As a highly data-rich field, it was only a matter of time before environmental, social and governance (ESG) reporting would cross paths with artificial intelligence (AI).

Their nascent marriage has certainly made a strong impression on Neil Robson, Regulatory Compliance Partner for Financial Services in the London office of US law firm Katten Muchin Rosenman. He says that in the past few months, he has seen an “explosion” of clients using AI for ESG tasks.

While ESG is not yet heavily regulated in the US, it is increasingly subject to EU law. And providing global businesses with region-tailored advice has become a key source of business for the firm’s London office.

“We’re aware of AI applications that are designed to help businesses review portfolios and stats, and generally make life easier for compliance and ESG professionals,” Robson says.

“A client may have compiled one set of reports under the guidelines from the Task Force on Climate-related Financial Disclosures (TCFD) and must then create another set to comply with the Sustainable Finance Disclosure Regulation (SFDR). In that case, the software is able to work out what it can take from the TCFD reporting and what it must do differently to meet the requirements of SFDR.”

In other words, AI is helping to drive some major efficiencies, Robson says: “There are lots of standards and rules out there – and once the software knows how to address each one, it can do so over and over again.”

Scattered sources

Perhaps not surprisingly, ESG-focused AI tools have become pet projects for some of the world’s biggest technology brands. IBM’s Envizi, Microsoft’s Project ESG Lake and Salesforce’s Net Zero Cloud are all in the process of making a splash.

Alongside those offerings is Briink: a generative AI (Gen-AI)-powered document analysis suite aimed at ESG teams. “There has been a massive ballooning of both the number of companies that must be assessed under ESG rules and standards, and of the amount of data they must collect and verify,” says Briink CEO Tomas van der Heijden, who co-founded the company in 2021.

“The problem is that 90% of that data is unstructured – scattered between sources such as website text, PDFs, Excel files, emails and so on. As a result, a significant amount of that data is not being assessed.”

Van der Heijden cites research from Boston Consulting Group showing that, in 2022, just 10% of companies measured their full greenhouse gas emissions: an improvement of only 1% on the previous year. “So, this is a key area where AI is coming in – assessing those large volumes of documents and datasets and using automation to support human analysis.”

However, manual processes are highly inefficient, van der Heijden says. “Some of our asset management clients have told us that, on average, it takes four to six manual hours to analyse the documents of a firm they may be interested in. We’ve also heard that figure from sustainable procurement teams at large, industrial manufacturers with many suppliers. But AI cuts those assessments down to an hour.”

Power hungry

But there is a downside. For a technology that is setting out its stall in ESG, AI itself is quite power intensive. In a recent blog, energy efficiency startup Zodhya worked out that it took developer OpenAI a whopping 1,064MWh to train the software at the heart of GPT-3, forerunner of its ChatGPT platform. Zodyha puts the output of ChatGPT’s efforts to fulfil users’ requests at 260.4MWh per day.

John Ridd is Co-founder and CEO of Greenpixie, which helps companies to measure and reduce the carbon emissions of their cloud computing networks. “When you onboard ChatGPT for enterprise use, there’s a very good chance that you will no longer be running it on your browser. Instead, you will run it in your own cloud environment, such as Microsoft Azure. That’s when you’ll start to incur the sustainability cost.”

Power consumption also differs by data type. “It wouldn’t be zero by any means, but text-based actions – where you’re querying and working off your own data – wouldn’t be as intensive as image-based work,” Ridd explains. “That’s where the computation can be a lot heavier. I recently learned that a single image produced on ChatGPT’s DALL-E platform takes the equivalent of one mobile phone charge to generate.”

Seeing things

Power use is not the only caveat to bear in mind, Robson warns: “With any form of Gen-AI, you will always run the risk of so-called ‘hallucinations’. That’s when a fairly minor flaw or assumption creeps into your inputs and the software then magnifies it – with the potential that it could create misleading outputs: reports that look truthful, but aren’t. My position is that the buck always stops with the human being.”

“Companies like ourselves have been developing techniques to safeguard against that risk,” van der Heijden says. “The main one is to encourage the software to focus. So, don’t throw an AI model at a set of documents and say: ‘Assess this material in the context of the entire internet.’ Instead, say: ‘Use the reasoning engine you have developed from analysing the entire internet, but concentrate only on what’s in front of you.’ 

“If you are analysing a sustainability report and ask, ‘What are the Scope 1 emissions of this company?’ and there’s nothing about Scope 1 in the files, the results should reflect that, rather than make up an answer that doesn’t exist.”

ICAEW Head of Tech Ian Pay believes the use of AI tools for ESG work is a double-edged sword. “Something with huge power and capability to put structure around unstructured data will make a big difference in ESG. AI has long been touted as part of the solution for assembling disparate data sources to meet inconsistent reporting requirements. It can also assist with the more forward-looking, predictive elements of ESG risk management.”

However, the environmental impact of Gen-AI cannot be ignored, Pay adds. “It was one of the most pressing questions raised at ICAEW’s Annual Conference last year – and right now, the developers don’t really have clear answers on how to mitigate those impacts. The key question when it comes to using Gen-AI should always be ‘Is this the right tool for the job?’ Given its intense consumption of energy, we must be mindful of using Gen-AI for tasks it is not best suited to.”