DC x Sus AI Futures: AI and the Future of Sustainability Reporting
In early April, Sustainable AI Futures and Digital Catapult hosted a workshop day in London on AI and the future of sustainability reporting.
Around fifty participants across industry, academia, and policy gathered to explore the challenges of reporting on the sustainability of AI, as well as the increasing use of AI within sustainability reporting. Speakers, panellists, and session facilitators included Chanell Daniels, Jo Lindsay Walton, Melissa Gregg, Oliver Cronk, Loïc Lannelongue, Massimo Contrafatto, Jamie Riley, Justine Porterie, Alexis Normand, and Shane Brownie. Slides from the keynote and some of the activities are available here.
A couple snapshots: It was a truly interdisciplinary, multi-professional crowd, and very exciting to hear the joyful and occasionally enraged buzz in the room, as teams thought through possible future scenarios for AI and climate, and roleplayed their imaginary start-ups through the perils, pitfalls, and possibilities of the years ahead.
It was also a real pleasure to hear sustainability professionals chatting about the impact of AI on the future of their role. One view was: Yes, AI is coming for our jobs, but that is okay! Sustainability teams were never meant to be so large in the first place. If you’re in sustainability and you want to continue with somewhat similar work in the future, stick close with finance and compliance functions.
What about the use of AI within sustainability reporting? It is clear that, despite many important initiatives of convergence and alignment, the typical sustainability professional still faces a dizzying array of standards, frameworks, and reporting requirements.
A huge amount of sustainability teams’ time is taken up with data collection and reporting, while ideas for driving change get de-prioritized. A substantial amount of sustainability teams’ time is consumed by locating data, cleaning it, reconciling incompatible formats and translating it into the categories required by different reporting regimes.
There appears to be a use case for AI here, helping sustainability teams to process messy and fragmented data sources, map information onto reporting requirements, detect anomalies, and monitor changing regulations and standards. When reporting workloads are high, more ambitious ideas for organisational change can easily be deprioritised.
But even setting aside the environmental impacts of these platforms themselves, there are some big questions. GenAI appears to be a big part of the story, so naturally users are concerned about hallucination, interpretability, and accountability. Sustainability platforms are seldom transparent enough about how they are leveraging AI in their products.
When an LLM needs to draw on a data source under the developer’s control, the most widely used approach is retrieval-augmented generation, or RAG. Relevant material (probably relevant) is taken from a pre-prepared corpus and inserted into the model’s context window before it produces an answer. Retrieval is usually based on embeddings, so it can identify semantically related passages rather than relying only on exact keyword matches.
RAG can improve the relevance and evidential basis of outputs, but it does not remove the non-deterministic core of generative AI. The model may still ignore, misread, distort or embellish the retrieved material, and the quality of the result depends on how sources are selected, parsed, divided, indexed, ranked and presented, among other factors.
Crucially, RAG is often misunderstood. We have repeatedly heard it described as a form of AI that “only looks up answers in the data you give it.” But RAG does not replace generation with lookup. It retrieves material and supplies it to a generative model, which still interprets, combines and reformulates that material probabilistically. The model draws on patterns learned during training—the big, expensive training on the huge datasets scraped from the internet—rather than relying exclusively on the retrieved sources.
Research into more grounded AI systems is developing quickly. . You can equip AI with deterministic tools, you can turn down the temperature to reduce the unpredictability of outputs, you can have LLMs devoted to double-checking the outputs of LLMs. There are a great variety of RAG methods out there, all with their strengths and weaknesses. All this means it’s all the more important that any company providing AI-powered sustainability management and reporting services is transparent about which methods, if any, they are using, and how. Sharing technical detail is the only credible and ethical approach–this applies to AI across many different spaces, but sustainability reporting should certainly be leading the way.
Sustainability reporting often is a messy, approximate art, where you make do with the data you have, and prioritize moving in the right direction, rather than obsessing over measuring everything perfectly. There is a risk that this is used to justify AI-powered bodges and fudges which feel similar (“Well, humans have to make stuff up too sometimes”), but may be far more pernicious. AI offers black boxes and dilutes accountability. Its estimates, workarounds, proxies, and mistakes are not the same as human estimates, workarounds, proxies, and mistakes.
Any use of AI within sustainability reporting needs strong controls: deterministic checks, structured and traceable lineages that provide explicit links between claims and inputted evidence, clearly defined abstention or escalation rules, tools to enable human review where necessary. Above all, providers of AI-powered sustainability solutions need to be much more open about how these systems work. Methods should be presented as auditable technical documentation, not marketing copy.
And a final signal boost: research into existing climate-related reporting is underway, and DBT is interested in the experiences of companies and investors. Get in touch with climatefinresearch@iffresearch.com
Some of the insights from the day will be collected in a short publication, RAI x ESG Compass. If you’re interested in contributing, or being involved in some other way, get in touch.