Generative AI has moved from pilot to production in record time. Boards expect impact. Regulators expect transparency. Stakeholders expect credible climate action. The question for leadership is simple: where does GenAI fit in your decarbonization story?
Most organisations are deploying GenAI to transform productivity, decision making and customer experience. However, every prompt, every model, every deployment consumes energy and carries an emissions footprint. If you lead finance, sustainability, technology or operations, this is now a governance issue, not only a technical one.
In this article, our experts have set out a leadership view of how GenAI affects emissions, where the real risks and opportunities lie, and how to build GenAI into your net zero strategy rather than alongside it.
Many enterprises still treat GenAI as a pure innovation topic. In recent surveys, only a small share of business leaders list sustainability impact as a top concern in GenAI deployment. That gap is becoming increasingly risky.
GenAI is scaling rapidly across functions. As usage grows, so does the underlying compute, data movement and power draw. At the same time, GenAI is one of the most powerful tools available to optimise operations, reimagine supply chains and redesign carbon intensive processes.
In other words, GenAI can increase emissions in one part of the organisation while driving reductions in another. Leadership teams need to understand this full picture before approving investments, setting KPIs and signing off climate disclosures.
There is a common belief that the main environmental issue with GenAI is training large models. Training does consume significant compute, but for most corporate users that is not where the bulk of emissions will sit.
From a life cycle perspective, there are five main emission drivers:
Large, general-purpose models are trained on vast datasets with substantial compute intensity. The emissions associated with this phase can be allocated across all enterprises that license and use the model.
Models tailored to domains such as software development, tax, legal or engineering also require significant design and build effort, which translates into energy use.
Most organisations now customize licensed models by securely embedding internal data, documentation and expertise. Fine tuning and retrieval pipelines add another layer of compute and energy consumption.
This is where the emissions profile changes for heavy users. Day to day usage, that is, user prompts and model responses, can cumulatively exceed the training footprint over time. Enterprise-wide adoption, multi-step agents and always-on assistants can amplify this effect.
Core platforms such as ERP, CRM, HR and collaboration tools now include embedded GenAI. As adoption grows, these hidden inference calls become part of your digital operations footprint.
Analyses conducted by large early adopters show a clear pattern. For heavy corporate users, model inferencing is already the dominant contributor to GenAI related emissions, while the allocated share of model training and customization is significantly smaller. Importantly, even in ambitious rollout scenarios, total GenAI emissions can remain a relatively small share of an organisation’s overall footprint, in some cases comparable to or lower than business travel.
The leadership task is therefore not to halt GenAI, but to consciously shape where and how it is used.
GenAI is also a powerful decarbonization engine when applied strategically. Used well, it can make operations more carbon efficient and enable decisions that were previously too complex to model. Examples include:
GenAI can automate analysis of energy use, maintenance logs and process data, flag inefficiencies and suggest low carbon alternatives, from route optimization in logistics to better loading patterns in industrial operations.
Combined with analytics, GenAI can run through millions of permutations of potential decarbonization actions, investment options and technology pathways. This helps leaders understand trade-offs and sequence actions in line with budget and impact.
As workflows are redesigned, GenAI can reduce manual work and in some cases replace less efficient legacy compute. The net result can be a lower overall IT footprint even with GenAI in the mix.
GenAI can accelerate data collection, quality checks and narrative generation for emissions reporting and climate disclosures, freeing specialist teams to focus on strategy rather than data wrangling.
The key is to treat these opportunities as core business outcomes, with clear metrics and accountabilities, rather than side benefits.
The emissions profile of GenAI is not static. Several technology trends are already reshaping the balance between performance and energy use:
Advances in model architectures and compression are enabling smaller, more efficient models that deliver strong performance with significantly less compute. For targeted use cases, these can be preferable to very large general models.
At the same time, organisations are moving towards multi-step, agentic workflows instead of single prompt responses. These orchestration patterns can increase inference calls, latency and energy use if not designed carefully.
Domain specific models for niche tasks can deliver better accuracy but may require additional training, maintenance and infrastructure. The resulting model sprawl can increase the overall compute footprint.
New processor classes such as application specific accelerators and tensor processing units can deliver more energy efficient compute for AI workloads than general purpose GPUs, especially at scale.
Early research indicates that quantum approaches could eventually enable certain AI calculations with a fraction of today’s energy use. While still emerging, these trajectories are important for long term strategy.
Leadership teams should view GenAI architecture decisions as climate decisions. Choices about model size, hosting, hardware and application design all influence the emissions trajectory.
To bring GenAI into responsible use, leaders can anchor on four core actions:
Build a structured approach, ideally within a life cycle assessment framework, to estimate emissions for GenAI initiatives. Include model training (where relevant), customization and inferencing, and allocate emissions across your overall carbon budget.
Assess the embedded emissions in pre-trained models, your own tuning and integration work, and day to day application usage. Consider both cloud and on-premise setups, data centre locations and energy mix.
Treat emissions as a design constraint from the start. When evaluating GenAI models, vendors and use cases, test alignment with your climate targets. In some situations, a conventional, lower compute solution may be more appropriate.
GenAI rarely acts alone. It interacts with IoT, advanced analytics, automation, blockchain based systems and sector specific technologies. Evaluate how these combined stacks can drive decarbonization, and where they might unintentionally increase demand.
By embedding these steps into governance processes, organisations can move from ad hoc GenAI experimentation to disciplined, climate aligned deployment.
IFRSLAB works with boards, CFOs, sustainability leaders and CIOs to operationalize responsible innovation. Our teams help organisations quantify the emissions footprint of GenAI, design architectures that are efficient and climate aligned, and use AI to unlock real decarbonization across operations and value chains. If you are looking to align GenAI strategy with net zero commitments, IFRSLAB can help you move from high level ambition to measurable, auditable outcomes.
Explore key data insights on corporate readiness for ESG adoption, revealing trends, challenges, and practical steps businesses can take to prepare effectively.
Learn how strategic leadership drives sustainable deployment of ESG initiatives, enabling responsible growth, accountability, and long-term value creation for modern businesses.
Discover practical ESG reporting alternatives for SMEs that are cost-effective, simple to implement, and help meet sustainability goals without complex enterprise systems.
Typically replies within a day
