look at 80 percent and half of it is packed.

50 percent cannot catch carton counts. Counting happens last.

80 percent cannot catch the early signal of dimensional drift. Dimensions show up at 50 percent. By 80 percent the tool has run tens of thousands of shots, the wear has happened, and the correction is no longer available.

Put the three rows side by side and the conclusion is plain: to change something, be there at 50. To stop something, be there at 80. To confirm direction, 30 is the only look you get.

50% Is the Only Window Where You Can Still Change Anything

Raise a problem at 30 percent and the usual answer is “it isn’t finished, let’s see when it is”. There is something in that — plenty of process deviation corrects itself downstream. But if the material is wrong or the orientation is reversed, 30 percent is the only point where the fix is free. Move to 50 percent and the rework volume doubles.

Raise a problem at 80 percent and the options are down to three, all of them expensive.

50 percent sits in a particular place: the line has settled, the problem has shown itself, and the order is not finished, so the correction is still available. Ask for it there and the factory stops the line and adjusts a parameter, at a cost it can absorb. Ask at 80 percent and stopping the line means stopping the batch.

If the budget covers one mid-production inspection, put it at 50.

After You Find the Defect: Repairable or Not

Finding the defect is the cheap part. Disposition is where the money goes. The fork is in the physical nature of the defect, not in what the factory promises.

The most expensive case is a wrong call: treating an unrepairable defect as repairable, sorting the whole batch, and discovering the sorted pile is scrap. Labour gone, delivery date gone. The test is physical — a sink mark is a moulding parameter and can often be polished out; a dimensional overrun is tooling or equipment and cannot.

One more trap. When a factory offers to sort 100 percent, ask what happens to the rejects. Scrapped means the cost has already landed. Mixed back into the next run means the problem has only been postponed.

The other question is who pays for the rework. The default answer is the factory, and the default outcome is a negotiation. Rework labour on a cosmetic defect is easy to absorb; scrapping a moulded part is not, and a factory facing a large scrap bill will look for a reason the defect is your specification’s fault. Settle the cost rule before the run starts, not after the sort begins. A purchase order that says who bears rework and scrap cost is worth more than an inspection report that says the batch failed.

None of this is exotic. It is the difference between an inspection that produces a decision and an inspection that produces an argument.

What a Mid-Production Report Should Contain

The report is not for you. It is for the factory. If the fields are thin, the factory can argue.

Batch and production order numbers that match the factory’s own scheduling records. Sample size and sampling method — random, stratified, or by shift. Defect counts by class, critical, major and minor separately, not one total. A photo reference number for every defect, with the defect position and a scale reference visible in the frame. The verdict: which AQL, what acceptance number, what was actually found. Line status: which models were running, on which line, across how many shifts. Corrective action and re-inspection conditions, written so that “fixed” has a definition.

Miss the photo references and half the report is worthless. The factory can say they never saw the defect you describe, and you have nothing to put against it.

Miss the sampling method and the factory can say you drew from the bad end. Same position, no answer.

Five Signals That Mean Stop the Line

Some problems do not need the report. Stop on the spot.

One, the same defect appears three times or more in the sample. A single unit can be chance. Three is a trend. On a 125-piece plan the acceptance number is 7, so three is already halfway to the line — wait for seven and it is too late.

Two, the defect sits on a safety-related part. Power section, load-bearing structure, anything a child can reach. There is no acceptance number for these. AQL 0 means none.

Three, the factory cannot explain the cause. No cause means the next batch repeats it, and there is nothing to correct.

Four, the critical operation has no work instruction, or has one nobody reads. That is not a quality signal, it is a management system signal, and it makes every subsequent batch unreliable.

Five, the line is asked to “hold on a second” while you are walking it. Whatever they stop to show you is not the real state.

When Not to Do a Mid-Production Inspection

Not every order earns one.

Sample orders and trial runs do not. At a few dozen or a few hundred units, a 50-piece sample is nearly a full check. Check everything instead — simpler and more accurate.

Rush orders need a trade-off. A mid-production visit takes half a day to a day, plus the report and the back-and-forth, so call it one to two days of real impact. If the whole lead time is ten days, those two days can cost more than the defects would. Put the money into pre-shipment instead.

Pure assembly with no processing has limited value. The quality risk in that model sits in incoming components, and a mid-production visit only shows whether things were put together the right way round. Incoming inspection is the better spend.

A supplier you have run with for years, with no major defects across many consecutive batches, can be checked less often. From every batch to one in three, and move the budget to a new supplier. On a new supplier the logic runs the other way: a factory audit in China before the first order answers questions a mid-production visit never reaches, because by then the tooling and the process are already fixed.

Questions We Get Asked

Mid-production inspection or pre-shipment — which one?

Both, for different jobs. Mid-production buys you time to correct. Pre-shipment inspection in China tells you whether this batch can ship. Pre-shipment alone means the goods are finished when you find out, leaving concession or rework. Mid-production alone leaves the tail of the run and the packaging unchecked. If the budget only covers one, take pre-shipment, and accept that nothing can be changed.

Does AQL 2.5 mean 2.5 percent defects are allowed?

No. AQL is an acceptance quality limit, a statistical boundary, not a permitted defect rate. What it actually means is that when a supplier’s long-run average sits at 2.5 percent, the batch is accepted more than 95 percent of the time — that is the first column of the table above. A batch that passes can still be worse than 2.5 percent, or better. Reading it as “2.5 percent is allowed” is the misreading that lets factories relax.

Can I cut the sample size to save money?

You can, and the miss rate goes up. Going from 200 pieces to 50 moves the pass rate on an 8 percent batch from 6.9 percent to 42.5 percent — six times worse. The cheaper route is not a smaller sample, it is a better bracket: special inspection levels on small lots, stratified sampling on large ones. Same money, better coverage.

Is the inspector arranged by the factory?

If the factory arranges the inspector, the report is close to worthless. When the person drawing the sample has a reason for the batch to pass, the AQL table does not matter. The point of a third-party inspection is not that somebody visited. It is that the person sampling has no stake in the result.

The report says pass and I find problems on arrival. Now what?

Start with the sample size and the defect breakdown. If it shows 125 pieces and 6 major defects against an acceptance number of 7, the batch was genuinely borderline, and you can ask for the rework and sorting records. If it shows 50 pieces and zero defects while the real rate is 8 percent, the problem is the sample size — that plan lets an 8 percent batch through 42.5 percent of the time. Next order, size the sample from the table.

How fast should the report come back?

Draft same day or next day, formal report within 48 hours is the common rhythm in the industry. Anything past three days is usually not a slow writer, it is a conversation that never closed.

Last Thing

Most of the conversation about mid-production inspection is whether to do it. Very little of it is about how many units to pull, which units to pull, and what to do with what you find.

Every number above comes from two places: the ISO 2859-1 sampling tables, and the acceptance numbers printed on them. None of it is estimated.

The inspection itself is carried out by a third-party agency; we do not walk the line ourselves. What we do is the other three things: put the inspection at the right point in the run, fix the sampling standard for your category, and translate what comes back into a decision you can act on — take the batch, rework it, or refuse it. That sits inside our third-party QC inspection service.

More on quality inspection

• AQL inspection explained — how to read the AQL tables, with a worked 1,000-unit example
• AQL sampling in China — how many pieces to inspect for your order size
• Pre-shipment inspection — what we check before your balance payment leaves