Xilva is on a mission to scale more forests than we've seen in our lifetime - and we're using AI. Here's how it’s enhancing due diligence, its current limitations, and my perspective on what lies ahead.

Author: Yeray Martinez

Realising the potential of Nature-based Solutions (NbS) faces several challenges: our collective undervaluing of natural systems, getting funding to the right projects, lack of trust on the demand side, and a developing, but insufficient pipeline of investable projects.

At Xilva, we work to overcome these challenges by using a data-driven due diligence approach rooted in our holistic framework - XILVA GRADE. The framework enables investors to ground decisions and trust in the right understanding, while also providing meaningful feedback to the projects assessed so they can improve and realise their potential.

While due diligence is critical for project success, for the NbS sector to succeed, we need to enable more forest projects, faster.

For this, we need to be open to new ways of working, and where relevant, embrace tools that enhance the efficiency and scalability of due diligence.

This is where we, like many others, recognise that artificial intelligence (AI) tools have significant potential and should be leveraged.

Our Approach to Working with AI

Transforming landscapes to be regenerative and to deliver positive impacts - such as carbon sequestration, biodiversity and social outcomes - requires innovation. We’ve found that an agile, Minimum Viable Product (MVP) mindset is critical for this; it's proven highly effective in developing dynamic and lively solutions.

This is particularly important in the context of forests, which are complex and ever-changing socio-ecological systems. To adapt to change, agile management techniques are essential.

So, when we consider applying AI in this context, we focus on the required processes that enable investment and project success. Due diligence and risk understanding are the pivotal steps for connecting funding with projects. These processes must be undertaken with the best available knowledge and data.

Therefore, our current focus of AI use is to enhance and speed up the risk understanding process, specifically through the application of Large Language Models (LLMs). And to do this we have had the invaluable collaboration of a team of researchers from the University of Zurich [See Footnote (1)].

What We’re Using AI For

Due diligence is a labour-intensive, time-consuming and costly process. It requires analysing and verifying massive amounts of information from multiple sources.

The most important thing is to identify where technology can add value to the process. At present, we’re focused on three things:

Speeding up due diligence to serve more investment processes, driving the sector's growth and enabling larger-scale impact.
Improving the quality and reliability of due diligence analysis and recommendations, boosting investor confidence in investment decisions as well as the sector overall.
Enhancing project developers' capacity to meet their goals through more insightful feedback and support.

When the goal is to provide evidence to aid an investor’s decision, this can be reinforced by the analysis of more data. Large language models (LLMs) offer tremendous opportunities to accelerate our data processing.

Benefits include:

Streamlining workflows
Processing more data for more nuanced insights
Increasing efficiency in identifying anomalies and risks
Benchmarking project results and collating key information from other projects

How We’re Using AI

Our current focus is on applying LLMs to enhance data analysis.

This said, we’re acutely aware of the limitations of these technologies. While LLMs seem to understand what they’re saying, they do not. So, we need to make sure that the analysis workflow is the right one to generate reliable results.

I recently heard a big data expert say that, “…it’s not so much about how much data you have, but what you want to learn from your data.”

That’s why, asking the right questions is essential for the right answers, and this is where our XILVA GRADE framework comes in. It’s really helpful because it’s designed to tackle the complexity of forest projects, distil key insights, and provide investors with a clear understanding of a project’s strengths and weaknesses. On top of that, it highlights relevant risks at the current stage of a project and offers guidance on how to manage them.

Essentially, XILVA GRADE is about assessing whether the potential value of a project is worth pursuing, and how to make sure that the value is, or can be unlocked. With this insight, investors and capital providers are equipped to make the right decisions.

In the context of the above, we’re developing our AI tool, combining the right prompting with the right analysis workflow.

We’re basing the workflow on the Retrieval-Augmented Generation (RAG) systems approach, a really powerful approach for increasing the reliability of results and minimising hallucinations.

However, we’ve realised that for the model to generate useful insights, it needs to understand the context of the information it retrieves from documents, as well as the criteria by which these insights should be evaluated.

This becomes challenging because a finding’s significance can differ depending on its source, for example, whether it comes from a project developer’s document, an external audit, or a research paper. Additionally, it’s important to assess consistencies and discrepancies across varying sources.

Setting up the right workflow for the model to deliver accurate results is key. We’ve made good progress with this, but it’s still a work in progress.

Where We’re Using It

We’re using LLMs in the first stage of our analysis to get an overview of available documentation and identify where relevant findings are located for each area of our analysis (based on XILVA GRADE).

We’re testing to form judgements on the findings, but only as a preliminary result and always subject to review. This process has also proved valuable in helping us improve our workflow prototypes and prompting strategies.

Then, we’re also using AI to challenge our final findings; to check that they’re consistent with all the data gathered and to assess if we’ve missed anything relevant.

Where We’re Not Using It

We’re not using AI to produce unsupervised results. While we recognise its potential for specific use cases like checking project eligibility criteria, this is still something under development.

Risks and Limitations

The reliability of output remains a primary concern. However, we are continuously refining our workflow, recognising that optimised input instructions are essential to achieving the right results.

We’re also exploring ways to address the interconnectedness of topics. While we anticipate these interconnections, our team is often surprised by how profound they are on the ground.

Working with these relationships requires nuanced judgement - you can’t narrow down or separate too much.

Currently, AI has limitations in handling these subtleties, contradictions and knowledge of sometimes very specific local contexts. These are difficult to resolve and require careful, manual assessment.

Beyond Technology: The Role of Trust & Relationships

There’s a growing need for more transparent and verifiable data to provide stronger evidence of the impacts of NbS.

AI can help with this. It doesn’t replace human insights, but it can support or challenge it.

However, for this, the right approach will always need the right questions, and often, additional lines of inquiry arise from discussions with stakeholders and personal interactions in the field.

It would also be simplistic to assume that investment decisions are only ever based on data.

A project’s narrative and trust between parties is an important aspect of investing in NbS, and trust isn’t a process - it’s developed through relationships over time.

We have clear evidence from both sides of these transactions that the human touch is essential. This is especially true for the long-term nature of forest projects, where projects can span decades.

Further, the mental shift required to reframe how we perceive and understand the value of natural systems requires more than data. It calls for deeper connections with nature; connections that are shaped through relationships and collective knowledge-sharing.

To the Future

AI will continue to accelerate workflows and strengthen due diligence, but it can't bridge trust gaps on its own, substitute for human judgement or inspire public support necessary to steer investor sentiment.

While we’re in the early stages of the technology, I see its most powerful role as a tool to speed up due diligence and provide richer insights that support human decision-making.

Footnote:

(1) Thanks to Liudmila Zavolokina, Kilian Sprenkamp, Jiayu Luo and Maria Letizia Jannibelli from the University of Zurich for their collaboration and support. More details of this collaboration can be found here:
Jannibelli, Maria & Luo, Jiayu & Sprenkamp, Kilian & Zavolokina, Liudmila. (2025). Using Large Language Models for the Assessment of Sustainable Forest Investment Projects. 10.24251/HICSS.2025.562.

How AI Improves Forest Project Due Diligence but Can’t Replace Judgment