Question 1

Encoded sustainability intelligence

Accepted Answer

AI is only as good as the underlying data and the expert review that trains the system. Watershed’s science team has decades of experience on peer-reviewed data projects – including foundational work in the LCA field and open-data initiatives. We have published public evaluation data-sets for AI systems. Questions to evaluate Sustainability AI:

What is the underlying data and how is it updated over time?

What evaluation data sets are used? Are they peer-reviewed?

How are sustainability experts involved in system development and validation?

When using AI for Scope 3, how do you ensure consistency across materials and industries?

Question 2

Safeguards for greater accuracy

Accepted Answer

The most reliable AI systems for sustainability are built with multiple agents to enforce accuracy. For example, a mapping agent and then a separate judge agent to prevent confirmation bias when evaluating the AI mappings. Without built-in and automated checks and balances, incorrect AI mappings can produce plausible yet wrong emissions values, causing silent failures. Questions to ask other solutions:

Does the technology take a multi-agent approach?

What are your hallucination safeguards?

Can they share regular accuracy testing scores?

Question 3

Transparency and explainability

Accepted Answer

Many AI models are black boxes — they provide a plausible-sounding result, but little information as to the source of that result. Without that transparency, it’s hard to trust AI and easy to miss mistakes. Watershed solves this by tracking the data points and assumptions the AI uses and exposing them so users can easily review and understand the AI's approach. Questions to evaluate Sustainability AI.

How can I understand how the AI achieved its result?

Does your platform show detailed steps explaining the models assumptions, data points, and calculations?

How can I efficiently review the most important assumptions the AI makes?

Question 4

Human control and refinement

Accepted Answer

Central to how Watershed builds sustainability AI is ensuring that it can be acted upon and improved by sustainability leaders. This includes the ability to efficiently review inputs, assumptions, calculations, and outputs, refine them with primary data to make the results more accurate and actionable, and log it all so it can be tracked and reviewed. Questions to evaluate Sustainability AI:

Can you easily review AI inputs, assumptions, and outputs?

How does the AI system show you the most important data points that need your review and refinement?

Can you easily override and refine AI inputs, assumptions, and outputs to create a better result?

How does the system record the AI and user actions taken in creating the final result?

Watershed’s approach to AI

Sustainability AI principles

Sample technical requirements for sustainability AI

Watershed research

Criteria for credible AI-assisted carbon footprinting systems: The cases of mapping and lifecycle modeling

ATLAS: A spend classification benchmark for estimating scope 3 carbon emissions

An LLM-based approach to creating a carbon emissions reduction levers library at scale

AI product advisory groups

Decompose your supply chain for smarter procurement decisions

Simulate the impact of potential decarbonization levers

Accelerate reporting