Where does AI go in a sustainability report?

In the European Union, a sustainability report follows the Corporate Sustainability Reporting Directive (CSRD) and its standards, the ESRS. The Omnibus I simplification package, published in the Official Journal on 26 February 2026, narrowed the scope: the obligation now covers only companies and groups with more than 1,000 employees on average and more than €450 million of net turnover, and Member States have until 19 March 2027 to transpose it. The revised ESRS↗, adopted by the Commission on 3 July 2026 on the technical advice of EFRAG, apply to financial years beginning on or after 1 January 2027, and companies already in scope may apply them from financial year 2026. The EFRAG draft, which the Commission took up, reduced the number of mandatory data points by 61%.

We searched for the words "artificial intelligence", "AI", "machine learning" and "algorithm" in the whole text of the revised standards, of the act adopting them and of the Commission staff working document: no occurrence. The standards therefore provide neither an indicator nor a disclosure item for AI. It enters through general items, depending on what the company does with it and on what its double materiality assessment retains.

The table follows the order of a sustainability report. For each item, it says what the use of AI may contribute and in which case. The numbering is that of the revised standards: energy becomes E1-7 and emissions E1-8.

The only provision close to AI is in the climate standard. For emissions from purchased goods and services, the first scope 3 category, it asks, where relevant, to disclose the emissions of purchased cloud and data centre services as a subset of that category, and the optional breakdown table that the standard proposes reserves a line for them. An assistant licence, a subscription, calls to a model API or rented GPU hours fall under that line: the provider runs the servers.

The figure depends on the method. Counted in euros, with a monetary ratio for IT services that covers the provider's whole chain, the item weighs nearly 30 times the electricity of the servers that answer. Counted in usage, the electricity of the servers answering the 200 employees of an office-based company equipped with Microsoft 365 Copilot, served in the European Union, weighs 0.007% of a median carbon footprint; this is a floor, without the manufacture of servers or the training of models. Thehub analysis of AI in the Bilan Carbone® sets out this calculation, the data to collect and the factors. Agents, AI served to the customers of a product or own servers change the scale of the item.

When the company runs its own servers, the electricity enters its energy consumption (E1-7) and its scope 2 emissions, reported location-based and market-based; air conditioning adds refrigerant leaks to scope 1. The emissions boundary follows financial control by default; the revised standards also allow equity share or operational control. For rented servers, an ESRS 1 rule takes precedence: the lessee reports in its own operations the impacts of using an asset it rents, and therefore its electricity. The water of an own data centre enters the water standard, with the share consumed in water-stressed areas.

The standards require the disclosure of material information, on impact or financial grounds, following a double materiality assessment, and the revised standards forbid disclosing non-material information except as additional information flagged as such. Materiality is judged by topic: if climate change is material, the emissions of purchased AI are included in category 1 of scope 3, and the question is how much detail to give. For the environmental footprint of AI, the question is first settled as an order of magnitude. In an office-based company, the electricity of the servers answering employees is very far from the large items of the footprint, such as purchased goods. It may become material when the company runs agents continuously, sells a product that queries a model on each request, or operates its own computing centres.

The social side of AI may count for more than its environmental side. In the summary of its 2025 integrated report, Capgemini mentions AI in connection with training, with more than 194,000 employees trained in agentic AI in 2025, and in a target for AI ethics training; the same page presents its environmental results without mentioning AI. The European standards on the workforce (S1), on customers (S4) and on business conduct (G1) accommodate this information, each according to the materiality of the topic.

For companies of up to 1,000 employees, the voluntary standard of 2026 provides for total energy and scope 1 and 2 emissions on a location-based basis; it has no line for cloud, and purchased AI appears there only if the company reports its scope 3. Operators of data centres with at least 500 kW of installed IT power report their energy and water indicators to a European database; this register is separate from the sustainability report and is published only in aggregate.

In our view at Projet Celsius, AI is reported like everything else: by its uses, in the item where the regulation places what it does. For the footprint, this means a figure in the cloud and data centre subcategory, calculated on usage with the grid electricity of the servers' region, rather than at the monetary ratio, which follows the price paid and not the usage; manufacturing and training, which this count leaves out, are added when a factor exists or are declared as a limit. A dedicated AI item is justified when the double materiality assessment retains it. Otherwise, its emissions remain in purchases; a company that wants to state the scope counted and its weight does so as additional information, flagged as not resulting from the materiality assessment. The gains that AI would bring to the company's climate cannot be deducted from its emissions: the standards require gross emissions, assess a positive impact without offsetting it against negative ones and exclude avoided emissions from reduction targets.