What AI does not know about your production environment (and how to fix it)

AI models seem to know everything. They write text, summarise reports, and answer questions on almost any topic. Yet they often stall the moment you point such a model at your own shop floor. It gives an answer that is correct in general, but not for your machines, your customers, and your exceptions.

The reason is simple: an AI model knows the world, but it does not know your production environment. This article explains what a model structurally does not know about your company, why that is a problem precisely on the shop floor, and how to solve it.

What an AI model does know

A large language model is trained on enormous amounts of publicly available text: manuals, standards, professional literature, forums, and general theory. That lets it explain well how a process works in general, which parameters exist, and what the common causes of a particular defect are.

For general questions, that is powerful. The model draws on the recorded knowledge of an entire sector. The problem arises the moment the question becomes specific to your company.

What an AI model does not know about your production environment

Everything unique to your shop floor that was never written down falls outside the model’s reach. Think of:

• The quirks of a specific machine, for example an older machining centre that drifts slightly at higher temperatures.
• Customer-specific adjustments you have applied for years but that appear in no manual.
• Exceptions experienced operators learned the hard way: if this happens, do that.
• The sequence and feel with which a process runs reliably, knowledge that lives in hands and heads, not on paper.

No model is trained on this, because this knowledge does not exist in text form. It lives in the people who have done the work for twenty or thirty years.

Tacit knowledge: the knowledge that exists nowhere

This experiential knowledge is called tacit knowledge, also known as tribal knowledge. The philosopher Michael Polanyi summed it up in the phrase: “we know more than we can tell” (Polanyi, The Tacit Dimension, 1966).

Tacit knowledge is hard to make explicit. An experienced operator knows when to deviate from the standard, but cannot always put that into words without being questioned about it. Precisely because this knowledge was never written down, an AI model cannot reach it. The model does not see the outcome of that knowledge in its training data and therefore cannot reproduce it.

This is the difference between explicit knowledge, which is written down, shareable, and findable, and tacit knowledge, which consists of experience, feel, and context. Models are good at the first and blind to the second.

Why this is more dangerous on the shop floor than elsewhere

When a model encounters a missing fact, it does not stop. It fills the gap with the most probable answer. That is called hallucination. In a chat conversation a made-up answer is annoying. On the shop floor it is expensive.

A model that invents a setting, a tolerance, or a work sequence sounds just as convincing as a model that really knows. You only notice the difference when there is downtime, rejects, or a customer complaint. A new operator who cannot verify the answer takes it at face value.

How to solve what AI does not know

The solution is not a bigger or smarter model. It is a reliable knowledge base from your own environment that you ground the model on. That happens in three steps.

• Capture. Get the tacit knowledge out of the heads of your experienced people. The pitfall is that this always fails the moment it creates extra work. Operators have no time and no appetite to type, and there is no computer on the shop floor. Capturing must therefore align with how they already work, for example by speaking during the work instead of documenting afterwards.
• Verify. Not everything that is captured is immediately correct or complete. A human review, where an experienced colleague confirms or corrects, prevents errors and assumptions from ending up in your knowledge base. This is the quality gate that makes the difference between a reliable source and a collection of loose remarks.
• Make accessible. Verified knowledge must be searchable and usable at the moment someone has a question. Only once that foundation is in place can you confidently let AI loose on it. The model then draws on your verified shop floor knowledge instead of a general guess.

Why this matters more in the coming years

Attention is shifting towards AI that performs tasks independently. The more you want to lean on that, the more the quality of your underlying knowledge weighs. An autonomous system acting on invented knowledge makes mistakes faster and at a larger scale than a human ever would.

The companies that preserve and keep shop floor knowledge current now are building the foundation on which every later AI application can rest. That becomes even more important when experienced people leave and knowledge transfer before retirement can no longer remain informal. Whoever skips that builds on sand. The technology gets better every year, but it cannot know what no one ever wrote down.

In closing

An AI model knows the world, but not your production environment. You do not close that gap with a better model, but by capturing, verifying, and making usable the knowledge of your own people. Then AI works with what is correct, instead of with what merely seems probable.

Taggl captures the tacit knowledge of experienced operators while they work, with human review as the quality gate. That creates a reliable knowledge base from your own shop floor.