Sharing Data Without Losing Control: Federated Learning and Data Spaces

Why the biggest AI lever in manufacturing sits outside your own plant — and how to use it without giving your data away. A status assessment drawn from 500+ projects across the DACH region.

AI × ERP series · Part 3
Part 1 showed why your own data are too scarce; Part 2, why they must be clean. Part 3 takes up the third way — cooperation in ecosystems — using manufacturing as the example: how federated learning and data spaces enable shared learning without giving up control.

The key answer in 60 seconds

You can learn from shared data without handing over your own.

The data of a single plant rarely suffice for robust industrial AI. Two mechanisms solve this without exposing trade secrets: federated learning trains the model where the data live — the algorithm travels, not the data set. Data spaces govern sovereign exchange between companies through binding rules.

The framework is in place: the EU Data Act has applied since September 2025, and with Catena-X and Manufacturing-X concrete data spaces for industry already exist. The question is no longer whether, but with whom and under which rules. 33+ years of consulting experience, 500+ projects, 100% vendor-independent.

"Our data belong to us — why should we share them?"

The question is fair, and the short answer is: you do not have to give your data away to benefit from data. As Part 1 of this series showed, industrial AI rarely fails on the model and almost always on data that are too scarce and too uniform. A single plant sees only its own machines, its own faults, its own season. For a dependable model, that is often too little.

In manufacturing this bottleneck is especially acute: predictive maintenance, quality prediction or energy optimisation need many comparable cases to become reliable. Those only arise when several operations pool their knowledge — without one looking into another's books. The third way out of the data bottleneck is therefore not a surrender of control, but a new form of control.

Two mechanisms: the model travels, not the data

Federated learning inverts the usual logic. Instead of tipping data into a central pool, the model is sent to the individual sites, learns locally and returns only the learned parameters — not the raw data. From many local learning steps a shared model emerges that has seen more than any single participant, without one data record ever leaving the plant.

Data spaces provide the organisational and legal bracket around this: who may see what, for what purpose, under which conditions? In industry they are already reality. The Catena-X data space connects the automotive sector, and the German Federal Ministry for Economic Affairs–backed initiative Manufacturing-X extends the principle to manufacturing as a whole — sovereign data exchange across company and industry boundaries, under common rules.

Three building blocks for shared learning without losing control

Anyone who wants to share data without becoming dependent or exposed needs three building blocks working together.

Only together do they create the effect: the technology (federated learning) keeps raw data in-house, the data space governs the cooperation, and the law makes the rules binding. None of the three building blocks carries the load alone.

Where it tips over: governance, dependency and the legal framework

The most common mistake is to treat data cooperation as a pure technology project. In fact, governance decides it: which data do we never share, which under conditions, which openly? Those who fail to draw these lines in advance give away more than intended in the rush of the first use case — or block every benefit out of caution.

The legal framework also demands attention. The EU Data Act has applied since 12 September 2025; the "data access by design" obligation takes effect for products placed on the market after 12 September 2026. In parallel, the EU AI Act is phasing in its obligations; the deadlines for high-risk systems were extended in 2026. Anyone setting up a data space should factor in both frameworks early — not as a brake, but as a guardrail that spares later rework.

Recommendation: three steps into the data space

Data cooperation does not begin with technology, but with a clear stance on what is shared and what is not. Three steps have proven themselves:

• Step 1 — set your data boundaries. Before any tool, determine which data remain trade secrets, which are shareable under conditions, and which can be cooperated on openly. This line is a management decision, not an IT one.
• Step 2 — choose the right data space. Assess which initiative fits your sector — Catena-X in the automotive environment, Manufacturing-X for broader manufacturing. Joining an established data space is faster and safer than going it alone.
• Step 3 — start with one use case. Begin with a clearly bounded case — such as predictive maintenance across several sites via federated learning — rather than a platform strategy. A proven benefit convinces more than any roadmap.

Our assessment: the third way is the most demanding — and the one with the greatest leverage. It calls less for a technical than a strategic decision: with whom do we share, under which rules, for what purpose? Part 5 of this series brings that decision together into a roadmap.

Next steps

When your own data are not enough for robust AI, it pays to look beyond the plant fence — technically secured and legally governed. You can find more on our approach to digitalisation, ERP and AI integration in our overview of consulting services.

So far in the series: Part 1 — why ERP data alone can't carry AI and Part 2 — data quality as the foundation.

Part 4 of the series to come soon: From the linear process to the learning system — AI agents in the enterprise (using project-based business as the example). 

Dr. Harald Dreher Managing Director, Dreher Consulting · 33+ years of consulting experience in the DACH Mittelstand · 500+ ERP, digitalisation and AI projects guided · 100% vendor-independent · Personally available to leadership for an initial conversation. Book an appointment directly with Dr. Dreher →