Walk into almost any mid-sized manufacturer in Michigan and ask where the process knowledge lives. Someone will point at a binder on a shelf, a shared drive folder named "SOP Final Final," and two people who have been there long enough that nobody bothers them with the questions anymore. That is the knowledge base. It is not glamorous, and it is not broken enough to fix when the plant is busy. It just sits there until someone wants an AI agent that can answer shop questions, and then the gap becomes obvious.
Agents do not invent process. They retrieve and apply what you already have. If the binder is unfinished, inconsistent, or only partly true, the agent will be unfinished, inconsistent, and only partly true, only faster and with more confidence. Scraping the SOP binder first is not a romantic digital transformation project. It is the boring prerequisite that keeps the later work from becoming an expensive embarrassment.
What "scraping the binder" actually means
People hear "scrape" and think of a one-afternoon PDF dump into a chat window. That is not the job. The job is to get every procedure that the plant claims to follow into a form a machine can search, quote, and cross-check, with enough structure that a person can tell whether the answer is current.
Start with the physical binders and the digital ones that pretend to be binders. Pull the revision dates. Note which procedures still say "see Jim" in the middle of a step. Note which ones describe a machine you sold two years ago. Note which ones exist in three versions with no owner. You are not digitizing for the sake of digitizing. You are building an inventory of what the plant believes, what the plant actually does, and where those two disagree.
The inventory matters more than the OCR. Optical character recognition will get you text. It will not tell you that the setup sheet on the shelf and the setup sheet in the quality folder describe different torque values. A person has to catch that. The AI work comes after the human has decided which version is real.
Michigan shops already know this pattern from quality systems and customer audits. You do not hand an auditor a pile of unmarked printouts and call it a controlled process. An agent is a worse auditor in one respect: it will not politely ask clarifying questions when the documents conflict. It will pick one and sound sure.
Why agents get requested too early
The agent pitch is easy to like. A supervisor imagines a tablet on the floor that answers "what is the first-piece check for this part" without walking across the building. A plant manager imagines onboarding that does not depend on one senior machinist being free. A vendor shows a demo against a clean sample document set and the room nods. Nobody in that room has opened the real binder lately.
The request arrives before the work because the request is exciting and the binder is not. Cleaning SOPs feels like homework. Agents feel like progress. In a region where labor is tight and tribal knowledge walks out the door when someone retires, the pressure to skip ahead is understandable. Skipping ahead still fails for the same reason every other shortcut fails: the inputs are wrong.
When a shop asks for agents first, what they usually get is a chat interface wired to a partial corpus. The partial corpus answers some questions well and invents answers for the rest by blending nearby documents. Operators notice within a week. They stop asking. The project becomes a slide in a quarterly review, then a quiet budget line that nobody renews.
The shops that get past that pattern treat the binder as the first deliverable. The agent is optional until the binder is honest.
A practical order of work
Begin with the procedures that get used under time pressure. Setup sheets, first-article checks, packing instructions, changeover sequences, and the safety steps people skip when a rush order hits. Those are the documents an agent would be asked about first, and they are also the documents where a wrong answer costs the most.
For each procedure, capture three things: the current approved text, the owner, and the last date someone verified it against how the work is actually done. If you cannot name an owner, you do not have a procedure. You have a rumor with formatting. If the last verification was "before COVID," treat the text as a draft until someone walks the process again.
Then convert. Scan paper. Normalize file names. Put everything in one place with a revision field that people actually update. Strip the "final_v3_USETHIS" naming culture. Keep the language plain. Agents and new hires both do better with short sentences and numbered steps than with paragraphs that try to sound official.
Do not wait for perfection before you stop. Aim for a corpus where the high-traffic procedures are current, the obsolete ones are marked obsolete, and the gaps are listed in a single place instead of discovered during a customer complaint. That is good enough to start retrieval work. It is also good enough that a human can supervise the answers without guessing which binder page the model saw.
Only after that inventory is stable does it make sense to talk about agents, copilots, or voice interfaces on the floor. At that point the vendor conversation changes. You are no longer asking "can your tool chat." You are asking "can your tool stay inside these controlled documents, cite the revision, and escalate when the answer is missing." That is a better question, and it filters vendors quickly.
What changes on the floor when the binder is honest
Supervisors spend less time being the search engine. New hires get a source that matches what the trainer says. Quality stops finding that production used an old printout because the shared drive had three copies. The plant still runs on people. The people just stop carrying the entire process library in their heads.
There is a regional angle here that gets underplayed. Michigan manufacturing is full of shops that have been through enough ERP projects, quality system rebuilds, and "paperless" initiatives to be skeptical of the next software story. That skepticism is earned. Scraping the SOP binder is one of the rare AI-adjacent projects that looks like work they already understand. Controlled documents. Revision discipline. Process ownership. The AI layer rides on top of habits the shop already needs for customers and auditors.
Patience is part of the method. A mid-sized plant does not clear a decade of half-maintained procedures in a sprint. Assign a few hours a week. Finish one cell or one product family before you open the next. Keep a running list of conflicts found and resolved. When someone asks when the agent is coming, show them the inventory and the open gaps. Most people calm down when they see the real constraint.
The short version for owners
If someone is selling you shop-floor agents and they have not asked to see your SOPs, they are selling theater. If your SOPs live in binders, email threads, and the memory of three people, you are not ready for agents, no matter how good the demo looked. Digitize, reconcile, and assign ownership first. Then decide whether an agent is worth building on top.
The binder is not glamorous. It is also not optional. Every useful plant AI project we have seen in this region eventually runs into the same wall: the process knowledge was never written down cleanly enough to trust. Scrape that wall first. The agents can wait until there is something true for them to stand on.