Every manufacturer I talk to has been told they need to do something about AI. Almost none of them have been told what, specifically, on a shop floor that runs on an ERP from 2009, a spreadsheet nobody wants to touch, and a guy named Dave who knows how the scheduling really works.

So here is the version without the conference talk. Where AI actually earns its keep in a manufacturing business, where it does not, and how to tell the difference before you spend money.

Start where a wrong answer is cheap

The useful way to sort AI work is not by how impressive it sounds. It is by what happens when the answer is wrong.

If a summary of a long spec document is wrong, somebody reads it, notices, and fixes it in a minute. If a quote sent to a customer is wrong, you either eat the difference or have an uncomfortable call. If a number that feeds your production schedule is wrong, you find out three days later when the wrong material shows up.

Start at the cheap end. Work where a person reads the output before it matters is the safest place to learn what this technology is actually good at, and you find out fast whether it saves anyone time.

Where it pays off in a plant

Paperwork that arrives in somebody else's format. Customer POs, supplier certs, spec sheets, RFQs. Somebody reads them and retypes the parts that matter into your system. That reading and retyping is the single most common place AI pays for itself in a manufacturing office.

Writing that nobody enjoys. Work instructions, job descriptions, the email explaining a delay, the first draft of a policy. A first draft in thirty seconds that a human fixes beats a blank page.

Finding things in your own documents. Ten years of quality records, maintenance logs, or customer history that nobody can search. Asking questions of that pile is genuinely useful, and it is the kind of thing that used to require a real project.

Making sense of the data you already collect. Most plants record more than they look at. Downtime logs, scrap rates, machine data sitting in a system nobody opens. AI is good at a first pass over that, and the answer gives you somewhere to point a person.

Where it does not belong

Pricing and quoting. This is the one I feel strongest about, because I tested it. I gave three AI tools the same software project and asked each for one price, twice. The six answers ranged from $42,000 to $128,000, and the cheapest quote included more of what the customer asked for than one that was more than twice the price. Nothing was made up and every number could be defended, which is exactly the problem. The whole experiment is here.

Your quoting has rules. Material plus labor plus setup plus margin, adjusted for the things you know about that customer. Rules belong in software, where the same inputs produce the same answer every time and you can point at the reason. A chat window will give you a different number on Tuesday than it gave you on Monday.

Anything a regulator or a customer audit will ask you to explain. Compliance rules change, and a model trained a year ago does not know that. We built hazmat shipping compliance validation for a distributor, and the reason it works is that it checks against the actual current rules, not against a model's memory of them. That one is written up here.

Anything that has to be identical every time. If two people ask the same question and get different answers, that is not a feature you can run a business on.

The part nobody mentions until it goes wrong

Your people are already using it. Somebody in the office has pasted a customer email, a price list, or a drawing into a free chatbot to get help with it. That is not a firing offense, it is what happens when you give capable people a tool and no rules.

So write the rules down. What is fine to paste, what is not, and which tool the company actually wants used. If you already pay for Microsoft or Google, you likely already have a version that keeps your data out of public training, and telling people that one sentence is most of the policy.

What to actually do first

Pick the single task that eats the most hours and produces something a person checks anyway. In most plants that is document handling in the front office, not anything on the floor.

Then run it for a month with the people who do that job, and measure the only thing that matters, which is whether the work takes less time and comes out at least as good. If it does, that task is worth building properly, so it happens inside your systems instead of in someone's browser tab. If it does not, you spent a month and learned something true about your business.

Do not buy a platform first. The vendors selling AI to manufacturers right now are selling the same three demos, and none of them know whether your bottleneck is the quoting process or the fact that production is written on paper and keyed in at the end of the shift. Find the bottleneck, then decide.

How to tell if someone is selling you nonsense

If you want a hand with this

How much help a company needs here varies more than anything else we do. Sometimes it is a conversation. You tell us what the week looks like, we tell you which one or two tasks are worth the effort and which ones to leave alone, and you go do it with your own people. That costs nothing and we are happy to do it.

Sometimes there is enough going on that it deserves a real look: sitting with the people doing the work, going through the systems, and coming back with a written recommendation about what to automate, what to leave alone, and what it would cost. That is a scoped, paid piece of work, and we would quote it before starting.

And sometimes what a company actually needs is not a project at all, it is for the team to know how to use the tools they already have, which is what our AI training sessions are for. If you would rather see what building it into your systems looks like, that is the AI and automation side.

Tell us what your office spends the most time on and we will tell you honestly whether AI belongs anywhere near it. Reach out here or call (507) 388-4748.