How the AI works
Grounded drafting, not generation.
Verdant uses language models for what they are genuinely good at: reading messy documents and mapping them to a framework. Every number it drafts is tied to a source line and a stated factor.
Four steps, each auditable.
Read the unstructured data
A language model reads the raw export the way a person would: a utility bill, a fuel-card CSV, a supplier ledger. It pulls out the activity data (the therms, gallons, MWh, kilograms, miles, and dollars) and keeps the exact sentence each one came from.
Classify into a framework
Each activity is classified under the GHG Protocol: Scope 1 for owned combustion and fugitive gases, Scope 2 for purchased energy, Scope 3 for travel and purchased goods. The model matches the right emission factor for the activity and region.
Compute, and show the work
Verdant converts activity data to tonnes of CO2e with the factor it chose, and records the arithmetic. The factor and the calculation are stored next to the figure, not hidden behind it, so a reviewer can check any number in seconds.
Ground every figure
Each figure links back to the verbatim source line it was read from. If an activity is not in the data, Verdant does not estimate it silently; it lists the gap. The model drafts the report; it does not invent the inputs.
How we keep it honest.
No number without a source
- Every figure carries the verbatim line it was read from.
- The emission factor and the arithmetic are shown, not assumed.
- Confidence is labelled: high for activity-based factors, lower for spend-based estimates.
- Missing data is named as a gap, never filled with a guess.
Model-agnostic by design
Requests go through a thin layer that targets an OpenAI-compatible endpoint. The model and the provider are configuration, so we can route to the best available model, or run on the model your security team has approved, without changing your data or your report.
VERDANT_MODEL · VERDANT_API_BASE · swap in two variables
What the model sees, and does not keep.
Your operational data is sent to the configured model provider only to draft your report, under a data processing agreement, and is not used to train their models or ours. You can redact identifiers before drafting. The full posture, including subprocessors and encryption, is on the Security page.