The carbon footprint of AI and information systems: KPIs and good practices

Information systems have an environmental impact that cannot be seen from the desk, but that grows with use: data centres, storage, networks, workstations and, increasingly, artificial intelligence. The good news is that a significant part of that impact depends on decisions the organisation itself makes: which model it uses, how much it uses it and which supplier it contracts.

Where the impact comes from

• Energy. Data centres consumed around 415 TWh in 2024, about 1.5% of the world's electricity, and the International Energy Agency expects their consumption to double by 2030, driven mainly by AI.
• Training versus use. Training a large language model generates a lot of emissions, but everyday use, with millions of queries, ends up exceeding it. The part a user organisation can manage is use: how much, and with which model.
• Water. Data centres use water for cooling. Per-query estimates vary enormously depending on the model and the data centre, which is why you should ask the supplier for the figure rather than rely on averages.
• Minerals and waste. Chips and specialised equipment require lithium, rare earths and other minerals, and end up as electronic waste.
• The rest of the estate. In-house servers, backups, printers and networks also add up. The measures in this article apply to the whole technology estate, not just to AI.

References for measuring

Reference	What it is for
UNE 0086:2025	Spanish specification with energy, carbon, water and performance metrics for AI systems in training and inference. The main reference for requesting data from suppliers
ISO/IEC TR 20226:2025	Conceptual framework on the sustainability of AI across its entire lifecycle
ISO/IEC 21031:2024 (SCI)	Software carbon intensity per functional unit: lets you ask for “g CO₂e per query”
ISO 14064-1 and GHG Protocol	The organisation's emissions inventory
AFNOR Spec 2314	Good practices for “frugal AI”
Directive (EU) 2023/1791	Requires data centres above a certain power rating to report on consumption, efficiency, water and renewables
ITU-T L.1510 (2025)	Environmental KPIs for digital infrastructure in six categories and three maturity levels
ITU-T L.1450 (2025)	Footprint of the ICT sector across the whole lifecycle and an emissions budget compatible with 2 °C
LCSP, art. 145	Legal basis in Spain's Public Sector Contracts Act (LCSP) for including environmental criteria in tender specifications

What an organisation using the cloud can measure

Consumption can almost never be measured directly. The realistic approach is to combine three sources:

1. Supplier data: energy and emissions per query or per user, PUE, WUE, percentage of renewables and processing region.
2. Dashboards from the contracted cloud: the major providers offer per-customer emissions dashboards (for example, Microsoft's for Azure and Microsoft 365).
3. Your own estimate for bespoke developments: based on the number of queries, the size of the model and the compute time, with open tools such as CodeCarbon.

A rule that avoids arguments: always state whether each figure is measured or estimated, and with which method, as UNE 0086 and ITU-T L.1510 require.

Sustainability KPIs by maturity level

ITU-T L.1510 proposes a simple method: set objectives, choose metrics, set a target for each one, and measure and publish the results over a rolling twelve-month period. It also grades the indicators into three levels, so you can start with a few and expand as you mature.

Indicator	What it measures	Unit	Level
Total energy consumption	Electricity used by the data centre, workstations and cloud services	kWh/year	Initial
Data centre efficiency (PUE)	Total data centre energy divided by IT equipment energy (the closer to 1, the better)	Ratio	Initial
Scope 1 and 2 emissions	CO₂e from the energy consumed by the systems	tCO₂e/year	Initial
Electronic waste managed	Decommissioned equipment that is reused or sent to an authorised waste handler (WEEE)	%	Initial
Systems retired for lack of use	Systems and functions switched off after the annual review	No./year	Initial
Renewable energy (REF)	Renewable share of the energy consumed	%	Advanced
Emissions per AI query	Carbon intensity of each interaction	g CO₂e/query	Advanced
Data centre water (WUE)	Cooling water per unit of IT energy	m³/MWh	Advanced
Carbon efficiency of the use case	Benefit obtained per kg of CO₂e (for example, documents reviewed per kg)	units/kg CO₂e	Advanced
Lifecycle footprint (scope 3)	Manufacturing, purchasing and end of life of equipment and services	tCO₂e/year	Leader

The most useful indicator for decision-making is the carbon efficiency of the use case, because it relates the impact to the value delivered: it lets you compare two solutions and justify, or rule out, a use of AI.

How to reduce the impact of AI

• Choose the right-sized model. Small models for classifying or extracting data; large ones only when needed. A specialised document extraction service is usually more efficient than asking an LLM to read every image.
• Do not use AI when a search, a template or a rule will do.
• Avoid generating images or video unnecessarily.
• Reuse frequent answers and batch-process anything that is not urgent.
• Choose cloud regions with cleaner energy, always within the EU so as not to create data protection problems.
• Retire systems that are not used and extend the useful life of equipment; at the end, reuse it or hand it over to an authorised waste handler.

What to ask suppliers

1. Energy and emissions per query or per user, with the calculation method (UNE 0086 or ISO/IEC 21031).
2. Data centre indicators: PUE, WUE, renewables and location.
3. A periodic report, for example annual, on the consumption and emissions of the contracted service.
4. Model options of different sizes and the justification for the one chosen.
5. ISO 14001 and ISO 50001 certifications and information on their waste management.
6. Use that data as an award criterion or a performance condition (art. 145 LCSP), without it outweighing security and data protection.

Building it into governance

Sustainability works when it is part of the process, not when it is a separate report. Three steps are enough to get started:

• Add to the record of each AI use case the chosen model and the justification for its size, along with energy, emissions and water data (measured or estimated).
• Have the body that authorises AI uses assess sustainability before approving each case.
• Integrate the KPIs into the data governance dashboard and publish them every year, starting from an initial measurement that serves as a baseline.

How to build the indicator dashboard step by step

1. Set the objectives. What you want to achieve: reduce the energy consumption of systems, choose suppliers with more renewables, retire systems that are not used, or measure the impact of AI before it grows.
2. Choose the metrics from the table above that address those objectives, starting with the initial-level ones.
3. Identify the source of each figure: electricity bills, the in-house data centre, the cloud emissions dashboard, supplier certificates or the equipment inventory.
4. Take a first measurement to serve as a starting point. Without a baseline there is no credible target.
5. Set a target for each metric, with a time horizon; ITU-T L.1450 helps set multi-year targets compatible with a 2 °C pathway.
6. Measure and publish over a rolling twelve-month period, always stating whether the figure is measured or estimated.
7. Review every year and move up to the next level when contracts and tools allow.

An example of a carbon efficiency calculation

Imagine a system that extracts data from documents to prepare them for review. In one month it processes 4,000 documents and, according to the cloud emissions dashboard, its consumption is equivalent to 20 kg of CO₂e. Its carbon efficiency is 200 documents reviewed per kg of CO₂e.

If the large language model that read each document is replaced by a specialised extraction service, and the figure rises to 600 documents per kg, the decision is backed by data. And if a year later efficiency falls without any change in use, it is a sign that something has changed in the service or with the supplier.

What matters is not the absolute precision of the number, but always measuring the same way so that you can compare.

Eight pillars for a digital sustainability commitment

Many organisations structure their commitments around pillars such as those proposed by ISMS Forum's Pact for Sustainable Cybersecurity (Pacto por una Ciberseguridad Sostenible). Applied to AI and information systems:

Pillar	Commitment	Indicator
Responsible governance	Assess sustainability when authorising each AI case	Authorised cases with sustainability assessed (%)
Transparency	Publish the AI systems in use and an annual report on consumption and emissions	Report published (yes/no)
Energy efficiency	Right-sized model, avoid unnecessary uses and use regions with clean energy	kWh and g CO₂e per query or per month
Circular economy	Extend useful life and retire what is not used	Systems retired; average equipment lifespan
Electronic waste	Authorised WEEE waste handlers	Equipment handed over to an authorised handler (%)
Collaboration and outreach	Share experiences and include sustainability in training	Actions per year
Sustainable standards	Measure with recognised metrics	Systems with indicators according to UNE 0086 (%)
Supply chain	Environmental criteria in tender specifications	AI contracts with environmental clauses (%)

What to ask when choosing an AI model

Before deciding which model to use for a case, these questions prevent most excesses:

• Is the task to generate free text, or to classify, extract or summarise? For the last three, small models or specialised services are usually enough.
• Does the answer need to be generated every time, or are there repeated questions whose answers can be reused?
• Does the process need an immediate response, or can it be batched during off-peak hours?
• Has the output of a large model been compared with that of a medium-sized one on the test bench? Often the difference in quality is minimal and the difference in consumption is large.
• Does the supplier offer several model sizes and consumption data for each one?

Common mistakes

• Using generic averages for consumption per query instead of asking the supplier for the figure.
• Measuring only AI and forgetting servers, backup storage, workstations and printers.
• Not stating the method: a figure without “measured” or “estimated” and without a methodology is not comparable.
• Letting sustainability outweigh security or data protection in tender specifications: choosing a region outside the EU because it has more renewables can create a bigger problem.
• Measuring once and forgetting about it: without an annual review, the indicators lose their meaning.

At Bidatia we help organisations build sustainability into AI and data governance, from model selection to indicators. If you are interested in the full framework, read which regulations apply to AI and data and how to organise data governance.