AI is now central to how sustainability work gets done.
Language models pull emissions data out of thousands of supplier disclosures that no team of analysts could read at that scale. Satellite imagery, analysed by pattern recognition, spots deforestation, fires and land use change in near real time across supply chains that span dozens of countries. Machine learning finds links between weather, insurance claims and asset performance that manual analysis would miss. Sensor networks produce continuous emissions data far more detailed than monthly reporting ever could.
This has been one of the most important shifts in corporate sustainability in the past five years. Organisations that have invested in it now have data that is more accurate, timely and verifiable than was previously possible at any price. Sustainability performance that companies simply report about themselves is fading, not because regulators have banned it but because the tools to check or contradict those reports are now widely available.
Agriculture and consumer goods supply chains show this most clearly. Olam, which sources across dozens of countries, has built satellite monitoring and farm level data platforms that give verified sustainability information at a scale and level of detail written supplier questionnaires never could. Satellite verification has changed what sustainability data means: no longer a statement of intent, but a measurement of where the company actually stands.
The paradox at the centre
At the centre of this transformation sits a paradox.
Training large AI models uses enormous amounts of electricity. Running the data centres needed for AI at scale uses far more. Cooling those facilities takes significant amounts of water, and manufacturing the specialised chips carries its own emissions and resource footprint.
Microsoft reported in 2024 that its emissions were about 29% higher than in 2020, and it has been open that much of the increase comes from building the data centres that AI and cloud growth require. Other large cloud and AI providers face versions of the same problem.
For organisations committed to cutting their footprint, the energy used by AI powered sustainability analytics cannot be exempt from the scrutiny applied to other infrastructure. It must be measured, managed and accounted for.
This is not an argument against using AI in sustainability. It is an argument against assuming that AI automatically delivers a net benefit, whatever the implementation choices.
Three disciplines for AI in sustainability
First, know your energy source. The footprint of AI depends heavily on the energy mix of the data centres running it. The same workload has a very different profile on renewable power than on a grid dominated by fossil fuels. Cloud providers increasingly publish the energy sources behind specific regions and services. If you use AI for sustainability analytics, treat the energy behind it as a sustainability variable, not a procurement afterthought.
Second, be disciplined about efficiency. The energy AI uses varies enormously with model size, retraining frequency, optimisation and architecture. Bigger models use more. Frequent retraining uses more. Running a model at scale over its lifetime can use more than training it did. Engineering choices translate directly into footprint, so apply the same discipline to AI’s own impact as to any other infrastructure investment.
Third, account honestly. If you report that AI analytics cut supply chain emissions by a certain amount, the emissions of the AI itself belong in that calculation. In most well designed implementations the net result will still be positive, but the net figure has to be calculated. Anything else is selective reporting that would not meet the standard now applied to other sustainability claims.
A leadership problem, not a technology problem
Beyond the operational detail, there is a wider lesson.
The sustainability transition is a complex systems problem full of uncomfortable choices. Treating each new technology as a solution without counting its costs has been a recurring habit of the voluntary era, and it is one reason the gap between sustainability intention and performance has lasted so long.
Mature sustainability leadership needs the opposite habit: being clear about what each decision costs as well as what it delivers, measuring both, and choosing on the honest balance.
The AI paradox in sustainability is not a technology problem. It is a leadership problem.
AI is neither the solution to the sustainability transition nor an obstacle to it. It is a powerful tool whose net contribution depends on how it is deployed. Organisations that bring AI into their sustainability work with that honest framing will build systems that perform well and stand up to scrutiny.
Those that assume AI must be a net positive, because it is new, sophisticated or fashionable, will end up with deployments that may or may not help, and that in some cases will undermine the very goals they are meant to serve.
Draws on Chapter Seven of my book, Sustainability Leadership: The Global Outlook.