The Financial Times piece on TDK – automating its way around Japan's population crisis – is being read as a technology story. Here is an uncomfortable observation: it is a knowledge story. The robots are the easy half. The hard half is the tacit failure-mode library that clocks out with every retiree, one Friday at a time. I learned that the expensive way.

A stamping line. A maintenance technician with thirty-one years on that press and an intermittent sensor fault nobody else could clear. He retired on a Friday in June. The following Tuesday the line stood for eleven hours before someone found his workaround – pencilled on a folded sheet taped inside his locker door. That sheet was the entire control plan. His leaving date had been on file with HR for years; nobody in quality had ever thought to ask for it.

Most plants can forecast their knowledge losses five years out. To the month. With a name attached. Demographic loss is the only failure mode a plant can predict with certainty – and the one our core tools treat as somebody else's problem.

The most expensive manual in any factory is the one that retires unwritten.

The only failure mode that comes with a date

PFMEA treats causes as probabilistic. You score occurrence, you score detection, you argue for an hour whether the rank is a four or a five. Retirement is not probabilistic – it is deterministic. Occurrence is certain; the only open variables are the date and your detection score, which without a plan is zero. HR holds the dates. Quality holds the process knowledge.

The two lists never meet.

Treat critical-knowledge holders the way you treat single-source components. If a critical chip arrived from one supplier with a dated end-of-life notice, you would already have containment running, a second source under qualification and safety stock on the shelf. Apply the identical logic to the setter who alone can restart the cut-off press: name, retirement window, criticality, containment. A living register, reviewed quarterly.

I led quality initiatives across a 2,000+ multi-site workforce, and the deepest single points of knowledge were never at headquarters. One setter in one plant. One toolmaker three hundred kilometres away. Neither knew the other existed. A register is how distance stops hiding risk.

Thirty years does not fit into two weeks

The standard handover – two weeks of overlap, a deck of contacts and passwords, a farewell speech – moves five per cent of what matters, and that is being generous. Tacit knowledge is pattern recognition: the sound a press makes a stroke before the die crashes, which supplier's lots run heavy in autumn, why the containment from 2017 exists and what breaks the day you remove it. None of that survives a slide deck.

What works is a structured harvest run over months. Pull the old 8D reports and QRQC logs carrying the expert's fingerprints and have them narrate the decisions behind them: why that containment, why that escape point, what they checked first and in what order. Record everything. Then reverse the roles – the successor runs live failures, the expert corrects. The archive is the script. The expert supplies the commentary.

Building the greenfield QA/QC department for 900+ people at SNOP taught me the same lesson from the other direction. There was nobody to inherit from. Every control plan, reaction plan and work instruction had to be written from a blank page, and that exercise shows you exactly how much knowledge in a mature plant has never once been committed to paper. We cut defect costs by 70% in that build-out on an unglamorous mechanism: every decision written down, every line of reasoning traceable.

The QRQC and A3 discipline from Witte Automotive is the second half of the answer. A failure documented the same day, reasoning attached, is portable knowledge. Documented three weeks later from memory, it is folklore.

Make the system, not the person, carry the memory

Same-day capture is what makes expertise leave-proof. QRQC inside 24 hours, A3 for the structural issues, standard work that stays alive – photos annotated in the expert's own words, updated when reality moves. At Witte we ran a Q-Wall: every significant failure on one wall, each card a failure somebody had already paid for. A new engineer could stand in front of it and inherit ten years of scar tissue in an afternoon.

Then change what the audits measure. Training hours are a compliance metric; knowledge coverage is an outcome. Ask the incoming shift why a control exists, what failure it guards against, what they would check first if it tripped. No answer means zero coverage, whatever the matrix says. And a retirement party is not a control plan – it is a celebration scheduled exactly where a verification step should be.

Key takeaways

  • Keep a knowledge-exposure register beside your PFMEAs: for every critical process, the holders by name, retirement window, criticality and containment. Review it quarterly – those dates move in one direction only.
  • Budget six to twelve months of structured harvest, not two weeks of overlap: the expert narrates the old 8D and QRQC archive decision by decision, on the record, then watches the successor run live failures.
  • Enforce same-day capture. QRQC and A3 written while the failure is fresh, with reasoning attached, is the only quality knowledge that survives the person who wrote it.
  • Audit knowledge coverage, not training hours – the test is whether the shift can explain why each control exists, not how many hours it sat through.

TDK's robots will hold the line in Japan. Automation replaces hands, not memory – and the curve is not uniquely Japanese. It runs through Czech tool rooms, German maintenance departments and Polish press shops on the same schedule, a few pages further back in the calendar. It will not negotiate. Plants that manage retirement like any other scheduled failure mode – containment before the date, a control plan around it, verification after it – keep their yield. The rest will rediscover, at full price, every failure their retirees had already paid for.