Recently, our CEO and Co-Founder, Gordon, gave a presentation at the ESG Imperative 2025 conference on “How AI is shaping the future of sustainability analysis.” Since many people seemed interested in it, we wanted to share his thoughts with you.

Keep reading for a quick overview of large language models (LLMs), followed by four use cases that show how AI can be used in sustainability analysis, as well as a discussion on some of the risks to consider when using and implementing AI.

Large Language Models

Since LLMs will be mentioned a lot throughout this article, it would be helpful to have an overview of them. If you’re already clued up in this area, then feel free to skip forward to the case studies.

What are LLMs?

LLMs are advanced AI systems that can understand and generate text similarly to how humans use language. They predict the next word in a sentence based on preceding words, enabling them to perform tasks like content creation, customer support, language translation, and more.

How do they work?

LLMs are trained on vast amounts of text, such as books, articles, and websites. The model is then trained to understand the context of each word in relation to others, enabling more nuanced comprehension of text. LLMs don’t just look at one word at a time; instead, they consider the entire sentence (and sometimes surrounding sentences) to understand the meaning of each word in context.

They then build layers of understanding, starting from simple world patterns and moving up to more complex ideas and relationships. Once the LLM is trained, it can generate responses. When given a prompt, which is what a user provides to an AI model to get a specific output, the LLM uses what it’s learned to suggest the most likely next words or phrases that make sense, creating a coherent and natural-sounding response. This is why LLMs are categorised as Generative AI or GenAI; they generate content. Many different GenAI tools are now available, including ones for generating images, videos, speech, and music.

Use Case 1: Controversy Detection

So, moving on to our first use case, which involves using AI for controversy detection.

It is very challenging for investment teams to monitor controversies at scale for thousands of companies in real-time across many geographies. AI can help us with this problem by analysing real-time news at scale.

LLMs are a massive improvement on traditional Natural Language Processing (NLP) techniques, largely because of their contextual understanding. When an LLM is analysing articles like the ones shown in Figure 2 about Walgreens, it most likely already “knows” what Walgreens does, what the DOJ is and what opioids are.

You can get LLMs to generate controversy scores, which traditional NLP models struggled with. Traditional NLP models were fairly good at calculating the sentiment of an article and picking out keywords, but they always struggled with assessing how bad or good an article was for a company. LLMs are now nearly as good as human analysts in analysing controversies in terms of the quality of output, but their processing speed is worlds apart compared to human analysts - they can analyse thousands of articles and pick out controversies in just a few minutes.

You can set up LLMs to be a lot more explainable than NLP techniques, you can ask them why they highlighted an article as a controversy or why they have given an article a certain sentiment or controversy score. LLMs are multilingual, so they can process news from all over the world - ChatGPT can “understand” over 50 languages. You can fine-tune these models to give even better results. For example, if you train an LLM on a dataset of human rights articles, it will become a specialist in analysing articles with a human rights lens.

Use Case 2: Using a PDF GenAI Chatbot for Company and Fund Analysis

In the financial world, we all spend many hours analysing reports of various kinds, such as company and fund reports, and we generally all rely on using Ctrll+F on the keyboard to search through them. It is probably obvious that Ctrll+F is not a smart search, it is just matching words or part of words, and there is no contextual understanding. Well, LLMs provide a much better solution than this since they are very good at reading and comprehending unstructured data like text from reports.

Now, it’s possible to build PDF GenAI chatbots to help us analyse reports at scale. We can use a Retrieval Augmented Generation or RAG approach to ensure that we get high-quality answers from the chatbot.

What is RAG?

Simply speaking, RAG combines the retrieval of external data sources with the generative ability of an LLM. The retrieval step ensures that the model has access to the most up-to-date or contextually relevant information before it generates its responses. We can give the LLM text from articles as we discussed in Use Case 1, or in this use case, text from PDFs in order to decrease the likelihood of hallucinations or irrelevant content being generated.

It is likely that many of you have played around with some form of LLM and have seen that if you ask an open-ended question, it can sometimes hallucinate or include filler words that don’t add much informational value. But if you are very explicit and detailed in your prompt and ask it only to use the given text, you will get a much higher quality answer.

Here are some examples of analytical tasks that you can get chatbots to assist you with:

1. You can ask the chatbot to carry out data extraction tasks and get it to pull out and summarise any numbers that you want (see Figure 3), like carbon emissions from a company report and output those numbers to Excel.

2. You can compare a company’s historical reports simultaneously and see how a company has changed over time.

3. You can compare multiple companies simultaneously, e.g. a certain peer group like UK banks.

4. You can customise it with your firm’s sustainability policy or responsible investing approach so the chatbot always refers to it when answering questions. We can also tell it to act as an ESG analyst and set up guardrails so that it does not answer anything non-ESG or sustainability-related.

5. You can bring in other data sources to complement the PDFs, such as the latest news for a company. This would require giving the chatbot the ability to pull in additional data sources.

These are just some examples, but hopefully, they illustrate how helpful GenAI chatbots can be for PDF analysis. The potential of this use case is hugely exciting.

Use Case 3.1: Company Greenwashing Analysis

Moving onto the third use case, which is about using AI to aid us with greenwashing analysis.

Over the past few years, prominent corporates and fund managers have been accused of greenwashing, which has resulted in reputational damage and financial penalties. Thus, it’s a very important topic in the sustainable finance world.

Assessing greenwashing risk at scale is challenging because we need to analyse a lot of data, but AI can help. Greenwashing is multi-faceted, but it boils down to assessing what a company says it’s doing and what it’s actually doing.

We will use P&G to illustrate a simple greenwashing analysis workflow. Firstly, we can use LLMs to extract P&G’s targets, such as SBTi (see Figure 4) and Net Zero Tracker targets, from company reports, websites or the news.

We can then forecast P&G’s carbon emissions to determine whether it is likely to meet its targets - for this, we can use a simple predictive model. You can see from the below image that P&G’s carbon scope 1+2 emissions are not on track to meet its 2030 target.

We can use LLMs to generate a greenwashing controversy score from the news, and we can see in Figure 6 that GaiaLens’ controversy detection system has flagged up an article describing how P&G has been accused of using misleading environmental claims and logos, indicating a high greenwashing controversy risk.

To conclude from this analysis, it does seem that P&G has a relatively high greenwashing risk.

Use Case 3.2: Fund Greenwashing Analysis

We can follow a similar workflow for funds. We can look at what the fund says it’s doing, i.e., its mandate, and what it’s actually doing, i.e., how it is invested. If a fund is sustainability-focused or has a sustainable label, we need to assess whether its investments align with its sustainability goals.

Firstly, we need to establish what the fund’s sustainability goals are, if there are any. If you are analysing many funds, this would take a lot of time, but we can use LLMs to extract this information from the fund prospectus in an automated, scalable way.

Then, we can use this information to analyse the fund dynamically. For example, if the fund is climate-focused, we want to analyse its climate performance, and if it’s more social pillar-focused, we want to analyse it with respect to social factors.

In this example, we are looking at a Climate Action Equity fund. We can see in the below image that the fund has a higher carbon footprint than the market benchmark, which is the MSCI ACWI, the fund’s comparator benchmark. This is surprising given that the fund is a climate action fund.

Analysing controversies again gives us up-to-date information. The next image shows that the GaiaLens controversy detection system has flagged National Grid (the fund’s sixth-biggest holding at the time of writing). The company has been accused of not living up to its environmental promises as it is not effectively using battery storage, indicating a greenwashing controversy risk.

To conclude, this climate action fund has a high greenwashing risk, according to this analysis.

Use Case 4: ESG AI Agents of the Future

So far, we have seen how good LLMs are at analysing text and improving our sustainability analysis workflows. One of the most exciting developments of LLMs is using them as agents to carry out some action. For example, you can give LLMs additional capabilities, such as the ability to search Google, manage files on your computer, or even send messages on your behalf to Slack or Microsoft Teams.

You may have seen recently that OpenAI just launched a new product called Operator, which is a browser agent that can book plane tickets for you, book an Uber, complete online grocery orders, and make dinner reservations for you - all from a single prompt.

I thought that it would be fun to think about how we could use agents to turn our sustainability analysis into actions. From an Asset Owner perspective, we could use AI agents to draft an email to a fund manager based on a significant controversy being detected or the greenwashing risk of a fund breaching a threshold.

We could give the AI agent even more context to draft the email, such as previous email chains with this fund manager, whether this has happened before, and whether the fund manager has already taken action. The AI agent could also automatically download a report summarising the fund greenwashing risk assessment and attach it to the email. Of course, the AI agent could also send the email and automate the whole workflow.

This type of automation carried out by AI agents may seem quite simple, but they could save people a lot of time the smarter they get and the more capabilities and access we give them. Hopefully, this example gives you a bit of an idea of how AI agents could be used in the future.

AI Risks to Consider

However, there are some risks to consider when using and implementing AI.

The first AI risk to consider is bias: AI systems can unintentionally perpetuate or amplify biases present in the training data. There are also biases created from the process of using human feedback to train and improve the LLMs, which is called Reinforcement Learning from Human Feedback (RLHF). You may have read that OpenAI’s ChatGPT and Google’s Gemini have been accused of having a left-leaning political bias. Alternatively, the Chinese DeepSeek model has implemented censorship measures for politically sensitive topics.

LLMs can hallucinate. In AI, hallucination refers to when LLMs generate false, misleading, or nonsensical outputs that are not grounded in real data. For example, LLMs can make up numbers or generate links that don’t exist, which you may have experienced. This happens because of the fundamental way LLMs work: they predict text based on patterns rather than verifying factual accuracy.

LLMs have caused IP and copyright Issues since the AI models are often trained on copyrighted materials, which can lead to IP disputes. Some of you may have seen that the New York Times has taken legal action against OpenAI, alleging that their AI tools have infringed upon the newspaper’s copyrighted content. OpenAI itself does not guarantee that generated content is unique or free from third-party claims. Another IP risk is that some AI tools may have claims to the IP that is created through the use of the tool, so it’s important to check their terms of use.

There may be privacy violations: some AI tools may inadvertently collect, store, or process sensitive personal data, which may breach privacy laws. The best tools allow you to choose whether you want the model to learn from your data or not, but sometimes it’s not always clear.

There may be cybersecurity vulnerabilities: AI systems can be exploited through adversarial attacks, and AI agents can be hacked, manipulated, or misused. With some chatbots that aren’t properly protected, you can quite easily override the system settings yourself just by using a prompt.

Finally, training and running large AI models can have a significant carbon footprint, which can affect sustainability goals. According to estimates by Goldman Sachs, a ChatGPT query requires nearly ten times the energy of a Google search.

Of course, this is not an exhaustive list, but it includes most of the main risks that are important to consider.

Please don’t hesitate to get in touch if you want to learn more about how GaiaLens has implemented AI to help clients with sustainability analysis. We have solutions for all of these use cases.