The headcount is denied. The design isn't.

Six hundred inspectors down to forty-eight. That was the number doing the rounds last week, attached to a rumour that Lao Gan Ma – the Chinese chilli-oil maker whose founder scowls from half the world's kitchen shelves – had replaced nearly its whole inspection workforce with AI cameras. The company denied it, then helpfully added that revenue had hit a record 5.4 billion yuan. Read both statements again and notice what neither contains: a single word about defect escape rates. No number, no trend, no promise. I have staffed an inspection organisation from a blank sheet of paper, so let me tell you what that ratio means on a filling line – and why arguing about it misses the point.

What six hundred inspectors are for

Headcount is an org-chart number, not a quality number. Six hundred tells me nothing without the lines underneath it: shifts, product variants, takt. Sizing inspection is arithmetic before it is strategy – takt time, multiplied by control points and shifts, divided by the sampling plan your risk analysis permits. When I built the greenfield quality department for a 900-plus-employee automotive plant, headcount was the last thing I calculated. It was an output of takt, complexity and the escape rate we had promised customers, never an input inherited from an org chart.

That is why "600 to 48" is not a cut. If the number is real, it describes a different job, not a smaller version of the old one. The residual team in a properly automated inspection shop are exception handlers clearing the rejects and no-reads a vision system produces; model supervisors watching for drift and triggering revalidation; MSA and calibration people keeping cameras and gauges honest; and QRQC leads who own root cause when the system flags a cluster. AI flags. Humans solve. Size that team from your failure modes and your false-reject rate and it might land at forty-eight. Size it from the old org chart and it will be wrong in whichever direction you were already prejudiced.

Headcount is what you count when you have stopped counting defects.

Where the machine wins, and where it fails

I have deployed machine vision and I will defend it without sentiment. For hue consistency, fill level, cap presence, label registration – the repetitive attribute checks that make up most food and bottle inspection – a tuned system beats human eyes, and it is not close. It inspects 100% of units instead of a sample and its attention does not sag in hour seven of a shift. A human doing visual checks at a three-second takt lets escapes through no matter how motivated they are. That is physiology, not attitude.

The failure mode is quieter.

A vision model that has lost its footing does not report itself unfit. A new chilli batch shifts the oil colour half a shade, a supplier changes the glass tint, a bank of lights ages – and a single model keeps scoring every jar with perfect, misplaced confidence. A tired inspector tells you they are tired. A drifting model says nothing. This is exactly why, on the multi-model orchestration platform I designed, I refuse to run a single judge: sixty-odd independent models in parallel, consensus synthesis on top, disagreement treated as a signal and routed to a person. In inspection as in security, the dangerous system is never the one that fails loudly. It is the one that fails silently and confidently.

The denial is the real defect

Now the more interesting half of the story: the denial itself. A statement saying "we did not reduce inspectors" answers a labour question and ignores every quality question an auditor or a customer actually asks. Under IATF 16949 or AS9100, nobody counts your inspectors. They open the control plan and read its revision history. They ask which PFMEA justified the automation, and what the validation evidence says – false-accept rate, false-reject rate, attribute agreement, revalidation cadence. Then they want the competence records of the forty-eight people who now own the reaction plans. A press statement is none of those things.

The market context makes that silence expensive. This is the same quarter in which phone makers are replacing handsets over screen defects and buyers of thousand-euro-plus foldables are finding hinges broken before the box was opened. Customers carrying warranty claims like those do not extend benefit of the doubt; when a supplier refuses to explain its system, they price in the worst control plan they can imagine. And the denial arrived stapled to a revenue record, which tells you who it was written for – investors, not customers. Transparency would have been cheaper: publish the escape-rate trend, the model validation, the named owner of drift. Nobody ever ate a press release.

So the next time a supplier brags about AI inspection – or denies it – skip the headcount question. Ask two instead. What is your escape rate, measured and trended, before and after the change? And who owns the model when it drifts – a named person, with a reaction plan and a revalidation cadence? If they can answer both, I do not care whether they employ forty-eight inspectors or six hundred. The system is real. If they cannot, the number was never the point. Rumour or denial, Lao Gan Ma's week contained the same defect: figures with no control plan behind them – the one escape no camera caught.