Insights
Capturing Tacit Knowledge in Manufacturing With AI
A polymer plant loses a line at 2 a.m. An operator with nineteen years on the floor listens to the extruder, adjusts something the manual never mentions, and the line is back in eleven minutes. Nobody writes it down. There is no field for it.
Those eleven minutes are among the most valuable assets the company owns, and they exist nowhere except inside one person. Manufacturing has always known this and never solved it. Siemens estimates unplanned downtime costs the world’s 500 largest companies some $1.4 trillion a year — roughly 11% of revenue — with an hour of stopped automotive production near $2.3 million. Meanwhile Deloitte and the Manufacturing Institute project up to 1.9 million manufacturing roles going unfilled this decade. The expertise walking out the door was never in the documentation. Workshops with 23 practitioners from 14 German manufacturers, published this January, found three barriers: the knowledge is situated, capturing it costs time nobody has, and the systems built to hold it were designed for compliance rather than for thinking.
So most organizations gave up on the knowledge and bought sensors instead. Predictive maintenance is worth doing. But it answers the narrow question — what is about to fail — and leaves the hard one untouched: what does the person standing there actually know?
Something has quietly changed: the cost of capturing informal knowledge has collapsed. A language model can take a messy spoken account of a 2 a.m. fix — half-sentence, dialect, hedged, full of “usually” and “unless” — and turn it into something searchable without stripping out the judgment. A European research team is now testing exactly this in a polymer plant: operators log disturbances as small cause-and-countermeasure cases, in their own words, inside the workflow rather than after it, with curation happening centrally. Structured, not sanitized.
This is less a technology story than a governance one. The moment capture becomes cheap, a question arrives that no plant manager has had to answer before: whose knowledge is it?
The Risk of Organizational Dehumanization
There is a version of this that fails. Record the operators, extract the patterns, feed a model, and the operator becomes a training input. That is not knowledge management. Research on organizational dehumanization — including a scale validated this July across 581 respondents, where instrumentality emerged as one of two clean factors — describes exactly this experience, and links it to lower engagement, higher turnover intention, and reduced discretionary effort. If contribution is what you want, instrumentality is a strange way to ask for it.
And there is a version that works, which differs in one detail. The knowledge keeps its author’s name on it.
Human Dignity as a Business Strategy
The operator is not a data source. She is the reason the case exists. Her name stays attached. When someone on the night shift solves a problem using her micro-case, she hears about it. When her method turns out to be wrong under new conditions, she is the one asked to revise it. The record becomes a body of work — something a person can be proud of, argue about, and build on. Which is exactly the behaviour the plant needs and cannot mandate.
This is what I mean when I call human dignity a business strategy rather than a sentiment. Attribution is not a courtesy in this design; it is the mechanism. It is what converts a one-time extraction into a nineteen-year contributor who keeps contributing, and lets the person joining next year learn from someone rather than from a database.
Every organization I have looked at holds intellectual capital it cannot see, mostly among the people furthest from its documentation. The tools to see it have finally arrived. What has not arrived is agreement on what we owe the people whose knowledge we are about to be able to read.
I suspect that question, not the technology, will decide who actually benefits.