By day, Alex Annaev is a sustainability manager at Shell. By night, he creates AI chatbots designed to help companies untangle their responsibilities around the EU’s Corporate Sustainability Reporting Directive (CSRD).
His CSRD Chat tool – which he stresses is entirely his own work and completely separate from his role at Shell – is trained on hundreds of pages from the regulator, hours of official guidance published in video format and any additional documentation that comes from the European Financial Reporting Advisory Group (EFRAG), allowing companies to ask it questions about the regulation.
He talks to IR Magazine about his motivation in building the tool, the challenges in creating a reliable chatbot for ESG and why AI holds huge promise for easing the sustainability data burden.
What motivated you to create this chatbot?
My motivation was twofold: enabling the efficiency of reporting practitioners and efficiency of small and mid-sized companies that are subject to CSRD.
I come from a sustainability reporting and assurance background – I was with the Big Four for eight years – and I used to do a lot of reporting manually, which often meant staying late at night in the office. I remember those days and I know the effort it takes so when you see CSRD, with more than 1,000 data points you need to analyze, that is quite a daunting task.
Not every company can afford EY and Deloitte to provide them with guidance, so I thought a tool that works like an expert assistant to provide some direction around the regulation would be really useful. This is about democratization of knowledge.
It doesn’t replace professional ESG or legal advice, though. In certain cases, you still need to look into the nitty gritty, but if you’re a small or mid-sized company – and of the 50,000 companies for which CSRD is applicable, there will be many that are not mature in terms of sustainability or lack resources – this can help to start the reporting journey and keep it going.
What are some of the key challenges for companies with this regulation?
Companies have a lot of questions: what is a double-materiality assessment? What do I need to do to complete that? When do I need to comply? How do I comply? There are lots of basic questions to start with. And if you go to the documents directly, you’ll see there are more than 600 pages: it’s impossible to digest.
This chatbot works as if you were asking an expert working in the reporting area – the response you get is delivered in a very human, understandable way. What ChatGPT is good at is at summarizing. This taps into the 600 pages, four hours of video content and all the additional guiding information that is being released by EFRAG.
What challenges did you face creating this tool? We hear a lot about trust issues with ChatGPT output, for example.
The ChatGPT interface is quite user-friendly, so it’s not that hard to create a tool based on that. Instead, it’s more about curating the content. There is a real risk with an open ChatGPT that just taps into the internet that you will get unreliable responses.
To eliminate this risk, I keep the tool trained on official resources only. This means you have to track all the news coming from the regulator and feed it into the bot to train it. What really takes time is converting official documents into the right format: it prefers Word documents to PDFs, for example, so I had to convert all the PDFs. And sometimes it just doesn’t want to take documents at all but it is improving and learning fast.
I’m trying to get feedback from users so I’ve asked them to send me questions – I’m just writing to people on LinkedIn for any question that is top of mind about CSRD – and I will come back to them. Then I review responses and cross-check them against the source to see whether it’s detailed enough, what further information could be given. If it’s not detailed enough, I give instructions to the chatbot to use a particular document, for example.
This is a public chat and some companies may have concerns about data privacy – and they could create their own chatbots within their own companies for different reporting tasks. But for some smaller or mid-sized companies, it’s easier to use something public, rather than build something internally. Of course, users should not share sensitive information with the chatbot.
How important is the framing of questions here?
It depends on the information you want to get. If you ask a basic question, such as, ‘What are the steps to conduct a double-materiality assessment?’, you will get a generic response and that’s good enough to start thinking and planning. But if you’re interested in a specific indicator, such as making disclosures about your workforce, and you’re wondering whether or not you need to include contractors, then you really need to be specific. And sometimes you need to ask a couple of follow-up questions to get into the information you need.
If you want to show this to your legal team, though, you need to go to the source document and actually read the disclosure requirements.
Does this tool help with directing you to the source as well?
It will give you the source but it will not give you the full text. It’s a simplified answer. If you want to stick to the legal language, you need to read the original documents.
From an IR-specific viewpoint, what is the biggest impact from CSRD, and how can a tool like this help IROs with that?
In the short term, this will help in improving the efficiency of specific reporting tasks. If you do a materiality analysis, for example, you need to scan quite a lot of documents to pull out the necessary triggers for impact, risks and opportunities. And generally, ChatGPT is good at that. If you need to do a compliance check to see how your report is aligned with the reporting requirements, that’s another area where [AI] can help.
Finally, what further potential is there in tools like this?
I’m actually working on another tool where you could upload your report and then scan it against the requirements to get a rating that you are 50 percent or 60 percent compliant, for example – and where your gaps are. Something like that can really help improve efficiency.
In the future, it could also help with interoperability, because there are so many different sources of regulation in different jurisdictions, let alone voluntary reporting standards. The intent is to harmonize reporting, but it is a slow process. Some companies keep reporting against GRI while being compliant with CSRD, but also report against TCFD and probably the Taskforce on Nature-related Financial Disclosures in the future. Companies also provide responses to CDP, EcoVadis, ratings providers and investor-side ESG analysts. Large companies may further wish to report against the IFRS Sustainability Disclosure Standards if their financial reporting is prepared according to IFRS.
To meet this myriad of reporting requirements, companies should be building the so-called lake of required ESG information and making sure it’s stored properly in the company, that it’s readable, robust data. The introduction of XBRL taxonomy by EFRAG will help with structuring and cross-referencing the data.
The role of AI tools will be to help pull out the right data from the lake to populate different reports or different requests from analysts and stakeholders – all from this bulk of information, while complying with the different formats, requirements and standards. This is likely to give back a lot of time and money to IR and ESG practitioners, so that they can focus on what really matters: driving sustainable progress within their organizations and building relationships with stakeholders.