Predictive maintenance in practice: what the Bosch Rexroth case shows

Predictive maintenance in practice: what the Bosch Rexroth case shows

Predictive maintenance in practice means that a model detects wear and failure patterns in sensor or control data, and maintenance moves repairs into planned windows. Documented cases show large effects, such as 70 to 80 % less downtime in pilots at Bosch Rexroth (company statement). Prerequisites: failure history, a narrow selection of equipment and a baseline before the start.

For decision makers in brief

• Effort: The typical entry point covers ten to twenty critical assets, existing control data and a pilot over three to six months.
• Cost: According to the industry report, there are no reliable public benchmarks for project costs per use case in the German Mittelstand, mid-sized companies with roughly 50 to 1,000 employees. Calculate on the basis of avoided downtime minutes, not vendor promises.
• Value: Narrowly scoped individual cases such as predictive maintenance pay off within three to twelve months (Deloitte 2025, consulting study). The average for all AI initiatives is two to four years.
• Risk: The most common mistake is a model that generates warnings but no maintenance orders. Without regular retraining, the hit rate also declines.
• Regulation: In our reading, predictive maintenance that does not intervene in the safety chain generally does not fall under the high-risk classification of the AI Act. This is not legal advice.

The Bosch Rexroth case: what is documented

First, for context: Bosch Rexroth is a documented case from industry and not a NexPatch AI project. The information comes from our analysis of 52 published practitioner conversations (September 2025 to August 2026) and is based on the company's own account.

Bosch Rexroth describes pilot projects on what it calls asset orchestration. Equipment is described via the Asset Administration Shell, a standardised digital representation of each machine, and the workflows around maintenance and intervention are modelled in BPMN 2.0, a widely used notation for business processes. The result according to the company: 70 to 80 % less downtime and 40 to 65 % time savings.

What is remarkable is less the size of the figure than how it was achieved. The case shows that predictive maintenance in practice is more than a good model. It connects equipment data with the process that follows a prediction. This is exactly where many projects fail: the model detects a pattern, yet nobody receives an order. The case thus shows a pattern that runs through the entire analysis. The lever often lies before the AI, namely in a clearly built process.

Two caveats apply. First, these are pilot projects, not a company-wide rollout. Second, our analysis reflects how the speakers present themselves. It is based on the published summaries, not full recordings. Only 14 of the 52 conversations mention a result with a robust figure at all. The case therefore serves as orientation, not as a planning value.

Documented cases compared

Bosch Rexroth does not stand alone. For predictive maintenance, the industry report collects several cases with different types of source. The type of source matters, because an independently verified figure carries more weight than a vendor reference.

The range extends from around a quarter of downtime avoided to peak values of 90 %. For your own planning, the lower end of the range is the more honest basis.

What the studies say on average

Individual cases impress but are not suitable for planning. Aggregated impact studies are the better basis. For predictive maintenance, McKinsey cites a range of minus 30 to 50 % downtime across many projects. Deloitte 2026 puts the figures for the process industry at plus 27 % equipment availability and minus 22 % maintenance costs per unit. Across key operational metrics, the average improvement potential according to Deloitte 2026 is around 20 %.

Both are consulting studies, and consulting studies are systematically more optimistic than representative surveys. That does not speak against the figures, though it does argue for a cautious approach in the business case.

The payback period is the real reason why this field is so often chosen as an entry point. Only 6 % of all AI initiatives achieve their return in under twelve months, and on average it takes two to four years (Deloitte 2025). Narrowly scoped predictive maintenance cases, by contrast, pay off within three to twelve months, in documented vendor references sometimes in under three months. The reason is simple: the value can be measured directly in machine hours. With our tool to calculate the ROI of an AI project, you can work through this logic with your own downtime costs.

Predictive maintenance in practice: prerequisites and pitfalls

Whether a project reaches the values in the table is rarely decided by the model. It is decided by data, selection and process.

What needs to be in place

• Time series data from sensors or controls, ideally connected via OPC UA.
• Historical failure and maintenance data covering at least 12 to 24 months.
• A prioritisation of assets by criticality, not by data availability.
• A maintenance process that can actually take up predictions.

Where it goes wrong

• Models without a link to the maintenance order system, which generate warnings but no orders.
• Too broad a starting scope: all assets at once instead of the ten most critical.
• No time for fine-tuning. Otherwise hit rate and false alarm rate deteriorate over time.
• Purely in-house development without existing building blocks. This lengthens implementation and increases costs significantly.

The first pitfall is the most expensive, and this is exactly where the Bosch Rexroth case comes in. Condition monitoring that only fills dashboards does not reduce downtime. The effect the studies measure only arises when a prediction triggers a planned intervention.

What a robust entry point looks like

The industry report describes how predictive maintenance in practice can sensibly begin, with a typical entry point that matches the documented cases. You select ten to twenty critical assets, use the existing control data and define a reference metric before the start, such as unplanned downtime minutes in the same period of the previous year. The pilot runs for three to six months. During this time the AI proposes and a person decides. Only when quality has been stable for weeks does the system take on more responsibility.

If you proceed this way, you can reduce downtime and also prove it. Without a baseline before the start, every success remains disputed, and without proof there is no second budget.

After that begins the part the industry talks about least: operations. Models deteriorate over time because equipment, tools and product mix change. If you do not define operational ownership, you lose the value again after about twelve months. This is why every pilot has to answer the question of who monitors model quality, who retrains and who decides on a shutdown. We describe what regulated AI operations as a managed service look like on a separate page.

We run predictive maintenance on our Orpheon forecasting platform, connected to ERP, MES and machines, including via OPC UA. The platform is a tool, not a substitute for the sequence described here. Our overview of AI use cases in manufacturing shows which further fields of application make sense after maintenance.