A few months ago, someone on our team asked a question that stuck with me: "Why does everyone keep re-asking the same five questions about Line 3?"
Turns out they weren't wrong. Shift leads, new hires, even a couple of engineers kept pinging the same senior operator for the same information: torque specs buried in a PDF from 2019, a troubleshooting step someone wrote on a whiteboard once and never documented, the reason a changeover takes longer on Tuesdays. None of it was secret. It was just scattered, and the one person who held most of it in his head was two years from retirement.
So we built something. Not a chatbot bolted onto a help desk. An internal AI assistant trained on our own documents: SOPs, maintenance logs, quality reports, a few engineering drawings we finally got around to digitizing. The goal was small and specific: let anyone on the floor ask a plain question and get a straight answer, without waiting on someone else's memory.
What we expected vs. what happened
We expected the maintenance team to use it the most. They did, but not in the way we predicted. Instead of asking "how do I fix X," most queries were "has this happened before, and what did we do about it." The assistant became less of a manual and more of an institutional memory, something none of our existing systems were built to hold.
Quality started using it differently too. Instead of flipping through old inspection reports to check whether a defect pattern had shown up before, someone types the question and gets pointed to the actual report, with a summary. Small thing. Adds up over a shift.
The part that surprised me most: adoption from newer employees was faster than from veterans. Makes sense in hindsight. A twenty-year operator has the answers in his head already. Someone six weeks into the job doesn't, and doesn't want to interrupt three people to find out.
What didn't work at first
Our first version was too general. We fed it everything we had and let it answer anything, and the answers were technically correct but often useless, too generic to act on. It took a real narrowing pass, tightening it to our actual documents, our actual terminology, our actual line numbers, before it started sounding like it belonged in our plant instead of a generic AI demo.
We also underestimated how much people distrust a system that "sounds too confident." A few early answers were phrased with more certainty than the underlying documentation actually supported, and that eroded trust fast on a floor where being wrong costs real time. We ended up building in a simple habit: the assistant cites which document an answer came from, every time. That one change did more for adoption than any prompt tweaking.
Where we landed
It's not replacing anyone's judgment, and it was never meant to. What it's replaced is the fifteen-minute detour to track down someone who happens to remember, or worse, the guess made because nobody was around to ask. Reporting and documentation work that used to eat hours now takes a fraction of that, mostly because the answer is already sitting there instead of buried in someone's inbox or a binder in the supervisor's office.
We wrote up the fuller build process, including what we'd do differently, here.
Curious how other manufacturing teams are approaching this. Are you building internal tools like this yourselves, buying something off the shelf, or still relying on tribal knowledge and hoping nobody retires too soon?
Top comments (0)