Guides / AI
Which parts of the paperwork you can hand to AI, which parts have to be ordinary code that does the same thing every time, and how to tell the difference before you spend money on either one.
Every week somebody sells a manufacturer an AI that's going to run their paperwork. Then it gets pointed at a real order, gets one number slightly wrong, and nobody trusts it again. The idea wasn't wrong. It just got aimed at the wrong half of the job.
I've spent about 25 years in manufacturing and owned and ran a machine shop for the better part of a decade. I now build the tools I wanted back then, and about ten of them run every day inside a real precision manufacturing operation. So this is a working answer rather than a theory.
Give AI the reading. Give code the arithmetic.
AI is genuinely good at anything messy and written by a human. A vendor's acknowledgment email that's laid out differently every time. A request for quote with the quantities buried in a paragraph. A drawing note. Twenty years of procedures nobody can find. That's the work that used to need a person's eyes, and it's the work AI does well.
What AI won't do is be exactly right every single time. It gives you its best answer, it doesn't give you the same answer twice, and it will not tell you when it's wrong. That's fine when it's reading. It's not fine when the output is a price, a quantity, a due date, or a record going into your ERP. Nobody wants to hear "the AI thinks the quote is nine thousand dollars."
So the jobs split cleanly. Reading a document, pulling out what matters, answering a question about your own procedures, drafting something a person will check: AI. Matching every line on a purchase order against every line on the acknowledgment, adding up a cost, writing a clean record into the system, deciding what gets flagged: ordinary code, written once, doing the same thing forever.
The clearest example I've got is a monitor I built for a precision manufacturer that sends out hundreds of purchase orders a week.
It watches every PO that goes out and every acknowledgment that comes back. When an acknowledgment lands, it checks it against what was ordered, line by line: price, quantity, date, part. It files the PDF where it belongs. Then once a day it sends one email with only the breaks on it, the handful where the vendor came back with something different from what was ordered. Everything that matched, you never hear about.
The only AI in that tool is the part that reads the acknowledgment. Vendors send those back in every format there is, so something has to make sense of the page. That's the messy human job, and AI does it. The rest of it, the comparing, the flagging, the filing, the digest, is structured automation. It didn't need AI and it's better off without it, because that's the part that has to be right every time.
Worth telling you what actually happened when it went live, because it's not what I expected. It was built to catch vendors getting orders wrong. It caught more mistakes made inside the building than outside it. That's usually how these go.
Same shape, different job. Requests for quote land in a shared mailbox. AI reads the request and the drawings and pulls out what it can: quantities, part numbers, descriptions, material, plating, heat treatment. That's the reading, and it's the part that eats an estimator's morning.
Then it stops and hands over. Routing and costing are the estimator's call, because that's judgment about how the part gets made and no software should pretend to have it. And before the quote can go out, a plain check runs over it and confirms every line actually carries the costs it's supposed to. No line goes out un-costed because somebody was moving fast at the end of the day.
The AI never touches a number. It just gets the person to the part only they can do, faster.
There's a middle case worth knowing about, because "AI can't be trusted with exact work" isn't the whole story.
You can put AI on exact work as long as you build something around it. The software scores how confident the read was. The clean, high confidence cases go straight through. The doubtful ones stop and wait for a person before anything commits. You get the speed on everything that's obvious and a human on the ones that aren't, and nothing wrong ever lands in your system quietly.
That's the difference between a tool and a chatbot you're trusting blindly. Same AI underneath. The engineering around it is the whole thing.
The usual story I hear is some version of this: we bought everyone accounts, and nobody uses them.
That's not a people problem. A general AI account knows everything about the world and nothing about your operation. It's never seen your part numbers, your customers, your procedures, your ERP, or the way your quotes actually get built. So the first honest answer it gives is generic, the person who tried it decides it's a toy, and the accounts sit there renewing.
The work that makes it useful is unglamorous. Getting your own knowledge into it, connecting it to the systems that hold your real data, and setting it up so it does a specific job somebody actually has to do on a Tuesday. That's the implementation, and skipping it is why the seats go unused.
Ask what the output has to be. If the answer is a document read, a question answered, or a draft written, that's AI. If the answer is a number, a record, or a decision that has to come out identical every time, that's code. Most real jobs are both, which is why they get built as one tool with AI in exactly one spot.
No. It changes how it gets set up, not whether it can be done. There are government cloud options that keep everything inside the compliance boundary you already operate in, and the work runs there instead. It's a setup question, and it's one I'll answer straight rather than dodge.
No, and you shouldn't. Everything above works on top of what you already run. Ripping out a working system to get better paperwork is the most expensive way to solve the smallest part of the problem.
It starts with a paid Diagnostic, where I map how the work actually runs today and tell you whether AI, a small build, or nothing is the honest answer. If there's something worth building, it's quoted up front. No hourly meter. I host and run it on managed infrastructure that works on top of your systems, and there's an ongoing care plan after that keeps it running and adjusts it as your work changes, with the term and the price agreed between us.
Where to go next. If you already bought AI accounts and nobody uses them, that has its own page: why most AI projects fail in small manufacturing. If the question is whether a drag-and-drop tool would do the job instead, here's where those stop working. And if the pressure behind all of it is that you can't fill the office roles, start there. The fuller picture of what implementing AI in an operation involves is on AI for manufacturers. If your problem is specifically quoting, that one splits the same way and is covered in how to quote machining jobs and on quoting software built around your pricing. And if you want a read on where your own operation actually stands before talking to anyone, the AI readiness assessment is free, takes about fifteen minutes, and nobody has to call you.
Both tools described above are deployed and running daily inside a precision manufacturing operation, along with several others: floor tools, background monitors, and reporting that lands in an inbox instead of waiting to be asked for.
See what I’ve shipped →First call’s free. About 30 minutes, a straight conversation about how the work actually runs today. If AI fits, I’ll say where. If what you need is a small piece of software and no AI at all, I’ll say that instead.