Why an environmental company uses AI – and why we'd be wrong not to

By our Co-CEOs, Matthew Brown and Rafi Cohen

‍

AI has a real environmental footprint. We take this seriously at Verna. But even though our mission is to accelerate nature recovery, we use AI extensively – and we believe we'd be wrong not to. Addressing the nature crisis means tackling many barriers – political, institutional, operational – but among the most difficult are analytical ones. These are data challenges so large, so complex, and in some cases so intractable, that choosing not to use AI isn't a neutral choice: rather, it's a choice to leave critical environmental work undone.

AI has a real environmental footprint

A single AI query can use anything from 0.3 Wh for a simple request up to around 40 Wh for the most intensive reasoning tasks. Water use is real too: one widely cited estimate puts it at 500 ml for every 10–50 responses in a US-average data centre.

In aggregate, these figures are small relative to transport or agriculture. But data centre impacts aren't evenly spread – at the local level, a single hyperscale facility can strain the power and water resources of its surrounding community.

The numbers are improving as models become more efficient and the grid decarbonises, but that doesn't mean they don't matter. The question is what you weigh them against.

Analytical challenges are holding back nature recovery

Nature is in crisis. Species extinction rates are tens to hundreds of times above natural background levels, and more than half of the world's GDP depends on nature.

Money isn't the biggest barrier to solving this. Over $120 billion a year already flows into conservation, and international frameworks – from the Kunming-Montreal Global Biodiversity Framework to the UK's Environment Act 2021 and the EU Nature Restoration Law – are driving action, alongside growing corporate commitments to nature. The problem is delivery.

The most underappreciated challenge to delivery is analytical capacity. Nature recovery requires decisions based on complex analysis, applied to data that is fragmentary, disputed, and incomplete.

Part of the problem is that there simply aren't enough skilled ecologists, or enough hours in the day, to plan, deliver, and monitor recovery at the pace and scale required.

But it's worse than a staffing issue. Some of these challenges are intractable at scale: not merely slow to do with conventional software, but impossible to do reliably at the volume, speed, and complexity required. No spreadsheet or GIS tool solves a problem that requires interpreting unstructured ecological data, cross-referencing dozens of heterogeneous datasets, or making sense of ambiguous ground-level evidence – all at a pace that matches the urgency of the global challenge.

It's proven that AI can deliver environmental outcomes

This isn't theoretical. Modern AI – the compute-intensive kind drawing environmental scrutiny – is already delivering results across conservation and environmental management.

OnDeck uses vision-language models to review thousands of hours of fisheries monitoring video, identifying species, catch events, and compliance issues without needing labelled training data. Only a small fraction of global fisheries are currently monitored, with the industry projecting a 10–12x increase in vessels needing video review. Every fishery looks different, the data is unstructured video, and the volume is beyond any human team. Their peer-reviewed research was presented at NeurIPS 2024.

Conservation Evidence at Cambridge has spent over 20 years building a database that summarises what works in conservation. They've screened 1.6 million scientific papers and summarised over 8,600 studies - evidence that people delivering nature programmes can use to understand what actually works on the ground. Now they're training large language models to accelerate this: reviewing papers, extracting findings, and adding them to the database at a speed no human team can match. With hundreds of millions of published academic papers – and conservation-relevant findings buried throughout – the task is genuinely intractable at scale without AI.

SpeciesNet, developed by Google and WWF, uses deep learning to identify wildlife from camera trap images – classifying roughly 2,500 species categories with 94.5% accuracy, trained on 65 million labelled images from conservation partners worldwide. Camera trap networks generate millions of images per year; identifying species manually takes months or years. SpeciesNet processes them in days. The model is deployed across Colombia, Peru, Australia, Tanzania, and the US.

These examples span different techniques – vision models, language models, video analysis – but share a common pattern: environmental tasks where the data and expertise exist, but the volume or complexity exceeds human capacity.

The real question is weighing the footprint against the benefits

What's the alternative to using AI to tackle the analytical challenges of nature recovery?

For some tasks, humans could tackle the work without AI – potentially more slowly and at greater cost, but it could get done. There is a whole class of analytical challenges, though, that are close to impossible without AI. Looking across what large-scale nature recovery programmes are attempting, now and in the near future, we see obstacles like:

• Quality-checking site evidence at scale. Biodiversity net gain and nature recovery projects generate thousands of ground-level photos from hundreds of sites, each needing expert review. Any ecologist can check one photo. But across a large portfolio – thousands of submissions arriving in tight reporting windows – no team can review them all. Without AI, either every photo gets rubber-stamped, undermining data quality, or people stop collecting that evidence altogether.
• Making sense of ecological data across a portfolio of sites. Nature recovery portfolios generate monitoring reports, condition assessments, survey data, and compliance submissions across dozens or hundreds of sites, in different formats, from different teams, and against overlapping deadlines. Patterns that only emerge at portfolio scale – systematic failures in particular habitat types, successful interventions worth replicating, seasonal windows closing on high-priority actions – go undetected. Without AI to cross-reference and prioritise across the full information stream, opportunities are missed and mistakes repeated.
• Connecting site decisions to the global evidence base. An ecologist encounters unexpected soil conditions on a restoration site. The relevant evidence exists – somewhere in hundreds of millions of published papers. No human can search that at the point of decision. AI doesn't replace the ecologist's judgement about what to do, but it gives them access to evidence they'd otherwise never find.

Human expertise remains essential. An ecologist's field knowledge, professional judgement, and contextual understanding can't be automated. But the analytical heavy-lifting around those decisions hits limits no human team can overcome at scale.

If we don't use AI to help with these challenges, then the work simply doesn't get done. No environmental footprint of AI usage… but no nature recovery either.

Verna is using AI to accelerate nature recovery

Verna exists to accelerate nature recovery globally. AI isn't what defines us, but it's a vital tool to achieve our mission.

Today, our Mycelia platform uses AI to review habitat classifications based on ground-level survey photos. It flags photos that may need an expert ecologist's attention. Thousands of users rely on this. Without AI, either every classification would need manual expert review – impossible at scale – or errors would go uncaught.

We're working now on AI solutions to some of the tough challenges discussed in this article. We want to enable anyone with large-scale nature recovery objectives, such as corporates with ambitious nature goals, to achieve them with speed, reliability, and confidence.

Our work with AI is guided by three key principles:

• We use AI where it creates genuine environmental value, not for its own sake.
• We are transparent about where AI is and isn't used in our products.
• We build systems to give expert humans superpowers, not to replace them.

Nature is too important to dismiss AI

The nature crisis won't wait for a perfectly clean AI model. Every week we don't use the best available tools is a week of nature recovery delayed. Environmental companies have a duty to use the most effective tools, including AI – thoughtfully, honestly, and in service of the mission. Not because AI is the answer to everything, but because for a critical class of analytical challenge, it's the only way the work gets done at the scale nature needs.

The common debate is: "AI has an environmental cost – should environmental organisations use it?" But the deeper question is: "What happens to the environment if this analytical work doesn't get done?"