The honest starting point

Manufacturing is the area where AI marketing runs furthest ahead of reality for small and medium businesses.

The impressive demonstrations — predictive maintenance, computer vision quality control, self-optimising production schedules — are real, and they are built on data most SMEs do not have. They need years of sensor readings, thousands of labelled defect images, or machine telemetry that a business without connected equipment simply does not collect.

That does not mean nothing applies. It means the useful things are less exciting than the brochure.

This article separates them.

What works without special equipment

Four uses that need only what Odoo already holds.

1. Better production scheduling

Given your orders, your capacity and your material availability, working out a sensible sequence.

Why it helps: scheduling by hand is genuinely hard once you have several products competing for the same work centres. A person does it by rules of thumb and habit.

What it needs: accurate work centre capacity and realistic operation times. If your routings say twenty minutes and reality is forty, the schedule is fiction.

2. Component demand forecasting

Predicting what raw materials you will need, from your production plan and history.

Why it helps: component shortages stop production, and the cost of that is far higher than the cost of holding a little more.

What it needs: accurate bills of materials and reliable supplier lead times.

3. Reading supplier documents

Bills, delivery notes and certificates read and turned into drafts.

Not manufacturing-specific, and often the fastest payback in a manufacturing business, because material purchasing generates a lot of paper.

4. Spotting unusual consumption

Flagging when actual component usage diverges from the bill of materials.

Why it helps: it catches BoMs that have drifted from reality, and it catches waste nobody had quantified.

What needs data you probably do not have

Three, and being honest about them saves money.

Predictive maintenance. Predicting a machine failure before it happens needs sensors on the equipment, collecting continuously, for long enough to have seen failures. Without connected machines, there is nothing to learn from.

Vision-based quality inspection. Detecting defects from images needs thousands of labelled examples of both good and defective parts. Collecting and labelling that is a project in itself.

Real-time process optimisation. Adjusting parameters as production runs requires machine integration most SME equipment does not support.

These are genuine technologies. They are appropriate for businesses with instrumented equipment and the volume to justify the setup. Presenting them as available to any manufacturer is where the marketing overreaches.

The prerequisite nobody skips

Your bills of materials must be right.

Everything above depends on this. Component forecasting from a wrong BoM produces wrong material requirements. Consumption variance against a wrong BoM flags normal usage as an anomaly. Scheduling from wrong operation times produces a plan nobody can meet.

Three specific things to check:

The BoM header quantity matches the product’s unit of measure. A product measured in grams with a header quantity of 1 will produce component quantities wrong by whatever factor. It throws no error — it just produces wrong numbers quietly.

Every component has a real cost. A component priced at zero makes the finished product look free, and every margin figure downstream is wrong.

Operation times reflect reality. Not what they were when the routing was written three years ago.

Where to start

1. Fix your BoMs. Check header quantities against units of measure. Verify component costs. This is unglamorous and it decides everything.

2. Record actual consumption, not just planned, wherever real usage varies.

3. Add document reading for supplier paperwork. Fastest payback, lowest risk.

4. Try component forecasting on your high-volume, stable materials. Measure shortages and stock value together.

5. Look at scheduling if you have real capacity constraints and multiple products competing.

6. Only then consider anything requiring new instrumentation, and cost it as the project it is.

Measuring it

Four numbers, and they must be read together.

Material shortages stopping production. Should fall.

Raw material stock value. Should not balloon while shortages fall. Anyone can eliminate shortages by holding everything.

Schedule adherence. Are you making what you planned, when you planned?

Consumption variance. Actual against BoM. Improving here means either your recipes got more accurate or your process got tighter. Both are wins.

What to be sceptical about

Any claim not attached to a specific task. “AI-optimised manufacturing” is not a feature.

Predictive maintenance without sensors. Ask what data it uses. If the answer is vague, it is not predicting anything.

Forecasts on short history. A component with six months of usage data will produce a forecast that looks identical in confidence to one with six years.

Anything replacing an operator’s judgement about safety or quality. Advisory is appropriate. Deciding is not.

The short version

For most manufacturers, AI’s value today is planning and paperwork, not the shop floor.

Better scheduling, better material forecasting, less time on supplier documents, and earlier warning when consumption drifts from the recipe.

None of it works if your bills of materials are wrong — and in most systems that have not been reviewed in a while, at least one of them is.

Start there. The shop-floor technologies are real, and they are a separate project with a separate budget.