Does your plant have something like this: a whole cabinet of technical documents, test reports, specifications, and material data, piled up over decades, that only a few senior people know how to navigate — and those people are about to retire?
That is exactly the problem I recently helped a manufacturer work through.
The owner's first question to me was: "If we hand all these documents to AI, can it become a Q&A system, so new hires can just ask it instead of asking someone?"
Sounds reasonable, right? But I have to be honest with you first — nine out of ten people get this step wrong, and they get it wrong in the same place.
What People Expect vs. What Actually Happens
What people expect: throw every PDF into ChatGPT or some AI tool, and it automatically gets smart and answers anything you ask.
What actually happens: you feed it decades of documents in inconsistent formats, mixed with scans and handwriting, and the AI answers wrong with total confidence — in ways you can't even spot.
Why? Because manufacturing documents have three traits AI hates:
- They aren't "articles," they're tables, data, and specs — AI is great at reading prose, but with your material property tables, test data, and tolerance specs, it often reads the wrong row or grabs the wrong value
- The same thing goes by several names — the shorthand the veterans use, the formal name in the spec sheet, and the code your customer requires may all refer to the same material, but the AI doesn't know that
- A lot of the knowledge isn't in the documents at all — it's in the veterans' heads — the document says "this process parameter is X," but not "why it's X, and when you need to adjust it"
If you just dump the documents in, you don't get an assistant that understands your plant. You get an intern who is very good at bluffing.
What I Actually Did
The system I built for that plant went live at this scale: 80+ products and 460+ material properties, all answered through RAG over their real documents, running in production.
But the number isn't the point. The point is the order I did things in — that's what decides whether it's usable:
Step one: not feeding the AI, but figuring out "who will ask, and what" I didn't start with that cabinet of documents. I first asked: where do new hires get stuck most? What do people on the floor most often interrupt the senior staff to ask? Because if you don't know what you need to answer, digitizing more documents is wasted effort.
Step two: breaking "documents" into "structures AI can read" I didn't feed in those tables, specs, and data sheets whole. I first broke them into structured data (which product, which property, which value, from which document), so every answer the AI gives can point back to the exact page of the source document — that way you catch it when it's wrong, and you can trust it when it's right.
Step three: handling "the same thing has several names" I built a mapping layer so the AI knows that the veterans' shorthand, the formal spec name, and the customer's code all refer to the same thing. If this layer is weak, the whole system is useless, because people on the floor don't ask questions using formal names.
Step four: honestly flagging "this isn't in the documents" The most important step: when the answer isn't in the documents, the system has to say plainly, "I can't find this — you might want to ask so-and-so," instead of making something up. A system that can say "I don't know" is a hundred times more trustworthy than one that always has an answer.
What Actually Changed Afterward
Not polished claims like "how much labor we saved." Three very practical things:
- New hires can ask questions without interrupting senior staff every time, and the answers point back to the source documents, so they can verify on their own
- In the years before senior craftsmen retire, what's in their heads has a chance to be systematically drawn out and recorded — that's the most valuable part
- The owner finally knows "what technical assets do we actually have," because to feed them to AI, you first have to take stock of them
If You're Thinking About This Too, Three Honest Points
- Don't buy a "general-purpose AI tool" and expect it to understand your plant. Your knowledge is your moat. General tools can't digest the details of your expertise — someone has to turn your documents into structures AI can read
- Figure out "whose questions are we answering, and what are they" before talking technology. I've seen too many projects with impressive tech that nobody uses, because they answer questions nobody on the floor actually asks
- The value isn't just a Q&A system — it forces you to take stock of what you have. Many plants only realize afterward that the biggest win was finally seeing the technical assets they've built up over decades
Manufacturing knowledge is one of the most underrated assets of this era, and one of the hardest for AI to replace — because it isn't on the internet. It only exists inside your plant.
The question was never "can AI do this?" It's "is there someone who can help you, in the right order, turn the veterans' know-how and that cabinet of documents into something AI can really absorb and the floor can really use?"
I've done this — all the way to production, in daily use on the floor. If your plant has a cabinet like that and you don't know where to start, I'd be glad to talk.