Are agents going to transform ESG?

There is a lot of hype around LLM Agents at the moment, with Jensen Huang claiming “AI agents [will be] a multi-trillion dollar opportunity” and with the current consensus that 2025 will be the year when AI agents cross the chasm.

At Briink we build AI tools for ESG and sustainability teams and have been investing heavily in agentic approaches. This is our learnings so far:

1. What could Agents do for ESG? If we define agents as AI systems that can plan, use tools and reflect on the output to iterate until they accomplish complex tasks — then clearly there is insane potential for agentic approaches to totally transform ESG. Like many other knowledge industries (like law or consulting), ESG is absolutely packed with complex, knowledge intensive tasks that currently can only be conducted by humans. For example, conducting ESG due diligence on a company which currently requires a human to manually collect and read large volumes of complex environmental information, extract relevant data, compare it to due diligence criteria, make an assessment of if this meets those conditions, summarise that as a report (and so on…) could be heavily supported by ‘ESG agents’ not only to make this type of tasks orders of magnitude cheaper, but also allow it at extreme scale — i.e. for all companies and in near real-time. We believe this is the way that all knowledge industries are heading over the next decade. 

2. But will this happen in 2025? As others have noted, agentic technology is still very new and while there have been a lot of cool demos and prototypes, bringing agentic systems to production was a challenge in 2024. However we strongly expect the technology for building production-grade agents (agent focused base models, agentic frameworks and toolkits etc.) to mature rapidly during 2025, if nothing else because so much money and focus is being poured into this topic this year.

3. What challenges remain? Despite the rapid progress, there are still large challenges for building robust agentic solutions, especially for complex end-to-end workflows such as those in ESG. The primary challenge here is how to make agentic systems reliably when dealing with the ‘messy’ real world of ESG. Like most knowledge economies ESG has a lot of edge cases and nuances which mean agentic systems need to be robust to a wide variety and variance of ESG data types and topics. We also still need to overcome issues with compound error when building very long, complex agent chains (e.g. small errors early in the chain can compound exponentially) and there are still issues with agentic systems getting stuck in loops or holes when trying to work through real world tasks in ESG. There is still a lot of work to do here, but we think these challenges can and will be largely overcome (potentially during 2025) and we’ll see the world’s first autonomous ESG agents for large workflows like end-to-end ESG research. 

4. How can we already benefit from agents in ESG? Using agentic approaches is not an all-or-nothing problem and we can already benefit from agents in ESG by using them to improve some tasks and sub-systems reliably. For example at Briink we use agentic approaches for more defined tasks such as ESG questionnaire automation, ESRS screening and ESG document retrieval, which dramatically improves the performance of those tasks (compared to static or zero-shot approaches). Eventually these reliable smaller tools and tasks will be orchestrated together to unlock more complex agentic workflows for ESG such as a full ESG Research Assistant. 

5. What doesn’t agents change for ESG? Fundamentally, agents do not change the fact that you still need to deeply understand the specifics of the ESG domain and the unique challenges of ESG teams in this space. ESG is a super complex and evolving field, with a lot of domain specific jargon, workflows, and regulatory requirements. Understanding these is still absolutely crucial to building AI systems that solve real problems in this space, agentic or not.

‍

One of the biggest constraining factors of ESG analysis to date has been resources. Most organisations have to decide on a subset of ESG topics they can cover for a targeted group of target companies. Agents may allow us to re-imagine this with technologies like Briink. What if you could do an ESG assessment not for the top 5% of your portfolio or supply chain, but for every company you look at and in real-time? What could we do in ESG if the ‘cost of intelligence becomes near zero’? 

‍