What tech teams need to know about navigating AI and sustainability

We’ve spent an increasing amount of time talking to development teams, infrastructure engineers and sustainability managers about AI’s environmental footprint. What strikes us isn’t the complexity of the problem but how familiar the solutions feel. Tech professionals have always optimised for efficiency. That muscle memory and instinct, honed through decades of squeezing performance from limited resources, is exactly what tech needs now to optimise sustainability.

Our recent report Powering Artificial Intelligence found that even in its baseline scenario, data centre energy consumption will nearly triple between 2023 and 2035, with AI hyperscale data centres being the primary growth driver. In the more aggressive adoption scenario, new data centre power consumption would require 24% of the projected increase in European renewable electricity over the same time-period.

To mitigate the expected growth in energy and environmental impact of AI, we need to look at optimisation and streamlined deployment. Recent studies show that almost all of the lifetime environmental impact of a generative AI model comes from its use-stage and inference, tech professionals have a real opportunity to address the largest area of AI’s environmental impact and optimise its use for efficiency and sustainability.

Sustainable AI isn’t a nice-to-have; it’s how we keep systems running without breaking budgets and the planet. Here’s what we’ve learned about making AI sustainable in practice.

The problem with measuring AI environmental impact

Last month, we watched a demo of a new generative AI application. The demo went well, but when someone in the audience asked about the environmental impact of their AI usage, the team had no way of answering. This happens constantly. Everyone is aware of the energy impact that AI data centres represent, and they know to ask about it. However, it’s a question that is (at the time of writing) unanswerable without a string of asterisks and caveats.

Historically, AI providers have not provided energy or water consumption metrics for their models. This means that all estimates of training and inference energy utilise assumptions and lack direct measurement, limiting the applicability and usefulness of results. It’s only recently that some providers have begun publishing the energy costs of training and using their models. This an amazing advancement, but even in these best cases, the lack of a standardised industry approach and methodology means usage of these figures faces a similar challenge to cloud sustainability dashboards: nuanced differences in methodology that make comparisons impossible.

However, this doesn’t mean you shouldn’t look to measure your AI environmental impact — measurement and reporting are key parts of optimisation. We all fundamentally understand that AI does have an energy impact, and that our contribution to that impact is connected to consumption. Since direct measurement isn’t available yet, using proxies to indicate directionality allows us to measure and connect improvements in efficiency to optimise AI impact.

Token consumption per model type, for example, matters. Every provider bills on tokens, but not every token is equal; tokens from chain-of-thought or reasoning LLMs have higher energy usage (based on academic research) than lightweight models. So, tracking token consumption should be done per model, and you can consider weighting parameters in your tracking to account for size on a per-model basis.

Infrastructure patterns can reveal waste. Cloud billing dashboards show usage by service and region. We can optimise these by looking for underutilised resources, inefficient batching and (where feasible) the potential to move workloads running in high-carbon-intensity times and regions to less intensive ones when scheduling and allocation can be implemented without business impact (not every compute task needs to be done right this second!).

But functional units matter most. Measure and optimise against the key metrics that matter to your business. By optimising your resource consumption against these key metrics (for example per transaction, end user, API call, revenue), you can monitor and show sustainability and efficiency improvements against the key KPIs that your organisation cares about, and track improvements in intensity against the same.

The point isn’t perfect measurement. It’s about establishing baselines, acting on insights, and tracking improvement. But where is the best place to look at optimising impact? With inference and usage accounting for up to 90% of the lifetime emissions of an AI model, optimising our usage can create AI systems that are as sustainable as possible.

Using existing fixes for AI sustainability optimisation

An IT team we worked with built an AI assistant that handled enquiries brilliantly in testing. Production deployment told a different story, with costs and token consumption far exceeding their projections. The culprit? A general-purpose complex AI interface analysing sentiment, retrieving information, generating responses and formatting output, all through one heavyweight reasoning model.

The fix involved techniques most developers already know. First, they broke the monolith into components — the size of the model is a key aspect of the environmental impact of AI systems. So instead of one model doing everything, they built an agentic, modular system with each component right sized for its job. Lightweight models for rapid retrieval, receiving distilled request information from a router agent communicating with the customer facing reasoning model designed for customer engagement and requiring advanced capabilities.

Then they applied standard optimisations: semantic compression to reduce input/output token counts; prompt caching to minimise redundant processing; context management to trim unnecessary history; parameter tuning to dial in the sweet spot between quality and resource consumption.

This sounds great, but agentic architectures are not a silver bullet solution and represent their own sustainability risks. Without proper guardrails, they can spiral out of control. Agents triggering each other endlessly, creating and cancelling prompts, burning resources on unproductive loops. Therefore, critical controls must be included in agentic workflows, with examples including manual triggers for important workflow transitions (not everything needs to be autonomous), rate limits and guardrails on agent interactions, and contextual logic determining when agents should proceed.

How AI can help with sustainability

Most client conversations about AI and sustainability start with a straightforward question: how can we use AI to meet our sustainability goals? The answers to this question are entirely dependent on the maturity of their tech estate in data collection and management.

A key statement that all tech practitioners need to understand is that sustainability data is just operational data. It needs the same controls and governance as have been applied in financial reporting for years. Sustainability teams without a robust data foundation spend hundreds of hours each reporting cycle consolidating energy consumption from disparate sources: different schemas, inconsistent update schedules and manual format conversion. Many reports and calculations end up rebuilding analytical foundations from scratch or siloed to single key individuals.

Many organisations still manage environmental data through spreadsheets and manual consolidation. This parallels the early days of financial reporting, and applying the same architectural best practices can transform these workflows. Automated ETL pipelines collect environmental data on consistent schedules with proper validation (no more manual reconciliation). Master data management ensures consistent definitions across sources, allowing for consistent and logged application of conversion factors. Semantic models enable analysis while maintaining auditability. Regulatory compliance becomes a structural feature, not a manual check.

For you

Be part of something bigger, join BCS, The Chartered Institute for IT.

Once proper foundations are in place, AI sustainability applications become possible, and we can start discussing how AI can make a measurable difference for sustainability.

One of the areas that has been coming up in recent client conversations is using AI to address software modernisation. Technical debt never seems to shrink; legacy systems accumulate patches and workarounds because fixing underlying issues takes more time than teams have. Full time refactoring isn’t feasible when pressure focuses on feature delivery, all while data inefficiencies and bolt-on patches waste compute, storage and carbon emissions. 

To address this challenge, we’ve recently been exploring the use of AI coding assistants to help developers modernise software catalogues through AI bolstered capacity. Recent research found that simply incorporating resource efficiency goals into the coding assistant used in the study produced measurably more energy efficient code (although the sample problems were simplistic). With the usage of AI assistants to modernise legacy software estates at scale, there could be significant sustainability benefits to be unlocked, alongside the other benefits represented by a modernised tech estate.

Active AI resource management in physical infrastructure represents another key area we’re discussing — usage AI connected to sensing capability to optimise infrastructure in real time with sustainability benefits as a measurable outcome. Modern data centre management platforms are prime examples of this capability, using AI to process data streams from cooling systems, power distribution, workload scheduling and environmental sensors simultaneously, and optimising system outputs for efficiency by adjusting airflow, redistributing workloads, and predicting maintenance needs before failures occur.

A key movement within these systems is shifting to carbon aware capabilities, such as connecting to electricity grid data and shifting resource intensive tasks like model training to periods when renewable energy supplies peak. AI has an unparalleled ability to ingest and connect data streams, and resource management represents a key archetype for AI used for sustainability, with similar opportunities for equivalent automation in IoT, supply chain management and other connected organisational systems.

AI has a huge capability to realise sustainability value, but often when it is actually aimed towards sustainability the systems only work when sustainability professionals guide their design. A recent example we saw had an IT team building an AI assistant to help salespeople communicate product sustainability credentials. Without sustainability professionals’ input, it overwhelmed reps with environmental data they couldn’t contextualise. After incorporating sustainability expertise into the design process, the requirement was identified for the final system to automatically connect product characteristics to prospects’ published commitments, and input was provided to enable the AI to effectively articulate this to sales teams. Salespeople could now speak confidently about sustainability without becoming experts themselves.

How to start with AI for sustainability

This is all valuable in theory — but how might you go about including sustainable AI in practice? Start by thinking about what you can accomplish within your realm of control.

• This week, you could test optimisation in existing AI deployments and talk to your sustainability team about their current data challenges. Ask their stance on quantifying AI impact.
• Over the next two months, you could implement measurement using available data and proxy metrics, then work with sustainability colleagues to translate these into actionable sustainability aligned optimisation initiatives. Include sustainability benefits in project business cases.
• Pilot AI-driven resource optimisation in a legacy system and evaluate the potential scale of resource improvements that could be addressed within your legacy software estates, including sustainability into modernisation arguments. Start evaluating the potential for existing organisational systems to be angled towards solving sustainability challenges and building a sustainability data foundation.
• Build for the long term by fostering genuine collaboration between technical and sustainability teams. They understand regulatory requirements and impact calculations. You control resource consumption through architecture and operations. Neither succeeds without the other.

The question isn’t whether sustainable AI will become essential. It already is. The question is when IT professionals will be recognised as crucial to sustainability strategies. Optimisation skills position you perfectly to solve environmental challenges while building more efficient, resilient operations. The principles that make you effective in technology make you uniquely qualified to drive sustainability and make an impact that matters in your day-to-day. If tech professionals can get Doom to run on a toaster, optimising tech and AI for sustainability is really just another challenge!