EA: There's an uncomfortable irony emerging in that the professionals tasked with decarbonising business operations are increasingly running their analysis on some of the most energy-hungry infrastructure in tech. Can you credibly advise on emissions reduction while your own toolkit's footprint goes unmeasured?

KB: It all comes down to staying mindful and in control of which models you actually use for each task. The industry's default habit - sending every task, no matter how trivial, to a trillion-parameter frontier model - is the AI equivalent of driving an articulated lorry to the corner shop. You don't need a trillion parameters to summarise an email.

Most of the work sustainability teams do with AI: extracting data from reports, drafting sections, classifying documents, can run on small, fine-tuned models at a fraction of the energy cost. A model a hundredth of the size, tuned to your specific task, will often do the job better and emit a tiny fraction of the CO2. But you can only make that choice if you're in control of your AI stack. If you're renting a general-purpose model through an API, you take whatever's behind the curtain, you can't right-size, you can't measure, and you can't account for it.

So the answer isn't to use less AI. It's to treat model selection the way you'd treat any other procurement decision with an emissions line attached: match the tool to the task, measure what you can, and own the infrastructure choices rather than outsourcing them blind.

 

EA: You've talked about a single complex query using as much power as an LED bulb left on for four hours. How does this kind of hidden energy use scale up, and why is it so often invisible to the people generating it?

KB: The scaling problem is one of invisibility multiplied by volume. That four-hour LED bulb figure sounds trivial for a single query, and that's exactly the trap. One person running one query is nothing. But an organisation running thousands of queries a day, across hundreds of staff, every working day of the year, has quietly built itself a new energy load that never appears on any meter it controls.

And that's the core of why it's invisible: the emissions happen somewhere else. When you drive to a site visit, you buy the fuel, the cost and the carbon are yours, visibly. When you send a query to a cloud-hosted model, the electricity is drawn in a data centre you'll never see, possibly on another continent, wrapped into a subscription fee that tells you nothing about consumption. The energy cost has been abstracted away by design. You're billed per seat or per token, not per kilowatt-hour.

It gets worse when you look at behaviour. Because each query feels free and instant, nothing discourages waste - people re-run prompts five times to tweak a sentence, use the largest model available for the simplest task, and leave AI features running in the background of tools they didn't even choose. There's no feedback loop. Nobody would leave a tap running because water feels free; with AI, the tap is invisible and the meter belongs to someone else.

For a profession built on the principle that you can't manage what you don't measure, that should be uncomfortable. The first step is simply bringing the load into view, knowing which models you're using, where they run, and what they draw. That's far easier when the infrastructure is yours: a model running on your own hardware, or in a deployment you control, has a power draw you can actually read. Rented intelligence hides its meter; owned intelligence puts it on your wall.

 

EA: The Summit's agenda has a strong focus on demonstrating value to multiple stakeholders and shifting the mindset from "cost" to "investment". How does that same thinking apply to AI infrastructure decisions?

KB: It maps almost perfectly, because AI infrastructure is where the cost-versus-investment distinction becomes brutally literal. Renting AI through subscriptions and API calls is a pure cost, an operating expense that repeats forever, scales with your usage, and leaves you with nothing. Five years of subscription fees buys you five years of access and zero equity. The moment you stop paying, the capability vanishes. That's the definition of a cost.

Building AI you own inverts that. Yes, the upfront number is bigger, there's development, fine-tuning, deployment. But what you're left with is an asset: a model trained on your proprietary data, embedded in your workflows, that gets more valuable the longer you use it.

The stakeholder framing sharpens this further. When you rent, the value accrues to the vendor - your usage data, your feedback, your dependence all strengthen their asset, not yours. When you own, every stakeholder can point at something real: the CFO sees a depreciating asset instead of an endless OpEx line, the sustainability lead sees infrastructure they can measure and account for, the board sees IP that adds to the company's valuation rather than a dependency that adds risk.

And the risk side is the part the 'cost' mindset always misses. Renting concentrates risk you don't control — the vendor deprecates your model, changes pricing, gets acquired, or decides your legitimate professional queries violate their usage policies. We've seen all four happen in the last six months.

 

EA: Throughout the Summit, delegates will be discussing how AI and digital tools can accelerate planning, consenting and risk modelling for infrastructure projects. Is there a tension between using AI to speed up sustainable delivery and controlling AI's own environmental footprint, and how should delivery teams think about that balance?

KB: AI doesn't compete with a sustainability professional's impact - it multiplies it. The professional who can interrogate ten years of environmental monitoring data in an afternoon, model twenty design alternatives instead of three, and spot the flaw in option twelve is simply a more effective agent of decarbonisation than the one buried in spreadsheets. And the emissions at stake on either side of that equation aren't remotely comparable. The compute behind a risk model is measured in kilowatt-hours; the design flaw it catches, the re-poured concrete, the redundant earthworks, the consenting delay that keeps a high-carbon asset running months longer, is measured in tonnes. When AI is pointed at consequential work, the emissions it helps prevent sit orders of magnitude above the emissions it generates.

That's the right way to frame the balance: not AI's footprint in isolation, but AI's footprint relative to the footprint a sustainability professional is there to eliminate. A tool that emits kilograms while enabling decisions that avoid tonnes is not a tension — it's leverage.

But that favourable ratio isn't automatic; it's earned through the discipline we've been talking about throughout. It holds when AI is deployed deliberately: right-sized models, consequential tasks, infrastructure you can measure. It erodes when a trillion-parameter model idles behind every trivial email, generating emissions with no prevented tonnes on the other side of the ledger. The unmeasured, indiscriminate use I described earlier is precisely how the ratio collapses.

So the balance delivery teams should strike isn't 'how do we limit AI to control its footprint', it's 'how do we keep our AI use on the right side of that ratio.' Deployed with intent, a sustainability professional amplified by AI prevents far more carbon than their tools emit.

 

You can watch Brainpool AI's latest Environment Analyst webinar Own Your Intelligence: Why 95% of AI Projects Fail, and What the Environmental Sector Can Do Differently on-demand here.