Anyone who has run a plant knows the shape of a quality engineer's week. Gauge studies Monday, PPAP paperwork Tuesday, a containment report written at ten at night because a customer found what your line didn't. Two stories collided this week, and both end at the quality department's desk. At IMTS, the robotics houses were demonstrating physical AI – machines that route, inspect and correct production with nobody in the loop. The same week, hiring data showed companies recruiting more technology and cyber-security professionals, not fewer, a year of apocalypse stories notwithstanding.

Here is my uncomfortable observation. The routine half of a quality engineer's week is already automatable – including the parts nobody likes to admit are routine. The other half just became the most expensive skill on the market.

What the machine takes, and what it can't

Across two decades in automotive and aerospace plants, I have counted where quality hours actually go. Inspection logging is automatable, and overdue for it. Report assembly – formatting CMM results, pasting welding SPC charts into a customer deck, drafting the skeleton of an 8D – automatable too, and nobody will mourn it. First-pass triage on traceability pulls is where the machine genuinely earns its keep. It reads ten thousand rows without getting bored at row four hundred.

What remains is narrower and far harder. Deciding when the automated answer is wrong, then standing behind that call while production waits. A manufacturing engineer told a reporter this week that the new system "can outthink me". Perhaps. Outthinking was never the scarce commodity on a shop floor. The scarce commodity is the person willing to overrule the machine at shift end, with the plant manager watching, on the strength of a control chart that says drift while the vision system says green.

An answer you can't audit is a liability wearing a confidence score.

The machine produces answers at line rate. Accountability still arrives at human speed. This is not an abstraction, either. Ask any pharmacist who spent this year rationing medication because a process upstream ran on trust instead of doubt.

The profile the hiring data is pricing

The hiring numbers make sense once you see what actually breaks. A drifted AI agent and a compromised one look identical from outside: same confidence, same dashboards, same serene green status. The only person who can separate them is someone who can audit the process and the code path in the same afternoon – who reads a control chart and a log file as the same instrument.

I am that profile, by accident of curiosity. Years as a quality director – SNOP, Witte Automotive before that – then the Certified Ethical Hacker exam, then the thirty-plus security certifications that followed, then a clickjacking flaw in T-Mobile's web estate, responsibly disclosed, which is how my name ended up on their public bug-bounty Hall of Fame. People call that a career change. It isn't. Quality management and offensive security are one discipline taught in two vocabularies: understanding how complex systems fail, then building them so they don't.

I distrust machines professionally. I built MultiPS, an orchestration platform that runs more than 60 AI models in parallel and synthesises their answers into a consensus, precisely because a single model's output deserves the same scepticism as a single supplier's first article. Consensus isn't trust – it's doubt, engineered at scale.

How to write the job ad

Start by deleting the "experience with AI tools" bullet. It filters for people who have watched demos. I can teach a demo in an afternoon. I cannot teach spine.

Building a QA/QC department from zero for a greenfield plant of 900-plus people taught me which profiles survive a launch year, because hiring mistakes that year were paid for in escalations. The survivors were never the candidates with the longest tool lists. They were the ones who asked, in week one, what happens when the system fails – and had a scar from the last time it did.

Interview accordingly:

  • Hand the candidate a dataset with a lie seeded in it – a shifted measurement, a duplicated batch, a capability index that is too clean. Hire whoever finds it and can explain why the data lied.
  • Ask them to walk an 8D in the first half hour and a red-team exercise in the second. Same discipline, different vocabulary.
  • Ask when they last overruled an automated system. "Never" disqualifies. So does a story with no cost attached.

The arithmetic closes the argument. At SNOP, a doubt-first posture – QRQC from day one, layered audits, people paid to say stop – took 70% out of defect costs and held critical customer escalations at zero for a full quarter. Run it backwards. One critical escape – containment, sorting, expedited freight, penalties, the 8D paper trail that follows – burns past €1 million before the dust settles. Call it fifteen years of a good auditor's salary. A plant cutting that chair to save €70,000 a year is underwriting the loss.

Key takeaways

  • Automate the logging, the report assembly and the first-pass triage – and defend the human whose job is deciding when the automated answer is wrong.
  • Replace "AI experience" with evidence of doubt: a time the candidate overruled a system, what it cost, and how they stood behind the call.
  • Interview across both layers – an 8D and a red-team exercise in the same hour – because a drifted model and a compromised one are indistinguishable from the outside.
  • Run the € arithmetic before cutting headcount: one critical escape can cost fifteen years of a good auditor's salary.

Tasks die; disciplines don't. Quality, security and AI engineering are converging on one job description because they were always the same job – knowing how systems fail and building them so they don't. Physical AI will run the line, and mostly it will run it well. But if nobody in your organisation is paid to doubt the machine, the machine will eventually be right about everything except your line.