Machine vision / Cost model
AOI False Call Rate: What 1% vs 5% Costs You Per Shift
A false call is the machine flagging a good unit as defective, and somebody has to judge every one of them. On a line inspecting 12,000 units a shift, a 1 percent false call rate produces 120 calls, about 40 minutes of verification. A 5 percent rate produces 600 calls, about 3.3 hours, which across three shifts and 300 days is close to 1.4 full-time operators employed to confirm that good units are good. That is the visible cost. The expensive one is what happens to the operator's judgement in hour three, because a verifier reviewing a pile of mostly-good images is back at the 80 percent human detection ceiling you bought the machine to escape.
The numbers at a glance
01 / Definitions
Three ways the rate gets quoted
Before costing anything, fix the definition, because vendors, operators and quality managers routinely use three different ones and then disagree about the same machine.
The distinction matters commercially. A machine with an 80 percent false-call share sounds catastrophic and may be perfectly manageable at low volume. A machine with a 2 percent unit-level rate sounds excellent and can consume two operators on a high-density board. Insist that any figure in a quotation states its denominator, and insist that it was measured over a full shift after tuning, not at acceptance on a golden board.
One more framing worth internalising: the false call rate, not the detection rate, is the operational constraint on an AOI system. As one analysis of AI-based classification puts it, verification overhead directly erodes the throughput gains that AOI was installed to deliver, and reducing false calls by even a few percentage points reclaims hours of operator time per shift.
02 / The per-shift model
1 percent versus 5 percent, in hours and ringgit
Take a line inspecting 12,000 units a shift, an operator review taking 20 seconds per call, and a fully loaded operator cost of RM 18 an hour. Three shifts, 300 operating days. Everything below is arithmetic from those four assumptions, so substitute your own and the table rebuilds itself.
| False call rate | Calls / shift | Review time / shift | Hours / year | Direct labour / year | Operator-equivalents |
|---|---|---|---|---|---|
| 1% (tuned) | 120 | 40 min | 600 | RM 10,800 | 0.29 |
| 2% | 240 | 80 min | 1,200 | RM 21,600 | 0.58 |
| 3% (drifting) | 360 | 2 hr | 1,800 | RM 32,400 | 0.87 |
| 5% (untuned) | 600 | 3 hr 20 min | 3,000 | RM 54,000 | 1.44 |
Direct labour only, at 20 seconds per review and RM 18 per hour fully loaded across three shifts and 300 days. This is the floor, not the true cost.
The gap between the first and last rows is about RM 43,000 a year in wages, which is real but not dramatic. The reason false calls get treated as a serious operational problem rather than a payroll rounding item is that direct review labour is the smallest component of the cost. Loaded cost models that include handling, disposition and the engineering support behind tuning put a single review nearer USD 1.85 of variable cost per call, around USD 2.21 once fixed disposition support is included. At roughly RM 8 to 10 a call, 600 calls a shift is RM 5,000 to 6,000 a shift, not RM 60.
Both figures are defensible because they measure different things. Use the RM 18 an hour table when you argue about headcount. Use the loaded per-call figure when you argue about whether a tuning project or an AI overlay is worth funding, because that is the cost the project actually removes.
03 / The per-board view
On dense boards the same rate costs an operator
High-density assembly changes the scale of the problem, because the denominator is inspection points rather than units. A board with 8,000 inspection opportunities running at 1,500 DPMO of false calls throws 12 false calls per board, and at 20 seconds of review each that is 240 seconds per board, which on a line running 30 boards an hour is two full-time operators clicking through photographs of good solder.
| False call DPMO | Calls / board | Review time / board | At 30 boards / hour |
|---|---|---|---|
| 250 | 2 | 40 sec | 0.33 operator |
| 500 | 4 | 80 sec | 0.67 operator |
| 1,000 | 8 | 160 sec | 1.33 operators |
| 1,500 | 12 | 240 sec | 2.0 operators |
The published case data lands in the same place. A documented SMT operation reviewing 600 flagged items a shift found 480 of them were false alarms, leaving 120 real defects buried in the pile. Another automotive contract manufacturer running multi-layer HDI boards saw its false call rate climb from 12 percent at process qualification to 30 percent within eighteen months, at which point false-call review was consuming roughly 125 hours a week across three shifts, equivalent to three full-time operators.
Three operators is no longer a tuning issue. It is a second inspection department, created by the machine that was bought to remove the first one.
04 / The real cost
Three costs bigger than the wages
There is a published attempt to total all of it. An Orbotech case model calculated the approximate annual cost of false detection on PCB inspection at nearly USD 7,000 per panel annually, on a linear model, so two extra false calls per panel doubles the figure. Whatever the exact number on your line, the shape is the same: false calls scale linearly with programme drift, and nothing about that curve flattens on its own.
Note the asymmetry that makes this hard. Escapes are visible, because customers report them. False calls are invisible, because they look like the machine doing its job. So the pressure on a quality engineer after an escape is always to tighten thresholds, and the cost of tightening never lands anywhere that gets reported. That is the mechanism by which a 12 percent programme becomes a 30 percent programme in eighteen months.
05 / Why it drifts
Five causes, all fixable
False call rates are not a fixed property of a machine. They are a property of the relationship between a machine, a programme and a process, and all three move.
| Driver | What it does | Fix |
|---|---|---|
| Lighting drift | Solder and PCB surfaces are highly reflective. LED ageing and ambient light change how the board looks to the camera, so a programme calibrated under one condition now reads everything as slightly wrong | Routine LED intensity calibration, enclosed conveyor sections to block ambient light, lens and diffuser cleaning, and trending of calibration-board results |
| Post-reflow warpage | A board that was flat during programming can bow 0.5 to 2 mm after the oven, shifting focus and measurement reference so components appear displaced | Programme against post-reflow boards, improve support and fixturing, and use 3D height data rather than 2D position where available |
| Threshold ratchet | Every escape triggers a tightening. Nothing triggers a loosening. Over time the programme is calibrated entirely by fear of the last complaint | Treat thresholds as a two-sided decision with a documented owner, and re-measure escape rate whenever a threshold moves |
| Programme age vs revisions | On high-mix lines, product revisions outpace programme updates, so the reference stops matching the board | Tie programme revision to engineering change orders, and budget the engineering hours as a standing cost rather than a project |
| Component shrink | As packages move from 0402 to 0201 to 01005, and from QFP to QFN to wafer-level CSP, the visual signature of a defect changes while fixed geometric rules do not | Move the judgement from rules to a model trained on your own images, which is what deep-learning classification is for |
Notice that four of the five fixes are maintenance and discipline rather than capital. Before anyone approves a new machine to solve a false call problem, the calibration log and the programme revision history are worth reading. A surprising share of 5 percent programmes are 1 percent programmes with an ageing light source and eighteen months of unrecorded threshold changes.
06 / Cutting it properly
Fewer false calls without more escapes
There is one wrong way to fix this and one right way. The wrong way is loosening thresholds, because every inspection window has a single knob that trades the two errors against each other: tighten it and you catch more real defects while flagging more good ones, loosen it and the false calls fall along with your detection. No setting minimises both. A false call rate improved by loosening is an escape rate worsened in silence.
The right way is to change what does the judging. Deep-learning classification sits on top of the machine's calls and learns which flagged images are genuinely defective, trained on the labelled reviews your operators are already producing. Because it attacks classification rather than sensitivity, it improves both error modes at once. The documented results are substantial: an AI overlay retrofitted onto existing AOI hardware took a contract manufacturer from a 30 percent false call rate to 3 percent, with the escape rate holding at zero over 90 days, and commercial tools report reductions in manual verification of up to 60 percent.
Sequence the work in this order and you will not need to argue about capital until step four. First, measure the real rate over a full shift, by component class. Second, Pareto it, because on most lines a handful of component types generate the majority of calls. Third, fix lighting, fixturing and the programme for those few classes. Fourth, if a residual rate remains on classes that rules cannot separate, add classification. The step-by-step method is set out in reducing AOI false calls without letting defects escape and in your AOI machine cries wolf.
Whatever you change, measure the escape rate in parallel. A false call project without an escape measurement is not an improvement project, it is an unmonitored risk transfer.
07 / On your parts
Trained on your images, tuned on your line
The reason generic recipes produce high false call rates is that they encode somebody else's board, lighting and tolerance stack. The last stretch of accuracy always comes from tuning against the real assembly on the real conveyor, and that work cannot be done from a datasheet.
CODETRACE machine vision and AOI is built on that principle. Models are trained on the customer's own good and defective parts, thresholds are set against your acceptance criteria, and JOVIS, our robotic vision inspection platform, contributes micron-level geometry analysis where a height measurement settles what a grey-scale threshold cannot. Systems are deployed, commissioned and tuned on site across Selangor, the Klang Valley and Batu Kawan, by the engineers who scoped them. CODETRACE is a member of the NVIDIA Inception programme.
If you already run AOI, the fastest useful exercise is not a quotation. It is one shift of measured calls with a component-class breakdown. That single data set usually shows how much of your false call rate is recoverable by tuning, how much needs classification, and how much is a fixturing problem wearing a software costume.
08 / Where to start
Measure one shift before changing anything
Log every call for one full shift and record three things per call: the component class, the operator's verdict, and the seconds taken. That gives you the true rate, the Pareto, and the review time you have been guessing at. Multiply it out to a year with your own shift count and loaded rate, and you will have a number that funds its own fix. On most lines the exercise costs one clipboard and one shift of attention.
Then run the escape check alongside it, because the two numbers are only meaningful together. Take the units the machine passed, audit a defined sample at a second gate, and record what got through. A false call rate without an escape rate is half a measurement, and it is the half that makes people feel better rather than the half that protects the customer.
For the wider comparison of inspection methods see AOI vs AVI vs manual QC, for the capital arithmetic what an AOI machine costs in Malaysia, and for what the escapes on the other side of this trade-off actually cost, the real cost of a defect escape. Inspection sits inside a broader factory automation plan, but this is the one number that decides whether the rest of it gets trusted.
An operator three hours into good images is not inspecting. Fix the calls, not the sensitivity.
FAQ / AOI false call rate
Questions, answered.
01What is a good AOI false call rate?
Mature programmes target single-digit false calls per board, or a false-call-to-true-defect ratio close to one. In practice reported false call ratios run anywhere from 30 to 80 percent of total calls depending on programme quality, which means most lines are far from that target. Expressed against units inspected, 1 percent is a well-tuned programme and 5 percent is a programme that has drifted. The number to trust is the one measured over a full shift after tuning, not the figure recorded at machine acceptance.
02What does a 5 percent AOI false call rate cost per shift?
On a line inspecting 12,000 units a shift, a 5 percent false call rate produces 600 calls that a person must judge. At 20 seconds each that is 200 minutes, about 3.3 hours, of pure verification per shift. Across three shifts and 300 operating days it is roughly 3,000 hours a year, close to 1.4 full-time operators doing nothing but confirming that good units are good. At 1 percent the same line produces 120 calls, 40 minutes per shift, and about 0.3 of an operator. The difference between those two programmes is roughly RM 43,000 a year in direct review labour alone, before throughput loss and escape risk.
03Why do AOI false calls increase over time?
Five causes account for most drift. LED intensity ages and ambient lighting changes, so the board looks different to the camera than it did at programming. Boards warp 0.5 to 2 mm during reflow, moving the measurement reference. Thresholds get tightened after an escape and never loosened again. Product revisions outpace programme updates on high-mix lines. And component packages shrink, from 0402 to 0201 to 01005, so the visual signature of a defect changes while the rules stay fixed. All five are fixable, and none of them are fixed by turning the sensitivity down.
04How do you reduce AOI false calls without letting defects escape?
Never by loosening thresholds alone, because that trades false calls for escapes one for one. The route that improves both error modes is classification: a deep-learning model reviews the machine's calls and learns which flagged images are genuinely defective, using your own labelled reviews as training data. Documented retrofits show the scale of what is available. One automotive contract manufacturer cut a false call rate from 30 percent to 3 percent with an AI overlay on existing AOI hardware while the escape rate held at zero, and commercial AI tools report reductions in manual verification of up to 60 percent.
05Who can tune an AOI programme in Malaysia?
CODETRACE tunes and deploys machine vision and AOI inspection on site across Selangor, the Klang Valley and Batu Kawan, for semiconductor, electronics and automotive manufacturers. Models are trained on the customer's own good and defective parts rather than a generic library, which is what separates a false call rate that falls from one that is simply hidden by a looser threshold. CODETRACE is a member of the NVIDIA Inception programme. A one-shift false call measurement and a Pareto of the components producing the calls is usually enough to say what is recoverable.