Search how to use AI in manufacturing and you will get the same article about fifteen times. Assemble a cross-functional team. Hire data scientists, or a consultant who has them. Audit your data maturity. Pick a pilot project. Start with predictive maintenance, or computer-vision quality inspection. Scale from there.
None of it is wrong. It is just written for somebody else — a plant with a thousand machines, an IT department and a capital budget with a line item for this. Read it as an owner with thirty machines and two people in the office, and the reasonable conclusion is that artificial intelligence is a problem for a few years' time.
That conclusion is wrong, and it is costing small manufacturers real money right now. The list is what's wrong, not the timing.
This guide is the version for a factory of twenty to fifty machines. What actually works, what to do first, and what you can safely ignore for another five years.
Why the standard advice doesn't fit you
Almost every "AI in manufacturing" article leads with the same two use cases: predictive maintenance and computer-vision quality inspection. They lead with them because they are genuinely impressive, and because the companies writing the articles sell them.
Both need something you do not have.
Predicting a machine failure means learning from failures you have already recorded, in detail, many times over. A plant running a thousand identical machines accumulates that history in a few months. You have thirty machines, several of them different makes, some of them older than the people operating them. The pattern learned from one does not transfer to another. You would need years of carefully logged breakdowns before there was anything to learn from — and nobody has been logging them that way, because why would you.
Camera-based inspection has the same shape of problem. It works when you can show it thousands of examples of good and bad. If your defects are rare, varied, and currently caught by an operator who has been doing this for twenty years, you are being asked to build a training set from scratch for a job that is already being done.
So the two headline use cases are real, and they are not for you yet. That is not a limitation of the technology. It is that your factory hasn't generated the evidence they run on.
Here is the part the articles skip: the AI that pays back at your size doesn't need a training set at all. It needs your data. And you already have it — it is just currently trapped on paper and in people's heads.
The one thing that has to be true first
Before any of this, there is a precondition that no amount of software will fix for you.
In most small factories, production is recorded by hand. An operator writes the day's output on a sheet. The sheet reaches the office. Someone keys it in on Saturday, or on the last day of the month, working partly from the sheet and partly from memory. Everyone knows the figure is roughly right and nobody would swear to it.
Now put an AI on top of that. Ask it what your cost per unit is, which product made money last month, whether machine four is underperforming. It will answer confidently. It will be answering from the sheet.
An AI cannot tell you something your data does not know. It can only be wrong faster, and with more authority.
This is the single most common reason AI disappoints in a small factory, and it has nothing to do with the AI. Fix the input first.
Fixing it is less work than it sounds. A machine that runs is doing something physically observable, and a small sensor can watch it without touching the control system, the wiring or the PLC. From "this machine ran for 6 hours 40 minutes" plus what it was set up to make, you get output — measured, not remembered. No operator has to type anything, which matters, because any system that depends on the floor entering data will quietly stop being accurate within a month.
That is the foundation. Everything below assumes it.
Where AI actually earns its keep at this size
1. Reading the paperwork that arrives
Every factory has a pile of purchase orders and supplier bills moving slowly toward a keyboard. Somebody retypes them. It takes hours a week, it is nobody's favourite job, and it is the reason your costing runs on last month's rates.
That last consequence is the expensive one. If your yarn rate went up 8% three weeks ago but the bill has not been entered, every quote you have given since is built on the old number. You are pricing from a past that no longer exists.
Reading a document is what modern AI is genuinely, boringly good at. Photograph the bill, and the fields come back filled in for someone to check and save. It is a good first project for reasons that have nothing to do with the technology:
- It repeats daily, so you find out quickly whether it works.
- You can check the answer. The paper is in your hand. Compare and you know.
- Failure is cheap. If an extraction is wrong, the person checking it corrects it, exactly as they would have typed it anyway.
- It compounds. Current rates make every downstream number — cost, margin, quotes — better without anyone doing anything else.
On our own floor, 354 GST invoices were read between May and July 2026; the invoice number and date were accepted without correction on all 354, and the total matched on all but four. The full numbers are here.
2. Reading your own record back to you
You know your top three customers. Do you know which customer bought 20% less this quarter than last, without going and looking? Most owners find out when the drop is big enough to notice in the bank balance, which is months late.
This is a job AI does well because it is reading rather than predicting. Given a real production and sales record it can describe what has been happening — order cadence, seasonality, which products are growing, which relationship has gone quiet — in plain language, without you building a report.
The value depends entirely on the record being real. Which is the point of the precondition above.
3. Taking that record outward
The one almost nobody expects from factory software, and often the one with the fastest payback.
Ask an owner how they find new customers and the answer is referrals, exhibitions, and agents. All three work; none of them is a lever you can pull in a slow month.
There is a fourth approach that is obvious and almost never done, because by hand it is miserable: look at who already buys from you, work out what those companies have in common, and go find more like them. AI is well suited to exactly that, and the output is checkable in the way that matters — each suggestion can name which of your existing customers it resembles, so you can judge it in seconds. More on that here.
4. Answering you in plain language
The least glamorous and most used. Instead of opening a report, you ask: what did machine 7 run yesterday? — which product lost money last month? — who owes me more than sixty days?
This sounds like a convenience. In practice it changes who can get an answer. When getting a number requires knowing which report to open and how to filter it, exactly one person in the building can do it, and everybody waits for them. When it requires asking a question, the constraint disappears.
The order matters more than the choice
If you take one thing from this: do it in this order.
- Measure output at the machine. Until this is true, everything downstream is a guess with better formatting.
- Point AI at your incoming paperwork. Narrow, daily, checkable. Your rates become current.
- Now ask it things. Cost per unit, per product, per customer — questions worth trusting because what's underneath them is real.
- Then look outward. New customers, better suppliers, using a record that now describes your factory accurately.
Done in that order, each step makes the next one work better. Done in reverse — the usual way, starting with an impressive demo — you get a confident system built on a shaky foundation, and you will stop trusting it within a quarter.
What you do not need
Worth saying plainly, because these beliefs are what stop most owners before they start.
- A data scientist. You are not training a model. You are using tools that are already trained.
- A machine learning project. Same reason.
- A server, or an IT department. This is bought as a subscription and set up for you.
- To replace your accounting software. Your books stay where they are and your CA keeps working the way they work.
- To change how you make anything. No new routine on the floor, no operator entering figures into a tablet.
- Eighteen months. A pilot on a few machines is a day. If it is going to take a year to see anything, that is the wrong project.
How to tell whether it is working
Set the test before you start, because a system that produces impressive screens and changes no decisions is the most common failure and the hardest to notice.
Reasonable tests, roughly in order of how quickly you should see them:
- Within weeks: is anybody still retyping supplier bills? Are the rates behind your costing from this month?
- Within a quarter: can you name your least profitable product without asking anyone? Did that answer surprise you? (It usually does — and that surprise is most of the return.)
- Within two quarters: have you re-quoted or dropped something because of a number you did not previously have? Has a customer going quiet been caught early enough to do something about it?
If the honest answer to all of those is no after six months, the system is decoration. Stop paying for it.
The short version
AI in a small factory is not the thing in the articles. It will not predict your breakdowns yet, and it will not inspect your product with a camera yet. Both of those are coming, and both need years of your own history first.
What it will do today is read your paperwork so your costs are current, read your own record back so you can see what is happening in your own business, look outward for customers who resemble the ones you have, and answer you without you opening a report.
That is not the exciting version. It is the one that pays for itself in a quarter — and the only precondition is that the numbers underneath it are measured rather than remembered.