Environmental planning, from weeks to minutes

Environmental planners at agencies like the US Forest Service spend weeks per project. They hunt for similar past work, draft categorical exclusions and environmental assessments, and coordinate timelines across teams and contractors. Expert work stuck in slow, manual tooling.

An AI that drafts regulatory documents has to earn its trust. If it invents a precedent or a citation, the result is worse than nothing, because a human still has to catch it. So the interface had to show what the agent was doing and where it got the answer from, not just hand back a document.

The core flow runs in three phases. The chat interviews the planner. A research agent works in the background, pulling real precedent projects and the right Categorical Exclusions. Only then does the workspace unlock document generation. The planner always sees what the agent found and why: chat on the left, project fields filling in on the right as it goes.

Around that core I designed the rest of the product. A project workspace on a keyboard-friendly Gantt chart with tasks and milestones. Task sheets where the agent drafts scoping letters, environmental assessments and decision memos, each with citations attached. A GIS map that renders uploaded shapefiles and tracks work unit by unit. A public portal for community comments. And a mobile app for fieldwork. All of it sits on a forest-green design system, organised the way the agencies are: organization, office, project.

In a regulated field, the interface has to keep the AI's work auditable. Every generated document traces back to a real source. The agent's progress stays visible. Nothing gets submitted on its own; the planner decides. That constraint pulled most of the design decisions in the same direction.