Can AI be green?

At one of Helbling Technik’s customer's production facility, a part can take up to 30 minutes to pass through the production line. Only then does it become clear whether the selected settings deliver the required quality. Hundreds of parameters influence the process. Finding the right parameter configuration for a new product can therefore take several hours and requires an experienced domain expert.

To address this, Helbling developed a machine-learning system that evaluates production parameters and quality data from the first manufactured part and proposes adjustments to the settings. In the implemented solution, the second part already meets quality requirements. This reduces the number of setup attempts from the low double digits to two. At the same time, the setup process becomes less dependent on the experience of individual experts.

Fewer setup iterations offer the potential to reduce start-up scrap and unnecessary machine runtime. But how significant is the environmental benefit? The process improvement alone does not answer that question; the additional resource use caused by the AI must also be taken into account.

The balance extends beyond data centres

Figures from the International Energy Agency (IEA) show why this additional demand matters. In 2025, data centres worldwide consumed around 485 TWh of electricity. That is more than eight times Switzerland’s final electricity consumption of approximately 58 TWh in the same year. In its central 2026 projection, the IEA expects global data-centre electricity consumption to rise to around 950 TWh by 2030. AI is an important driver of this growth, although these figures cover all data centres, not AI applications alone.

For an individual application, two perspectives are useful. “Green in AI” focuses on the AI’s own resource use and on ways to reduce it—for example, through appropriately sized models and hardware, efficient architecture, or avoiding unnecessary processing. “Green by AI” looks at the savings that AI can enable in the wider process.

Taken together, these perspectives change the basis of comparison. What matters is the resources required to perform the same task, not just the resources consumed by a single model computation. In manufacturing, a part that meets defined quality requirements can serve as the common reference unit. The AI-supported process can then be compared with the previous process or with a suitable non-AI alternative over a defined period and within consistent system boundaries.

Quality is part of the environmental assessment

This shows why a smaller model is not automatically the better solution. If, for example, a more resource-efficient model used for quality inspection rejects good parts more often, the energy saved in computation may be offset by additional material losses. Conversely, greater model accuracy does not justify any level of additional resource use. What matters is the improvement the model actually delivers in the process.

The environmental question therefore comes back to requirements. What level of quality must the solution achieve? Which errors are acceptable, and which lead to scrap or rework? During a feasibility phase, Helbling clarifies these requirements and assesses whether the available data contains the needed information. More data is of little value if it does not capture the relationships that actually matter.

On this basis, technical approaches are compared using automated tests and experiments. This includes non-AI solutions as well. For clearly defined tasks, fixed rules or conventional algorithms may be sufficient. Machine learning is an option when relevant relationships can be learned from data—for example, for quality prediction. Language models can be useful for processing text.

The tests show which variants can reliably achieve the required quality. To make an environmentally sound selection, their resource use and the consequences of their decisions within the process must also be assessed.

Measuring what is added and what is avoided

For the production line, this means asking: How much material loss and machine runtime are actually avoided? How often does the process need to be set up? And what additional resource use results from data processing, model training, operation and the hardware used? Only this information enables a robust assessment.

Depending on the infrastructure, the software’s energy demand can be measured or modelled. In addition to computing components, other consumers such as storage, networks and cooling must also be taken into account. Where detailed data is unavailable—as is often the case with cloud services—estimation models based on transparent assumptions and stated uncertainties can be used.

Energy consumption, material use and greenhouse-gas emissions must be considered separately. Lower electricity consumption does not automatically translate into a proportionally lower climate impact, because how the electricity is generated also matters. Similarly, material savings cannot simply be offset directly against additional electricity consumption. The respective environmental impacts must be assessed.

A computational result must become a process improvement

The overall balance also depends on how the solution is integrated into operations. A proposed parameter correction does not save material on its own. That effect can occur only if the recommendation is used in the workflow and leads to an actual process improvement.

In the setup process described above, this requires bringing production and quality data together, assessing deviations correctly from a domain perspective, and providing suitable parameter adjustments. Process knowledge determines which measurements are meaningful and which interventions make sense. The software architecture ensures that the required data is available and that recommendations reach the workflow where they can be acted on. More computing power alone does not solve these tasks.

To achieve this, Helbling combines process engineering, data engineering, AI/ML and software engineering. Their interaction connects data and models with the real-world process. This creates the conditions needed for a technically possible improvement to take effect in operation.

Conclusion: Net impact is what matters

AI can be green if the environmental impacts avoided through its use outweigh the additional resources it requires. An efficient model can contribute to this outcome, but it cannot replace a system-level assessment—any more than improved process quality can.

When setting up a production line, the potential lies in the iterations that are no longer needed. The relevant question is therefore not only how much energy a computation consumes, but how total resource use changes on the path to a part that meets the quality requirements. A model can become more efficient and still represent the wrong optimisation if the real-world process does not improve as a result.

Authors: André Wangler, Stephan Schlatter

Main Image: AI-generated