Scaling Sustainability Impact: How AI Accelerates Environmental Assessments

Solving the unstructured data challenge

The bottleneck in most sustainability software isn’t the software — it’s the data upstream of it. Extraction from supplier PDFs, ERP exports, and invoice archives still happens mostly by hand. In this session, Josine oude Lohuis (CPO, Spirefly) explains why traditional tools have struggled to keep pace with CSRD and Scope 3 timelines, and how a new generation of AI is changing that.

From raw data to impact

Sustainability teams don’t need smarter chatbots. They need systems that understand regulatory context and handle the underlying math correctly. Spirefly’s approach works in three layers.

Data extraction. Large language models now understand the meaning of a sentence, not just its keywords. Whether it’s “200kg of recycled cotton” buried in a purchase order or a complex supplier report, the system pulls exactly what’s needed for an auditable assessment.

Agentic workflows. LLMs are powerful at language but unreliable at calculation. Spirefly gives its AI agents access to specialised tools — Python scripts, Excel formulas, regulatory databases — so footprints stay traceable and scientifically sound.

Human-in-the-loop. The final validation stays with your team. AI handles the repetitive extraction and calculation loops; humans provide the expert judgment that regulators expect.

Beyond static dashboards

Dashboards tell you what happened. They don’t tell you what to do about it. The Spirefly Workbench takes a different approach: you set a goal (say, completing your CSRD submission by a given deadline) and the system builds a plan toward it, surfacing the specific data gaps and the actions most likely to close them.

The session walks through a live example, showing how biodiversity and land-use assessments layer on top of carbon data within the same workflow.

Tips for getting started

A few practical suggestions from Josine for sustainability teams exploring AI tools:

• Give the system context. Tell it your role, your targets, and the regulatory framework you’re working under. Vague prompts produce vague outputs.
• Separate text tasks from calculation tasks. Use AI for extraction and summarisation; use purpose-built tools for the numbers.
• Ask it to check its own work. Most modern systems can flag where their confidence is lower — use that signal.