Summary

AI in Water Risk: The Model Provides Capability. The Framework Provides Auditability.

Why non-specialized models are a liability for capital decisions in corporate sustainability.

There's a question we get asked a lot from technical teams: "How do you overcome hallucinations from models?"

The answer isn't just a better model. It's what you build around one.

Even top-tier models show measurable hallucination rates. For brainstorming? Fine. For site selection decisions involving billions in capital? Not acceptable. Decision-makers need traceability, auditability, reproducibility - they need to know where every claim comes from and why it's credible.

At Waterplan, we built a production-grade water risk intelligence system that synthesizes five distinct evidence sources - from satellite data and government pipelines to AI-powered research in local languages - into auditable, locally-grounded risk scores across six dimensions, with global coverage.

The full post walks through a real example showing exactly how it works: location resolution that maps coordinates to actual hydrological features (not just a state name on a map), multi-language search queries that surface Dutch PDFs English-only searches miss, multi-stage filtering that rejects irrelevant sources before synthesis, and a concrete scored example from a real site.

The model provides capability. The framework provides auditability.

Want to read the full technical breakdown?