AI in manufacturing: where it delivers results and how a mid-sized company should start

AI in manufacturing covers two very different kinds of project. The first concerns the office around production: orders, documents, planning and reports. The second concerns the line itself: quality control, machines and their parameters.

Office projects do not interfere with how machines run, use data the company already has and deliver results quickly. Line projects need data from sensors and cameras, a connection to PLCs (the computers that control the machines) and a longer pilot. The table below shows what each area needs and what is worth measuring.

Customer orders arrive by email, as a PDF, an Excel spreadsheet or a printout from the customer’s purchasing portal. Someone in the office retypes them into the ERP, looks up products in the master data, checks prices and confirms the delivery date. Customers use their own names and numbers, so retyping takes knowledge and care.

An AI assistant reads the order, recognises products even by the customer’s own names and matches them to item codes in the ERP. It prepares a ready draft order and, when data is missing, such as a quantity or a date, flags it for clarification. An employee checks the draft and approves it with one click.

Do the calculation on your own figures. Suppose the office handles 60 orders a day, and retyping and checking each one takes 6 minutes. That is 6 hours a day, or three-quarters of a full-time job, plus the cost of mistakes in item codes and quantities.

This kind of solution is often built in Microsoft 365 using Power Automate, a tool for moving data automatically between programs, and Copilot Studio, where AI assistants are created. You can also build a separate application connected to the ERP.

In many mid-sized companies, the production plan is drawn up in Excel or an ERP module, but the key decisions are made in the head of an experienced planner. When that person is ill or leaves, the company loses knowledge that nobody has written down.

AI does not replace the planner; it shows them the consequences of their decisions. It brings orders, material availability, machine load and customer deadlines together in one view. The planner can check what will happen if an order is moved by two days and gets an early warning when a deadline is at risk.

A mid-sized company rarely needs an expensive specialist production scheduling system for this. Often it is enough to tidy up the data in the ERP and build a simple view with suggestions on top of it.

In many plants, the production manager puts together a weekly report by hand from several spreadsheets. The management board gets the data late and cannot react during a shift. Power BI, Microsoft’s tool for reports and charts, pulls data automatically and refreshes it on a schedule, even several times a day.

The most commonly measured metric is OEE, or overall equipment effectiveness. It is the product of three figures: machine availability, performance and product quality. For example, 90% × 85% × 98% gives an OEE of about 75%, and the report shows which of the three components is pulling the result down.

The data can come from the ERP, from a manufacturing execution system (a program that records how work progresses on the shop floor) or from simple data loggers on the machines. An AI assistant can also point out unusual changes and answer the management board’s questions, but it is always worth checking the figures in the source report.

A camera with an AI model can spot scratches, missing parts, discolouration or wrong labels faster than a person at the conveyor, and without the drop in attention at the end of a shift. This is one of the areas where AI has a clear technical advantage. However, it only works well in repeatable conditions.

The most common causes of failure are variable lighting and too few examples of defects. The model has to see many images of each type of defect, and rare defects are hard to collect. The system also has to be connected to the line PLC so that a defective product is rejected automatically.

Start with a pilot at one workstation, ideally where defects most often reach the customer. Throughout the pilot, run manual inspection in parallel and compare the results. Only that comparison will tell you whether the system is better than your current inspection.

Failure prediction means that an AI model analyses sensor data, such as vibration, temperature or power consumption, and warns of an approaching fault. In practice it is often confused with simple monitoring, which only raises an alarm when a threshold is exceeded.

For a model to predict anything, it needs history: sensor data from many months and a record of breakdowns with their causes. If the department responsible for keeping machines running does not record the causes of breakdowns, the first step is to change that habit, not to buy a model.

Start with a few machines whose failure stops the whole line and costs the most. At first, the goal is early warnings for the maintenance technicians. A full replacement schedule based on forecasts is only the next stage.

Not every manufacturing company is subject to cybersecurity regulations. The amended Polish National Cybersecurity System Act, which implements the NIS2 Directive, covers only selected manufacturing sectors. A size threshold must also be met: as a rule, at least 50 employees, or turnover and balance sheet total above EUR 10 million, taking linked companies into account.

Medium-sized and large companies in these sectors are important entities. If your company meets both conditions and has not applied to be entered in the list, the matter needs urgent attention, because the deadline passed on 3 October 2026. AI projects that connect the office network with the machine network must then fit within the security rules required by 3 April 2027.

The AI Act rarely applies to office projects in manufacturing, apart from the Article 4 obligation to support employees’ knowledge of AI. It matters more to manufacturers of regulated products, such as medical devices, in which AI performs a safety function. The list below sets out the manufacturing sectors covered by NIS2, and the table summarises the deadlines.

The safest approach is to start with one process and make one decision after the pilot. The plan in the list below is an example of how the first three months might be laid out. It always depends on the number of systems that need connecting and on the availability of people on the company’s side.

The most important thing is a person on the plant’s side who is responsible for the process, such as the customer service manager or the production manager. Without them, the project will stall on questions about exceptions that nobody has time to answer.

The worst start is buying a ready-made “Industry 4.0” platform, meaning an extensive system for digitising the whole factory, with no chosen process and no data. The company pays for licences, and the rollout ends with a demonstration that never makes it into everyday work.

It is also too early for AI when every department keeps the same data in its own spreadsheet and nobody knows which one is current. The same applies when there are only a few orders a day and nobody records the causes of breakdowns. In that case, you first put data and processes in order, and only then choose the tools.