Siemens’ Erlangen factory: how to attribute results when AI is only one component

On October 7, 2024, the World Economic Forum added Siemens’ Erlangen factory to its Global Lighthouse Network. The site profile credits its Green Lean Digital strategy with a 69 percent improvement in labor productivity, a 40 percent reduction in time to market and a 42 percent reduction in energy use. These are industrial outcomes, not laboratory scores. They are also the combined product of artificial intelligence, digital twins, robotics, analytics, systems integration and process redesign.

That final point is the lesson. In a complex transformation, a factory-wide indicator does not by itself show how much each algorithm contributed. The transferable skill is building an attribution chain: start with a specific decision, define the local metric it can change, verify the mechanism, and only then connect it to plant-wide productivity, quality, energy or lead time. Without that chain, “AI raised productivity by 69 percent” would be a stronger conclusion than the sources support.

What Erlangen is and what was measured

The plant makes SINAMICS drive products in a high-mix, medium-volume setting. According to the World Economic Forum site profile, it deployed more than 100 AI algorithms and extensive digital twins on a flexible, modular IT architecture. The improvement period was four years. WEF presents the three headline figures as outcomes of the site strategy, not the output of a single model.

Siemens issued its announcement on October 8, 2024 and used similarly aggregated language: AI, digital twins and robotics increased productivity and reduced energy use. The company said AI had been deployed across more than 100 use cases. “Algorithms” and “use cases” are not necessarily the same unit. One application can combine several models, while a model can support more than one task. A portfolio audit should not substitute those terms for one another.

Lighthouse status is not a certification that every buyer will obtain the same outcome. It is a selection of sites showing digital transformation at scale. Erlangen is also “customer zero”: the factory’s own page says it tests Siemens solutions in real production before they reach the market. That creates a valuable validation environment, but it also brings vendor expertise, resources and a commercial interest that should stay visible.

Breaking down an aggregate number

The 2025 Global Lighthouse Network report goes one level deeper. Highlighted applications include AI-enabled closed-loop electrical testing, with a 51 percent reduction in false positives; an end-to-end analytics platform for new semiconductor manufacturing operations, associated with a 19 percent improvement in production yield; product digital twins used to train visual-inspection and perception-robotics models, linked to a 50 percent reduction in field failures; automated outbound logistics, with five times the labor productivity in the affected area; and an additive-manufacturing network for spare parts, with an 80 percent reduction in lead time.

Those metrics are more auditable because each has an intervention and scope. Reducing false positives in an electrical test can prevent reinspection and holds on good units. Improving yield means obtaining more conforming output from process inputs. Shortening spare-part lead time can reduce maintenance waits. None is automatically equivalent to the plant-wide 69 percent productivity increase, but each describes a mechanism through which the program may contribute.

The word “productivity” always needs a denominator. It may refer to units per labor hour, value added per employee, cell throughput or direct time saved. WEF identifies labor productivity for the headline figure and confines the fivefold improvement to the affected logistics area. Merging those measures would create a false story: a fivefold local gain did not necessarily occur across the factory.

Energy requires the same care. Siemens reports 42 percent lower consumption over four years and, inside the semiconductor project, a special energy-management system that reduced consumption by more than 50 percent. These are different boundaries. Before comparing them, ask whether the figure is absolute or per unit, which buildings and processes it covers, what the baseline year was, and whether output volume or product mix changed.

How a decision travels from model to process

Consider electrical testing. A system classifies signals and flags a unit as potentially defective. The first metric is an error matrix: real defects caught, defects missed, good units held and good units released. Cutting false positives saves review, but it is safe only if false negatives do not rise. The next level measures cycle time, inventory buildup and diagnostic work. Only then can the contribution to productivity or lead time be examined.

For bin picking, Siemens simulates a robot, bin and parts through digital twins to train perception and motion before acting on real hardware. Simulation lowers the cost of generating scenarios and trying configurations, but it introduces a classic question: the gap between the simulated and physical worlds. Reflections, wear, part positions, lighting and tolerances change. Deployment needs tests with real components, confidence limits and a safe fallback when the system cannot recognize the scene.

Information-technology and operations-technology integration makes that loop possible. Test, order, equipment and quality data need shared identities, timing and meaning. If a signal cannot be connected reliably to the correct unit, product version and operation, the model learns ambiguous correlations. Architecture and data governance are not administrative preliminaries; they are components of the production system.

A digital twin is not a magical copy either. It can represent geometry, physical behavior, process sequence or a combination. Its value depends on which variables it contains, how it is calibrated and when it is updated. For each decision, teams should document what reality is omitted. A model sufficient to prevent a collision may not predict wear; one useful for planning space may not validate a robot path.

An operator is not a generic “control”

Saying that a human is in the loop is insufficient. A design should specify who receives an alert, what information that person sees, how much time is available, what can be changed and who is accountable. In inspection, a specialist may review uncertain cases; in maintenance, confirm a cause before stopping equipment; in planning, try an alternative in the twin. Each function needs different authority and training.

Automation may free people from repetitive work, as Siemens argues, but it can also shift effort toward supervision, labeling, exception handling and model maintenance. A fair evaluation counts those hours. If an indicator records only manual time removed while ignoring the new work, it overstates value and hides an operational dependency.

Safe adoption also requires reversibility. Teams need a way to return to a known rule or process when data are missing, the product changes or the model leaves its validated range. Recording model version, input, recommendation, human decision and outcome makes incidents reconstructable. In manufacturing, traceability is not a later report; it is the condition for improving without losing control.

A method for transferring the case

The first step is to choose a decision, not a technology: release a test, prioritize maintenance or select a part. The second is to establish a local baseline for volume, quality, energy, time and exception cost. The third is to define model and process metrics separately. The fourth is to pilot against a comparable group with stopping criteria. The fifth is to measure the entire operation, including human oversight, and document other changes made at the same time.

Scaling can then proceed in modules. One hundred use cases should not be the opening target; they are the result of data, architecture and discipline that make the cycle repeatable. A portfolio should retire applications that do not create value rather than preserve them to inflate an innovation count. It should also report periods, denominators and uncertainty, not percentages alone.

Erlangen shows that digital transformation can sustain measured results in a real factory. It does not show that AI in isolation produced every point of the 69, 40 or 42 percent. That reservation does not weaken the case; it teaches readers where value is actually created. A factory learns when it can connect a prediction to an action, that action to a process change and that change to a metric whose baseline it retained.