Predictive maintenance in manufacturing: fix it before it breaks

Machines often warn before they fail. A bearing starts to vibrate differently, a motor draws a little more current, a gearbox runs a few degrees warmer. Many factories only notice when the line stops.

In a large plant, that adds up to around 27 hours a month (Siemens, 2024). Predictive maintenance is the discipline of listening earlier: reading the drift in machine data and fixing the problem in a planned stop, before it becomes an unplanned one. This article goes deeper into what we showed in AI Xplained EP.01: what predictive maintenance is, what the evidence says, where European manufacturers stand, and how to start with one machine and 90 days.

What an unplanned stop really costs

Siemens’ 2024 study estimates that unplanned downtime costs the world’s 500 largest companies 11% of their revenue, almost $1.4 trillion a year. An average large plant loses 27 hours a month, down from 39 in 2019, and an hour of a stopped line in a large automotive plant now costs $2.3 million (Siemens, 2024).

Most European manufacturers are not automotive giants. The average manufacturing company in the EU has around 14 people (Eurostat, 2023). The logic holds at any size, because the cost of a stopped hour is never just the parts that were not made. It includes operators waiting, overtime to catch up, rush orders for spare parts, scrap from restarting the process and, often, a late delivery.

A simple formula puts a number on it:

Hours of unplanned stops per month × cost of one stopped hour × 12 = annual cost

As an illustration, a plant with 12 hours of unplanned stops a month and a stopped hour that costs €800 loses €115,200 a year. The figures are an example, not a benchmark. The formula is the point: most plants have never multiplied it out.

Three ways to run maintenance

Reactive. Fix it after it breaks. It needs no planning, which is why it persists, but every failure arrives at the worst moment, with rushed parts and overtime.

Preventive. Fix it on a calendar or a counter: every 500 hours, every quarter, every 10,000 cycles. It turns many surprises into planned stops, but it assumes parts wear out with age. Many failures do not follow age, so a calendar both replaces healthy parts and misses faults that develop between services.

Predictive. Fix it when the data says so. The machine’s own signals decide when an intervention is needed. According to the US Department of Energy, a properly functioning predictive programme saves 8% to 12% over a programme that relies on preventive maintenance alone (US DOE, 2010).

The three are not mutually exclusive. Most plants will keep reactive maintenance for cheap, non-critical equipment and preventive maintenance where wear really is age-related. Predictive maintenance earns its place on the machines whose stops cost the most.

How it works

Most production lines already record more than people think: temperature, vibration, current and power draw, pressure, cycle counts and alarms, stored in PLCs, SCADA systems or machine controllers. Nine in ten large manufacturers (87%) already collect data that makes predictive maintenance possible (Siemens, 2024).

A predictive model uses that history to learn what normal looks like for each machine, under each operating condition. The same motor behaves differently at full load and at idle, with one product and with another, so the baseline has to be specific. Once the baseline exists, the model watches for drift: a slow change away from normal that does not yet trip any alarm.

Reliability engineers call this window the P-F interval: the time between the moment a developing fault becomes detectable (P) and the moment the machine fails (F). Threshold alarms usually fire close to F. Predictive maintenance aims to detect close to P, when there is still time to order the part, schedule the stop and do the repair properly.

The last step is human. The model flags the drift, a technician checks it, and a person approves the work order. Predictive maintenance does not replace the maintenance team; it gives it time. We wrote about designing those approvals without slowing the operation in Human in the loop.

What the evidence says

The headline numbers are consistent across studies, with one caveat: they are averages from companies already running predictive maintenance, not what a first pilot delivers.

• Downtime and machine life. Predictive maintenance typically reduces machine downtime by 30% to 50% and increases machine life by 20% to 40% (McKinsey, 2017).
• European users. In a survey of 268 companies in Belgium, Germany and the Netherlands, companies implementing or piloting advanced predictive maintenance reported, on average, 9% more uptime, 12% lower costs, 14% lower safety, health, environment and quality risks, and 20% longer life for ageing assets (PwC and Mainnovation, 2018).
• Adoption. Almost half of the firms surveyed by Siemens now have predictive maintenance teams, twice the share in 2019, and the hours lost per plant have fallen over the same period (Siemens, 2024).

The same European survey shows how far most companies still are from that level. Only 11% had reached advanced predictive maintenance. Most relied on visual inspections (25%) or instrument inspections (42%), and 20% had real-time condition monitoring without the predictive layer (PwC and Mainnovation, 2018).

Where European manufacturing stands

AI adoption in Europe is still low. In 2025, 20% of EU companies with 10 or more persons employed used AI, with wide gaps between countries: 42% in Denmark against 5% in Romania (Eurostat, 2025). Size matters too: 17% of small companies used AI, 30% of medium-sized ones and 55% of large ones.

How AI is used also depends on size. Among EU companies that already use AI, a third of large firms (33.5%) apply it to production processes, against 19% of small ones (Eurostat, 2025). Small firms are the furthest behind on the factory floor, and most European manufacturers are smaller still: almost 85% of EU manufacturing companies have fewer than 10 people (Eurostat, 2023).

The reasons are revealing. Among EU companies that considered AI but did not adopt it, 71% point to a lack of relevant expertise, 53% to unclear legal consequences and 49% to data protection and privacy concerns (Eurostat, 2025). In the European maintenance survey, 63% of companies with no plans to adopt said a good business case could not be made or that it was not relevant to their business, and 23% said they did not have enough data (PwC and Mainnovation, 2018).

Know-how, a business case and data. All three can be tested on a single machine before anyone commits to a plant-wide programme.

What it needs: three things

Machine data. Temperature, vibration and power are the usual starting point, and most lines already record at least one of them. Where a critical machine has no sensors, vibration and temperature sensors can usually be retrofitted without redesigning the line. The data does not need to be perfect to start, as we explained in Your data doesn’t need to be perfect.

Maintenance history. Past failures and repairs, even the ones logged on paper or in a spreadsheet. History is what lets the model link a pattern in the signal to a real failure, and what lets the team judge whether an alert makes sense.

Someone who acts. An alert nobody owns is noise. Every alert needs a person who checks it and decides, and a route into the work order system the team already uses. A model that flags too often gets ignored, so the thresholds are tuned with the maintenance team, not for it.

The legal side is simpler than it looks

More than half of EU companies that considered AI and did not adopt it point to unclear legal consequences (Eurostat, 2025). For predictive maintenance, two European rules matter, and both are more helpful than they sound.

The Data Act. Since 12 September 2025, users of connected products, including industrial machines, have the right to access the data their equipment generates, free of charge and, where technically feasible, continuously and in real time. They can also ask the data holder to share it with a third party of their choice. Connected products placed on the EU market after 12 September 2026 must be designed to make that access possible (Regulation (EU) 2023/2854). In practice, it becomes much harder to keep the data a machine produces about itself locked inside a supplier’s service contract.

The AI Act. Predictive maintenance is not on the AI Act’s list of high-risk uses. It would become high-risk if the AI system were a safety component of a product that needs third-party assessment under EU product legislation, such as some machinery, or a safety component in running critical infrastructure such as electricity, gas or water supply (Regulation (EU) 2024/1689). A model that recommends when to service a motor, with a person approving the work order, sits well outside that category. Specific cases should still be checked with legal advisers. Our article on the AI Act for operations leaders covers the rest.

How to start: one machine, 90 days

Day 0. Pick the machine. Choose the one whose stop costs the most per hour, usually the bottleneck of the line. One machine is enough to prove the case. We described how to choose in How to choose the first process to put AI to work.

Days 1 to 30. Connect what exists. Sensors, PLC logs and the maintenance records the team already keeps. Fill the gaps that matter, nothing more.

Days 31 to 60. Learn its normal. The model builds the baseline for each operating condition. Alerts go to the maintenance team for review, and the team helps decide what counts as real drift.

Days 61 to 90. Act and measure. The team acts on the alerts that matter. Compare unplanned stops, hours lost and their cost with the same period before.

Back to the example: a plant losing €115,200 a year to unplanned stops that cuts them by 30%, the lower end of McKinsey’s range, saves €34,560 a year. A 90-day pilot on one machine will not reach that by day 90, and it does not need to. Its job is to prove, with the plant’s own numbers, that the drift was visible before the failure. Setting that measure before the project starts is the subject of How to measure the return on an AI project before you start.

Fix it before it breaks, not after

Predictive maintenance is not a new idea. What has changed is the cost of doing it: the data is already on most shop floors, the models are mature, and European law now gives manufacturers the right to the data their machines produce. The biggest gap that remains is know-how, and that is the easiest one to close, one machine at a time.