That's tribal knowledge: institutional expertise that exists in people's heads rather than in documents or systems. Every manufacturer over a certain age runs on it. The problem is what happens when the heads leave.
What does it cost when knowledge lives in people instead of systems?
It costs you twice: every day in interruptions, and catastrophically at retirement. Day to day, the expert becomes a human help desk: their actual work stops every time someone needs an answer, and the person asking waits in line. A sick day becomes a work stoppage for everyone downstream of that person's knowledge.
Then there's the cliff. When a 45-year veteran retires, four and a half decades of judgment (which configurations work, which customers need what, why the standard is written the way it is) leaves the building in an afternoon. The demographic math makes this worse, not better: the deepest expertise in mid-market manufacturing is the closest to retirement, and replacements stay a fraction as long. Every departure is a withdrawal from an account you can't refill the old way.
How does AI in manufacturing capture tribal knowledge?
By making the knowledge itself answerable, instead of routing every question through a person. AI in manufacturing, in this context, means a system that ingests what your company already has (documents, standards, specs, ERP records, the files on the shared drive) and lets anyone ask questions against it in plain English.
The veteran's role changes from answering the same questions forever to getting their domain into the system once. The new hire's question, "how is this machine configured, and why?", gets answered by the system, with the source document attached. The expert gets interrupted for the genuinely hard cases, which is what their time was for all along.
Who does the capturing? Doesn't this die under its own weight?
It works when ownership is explicit and small. The manufacturer above appointed knowledge masters, one or two people per department, about ten across the company, each responsible for getting their domain's knowledge into the system and keeping it accurate.
That structure matters more than the technology. Capture-everything initiatives with no owner produce a junk drawer. A named person per department, accountable for their slice, produces a working reference that people learn to trust. The veterans, notably, tend to like the role: it's recognition that what they know is an asset, and a say in how it gets recorded.
What changes for new hires?
They stop depending on the org chart to learn the job. A new hire's first months are mostly questions, and in a tribal-knowledge company, every question requires knowing who to ask and being comfortable asking. When the knowledge is queryable, they find answers themselves, and ramp without consuming a veteran's calendar.
The senior people don't disappear from the process. At the manufacturer above, newer service staff ask the system first, then confirm with the senior lead when it matters. The veteran becomes the verifier rather than the bottleneck. That's the right division of labor: the system handles recall, the person handles judgment.
When should you start?
Before the retirement, not after. Capture works by drawing on the expert while they're still in the building: their documents, their corrections, their review of what the system says. Once they're gone, you're reconstructing from fragments.
A practical test: list the people whose departure would genuinely hurt, the ones whose heads hold how things actually work. If the list has names on it, the clock is already running on each of them.