The Unqueryable Factory

Machine 7 has a name for that sound.

Not in the manual. Not in the CMMS. Nowhere in the MES or the maintenance history or the twelve pages of standard operating procedure that the plant manager keeps in a binder by the workstation. The name lives in Karl, who has operated Machine 7 for nineteen years, who has heard every iteration of that sound across every ambient temperature and every production run, and who can tell you – without looking at a screen – whether it's the compressor cycling normally or the first warning that the vacuum pump is twenty minutes from failure.

Karl retires in September.

Every meaningful conversation I'm having about agentic AI in manufacturing eventually hits the same wall.

The wall is not the model. The model is good. It is not even the integration – though that is real, and I have written about the MES that never got an API. The wall is upstream of all of that. The wall is what you can actually query.

The dominant narrative about agentic manufacturing assumes that the factory's knowledge lives in its systems. A database, a historian, an MES record, a maintenance log. The agent queries. The agent reasons. The agent acts. This assumption works for operational data – shift reports, OEE numbers, scheduled maintenance intervals. It breaks completely at the judgment layer.

The judgment layer is where Karl lives.

29% of the aerospace and defence workforce is over 55. Ten thousand Baby Boomers retire every day. Fortune ran a piece in May calling institutional knowledge "the hidden bottleneck holding back manufacturing" – not machines, not capital, not AI readiness. Knowledge. And 41% of organisations, according to industry research, rarely or never attempt to collect expertise from retiring employees before they leave.

This is not a workforce-planning problem with an AI solution. It is an AI problem that starts before anyone has installed a model.

Karl's knowledge about Machine 7 is not in any system. It never was. It was never meant to be. It lived in the moment of learning – the first time he heard that sound and got it wrong, the second time he understood, the third time he predicted the failure forty minutes before it happened and saved a shift's worth of production. Nobody decided to write that down. Why would they? Karl was right there.

The MES knows the nominal maintenance schedule. It does not know that Machine 7 needs a fifteen-minute warm-up cycle when ambient temperature drops below 12°C. It does not know that the gearbox replaced in 2019 uses a non-standard bearing – SKF 6205-2RSH, not the spec'd 6205-2Z – because the original was backordered and Karl approved a substitute and never got around to updating the record. The MES holds the procedure. Karl holds the exceptions. The exceptions are most of the job.

Here is the tension nobody in the agentic AI discourse wants to sit with.

You can connect your agent to every system in the plant. Give it access to every table in the MES, every sensor reading from the SCADA stack, every work order in the CMMS. And when Machine 7 starts making that sound, the agent will check the maintenance schedule, find no flag, and do nothing.

iBase-t has started embedding AI capture into MES workflows. Glean published a playbook on knowledge transfer from retiring engineers. These are the right instincts. But they are enterprise-grade solutions for organisations that already have the systems. The plant manager with three months until Karl leaves does not have time for a platform deployment.

There is a precedent for this kind of problem. In the 1990s, knowledge engineers spent months in structured elicitation sessions with domain experts – painstaking conversations designed to extract tacit rules and encode them into expert systems. It worked, occasionally. It was slow, expensive, and fragile enough that only the largest organisations attempted it seriously. The knowledge transfer problem is not new. The tools have just changed.

The uncomfortable version of all this is that the agentic factory, as currently imagined, is a building designed around a library that nobody filled. The integration layer is being built. The protocol bridges are being built. The calling surfaces are being exposed. And the most important knowledge in the building was never written down.

The shift that's happening now is not in the model. It's in the capture layer.

Voice transcription accurate enough to reliably structure speech arrived in 2024. LLM extraction capable enough to take a rambling walk-and-talk and produce structured, queryable output arrived in 2025. The combination – sit with Karl, walk the floor, record two hours of conversation, and emerge with a knowledge document that an agent can query and a new hire can read – is something that is possible today and almost nobody is using inside a factory.

Not the formal knowledge-engineering exercise of the 1990s. Something faster, more honest about its incompleteness, designed from the start to be read by agents as much as humans. The output is not a perfect expert system. It is a queryable approximation of Karl's judgment, captured while Karl is still there to correct it.

The factory that gets this right does not need a better model. It needs to spend two hours walking the floor with Karl before Karl walks out the door.

The agent can query the database. It can't query the man.