New public procurement rules ask companies to do more environmental assessments: Here’s how AI can help

Starting in 2025, when companies compete for Norwegian government projects, their climate and environmental impact will count for at least a third of the decision.

This also applies in other contracts. Thus, companies must calculate the real carbon footprint of all their products, for example. International standards say how to do that, so products can be compared easily without cheating (greenwashing). Companies must also collect a lot of data and do fast environmental assessments to be ahead of their competitors. Artificial intelligence (AI) can help to get the job done, but companies do not know how. This article guides you on how to select the right AI tool for making high-quality environmental assessments.

For this, we need a lot of data. Unfortunately, most of the data available from Norwegian companies is not good enough. This means that customers cannot fully trust the carbon footprint of the products they buy. Hence, companies must spend more to improve the quality of their data. Companies do not realise that the data they use to calculate the environmental impacts can also be used for improving their operations (e.g. sales, purchases, and maintenance). In fact, improving their data quality can lead to huge financial benefits for these companies in addition to an increase in trust by their customers.

However, data with bad quality is just one aspect. Choosing the right AI tool is another. AI is a box with many tools. Some are good for collecting or processing data, and others are good for environmental impact calculations. Without being an expert in AI, it is very difficult to know which AI tool to use for what. We looked at eleven AI tools and listed which tool is used for what step in environmental assessment. We also checked them against the five AI questions. None of the AI tools perfectly answered the given questions. Some phases have multiple potential AI tools, but you must look at how they score against the five AI questions and then select the best. You don’t take that decision alone. It is best to involve people who know about: AI and data management, environmental assessment, and the manufacturing process for the product. This means people select the AI tool.

The practitioners must quality check the AI tools at different steps of the environmental assessment.  This quality check is crucial. Without humans in the loop, the AI models will give misleading results that can have big consequences in decision-making. Many people have reported mistakes when using AI tools like ChatGPT. These mistakes may be a result of flaws in the AI model. They may also be intentional by the model creators to push their narrative. AI tools that do calculations or comparisons using data also need quality checks. This is because the data can become outdated. For example, due to changes in the manufacturing process of the product.

The human and the AI tools can work together to produce faster, accurate, and more transparent environmental assessments. Our framework provides hands-on guidance on how to do this. It is a step towards hybrid intelligence and better decision-making for a sustainable future.

What does this mean for an environmental assessment practitioner? The first step (goal and scope) of environmental assessments involves deciding what parts of the value chain to include in the assessment. The last step (interpretation and reporting) is where you interpret the results. Because these steps are mostly text-based, you can use an AI tool like ChatGPT.  The second step (life cycle inventory) involves handling missing or poor data and arranging them into a usable format. Therefore, we recommend AI tools like artificial neural networks and Bayesian networks. The third step (impact assessment) mainly involves mapping the data to any impacts they are likely to have on the environment. So, we propose decision-tree-based models and artificial neural networks.

You can read the full article here: https://doi.org/10.1007/s40831-025-01305-x. The referenced paper was part of the Flex4Fact project and received funding from the European Union’s Horizon Europe research and innovation programme under grant agreement No 101058657.