Related ToolsChatgptClaude

OpenAI and PNNL Build Benchmark for AI-Assisted Federal Permitting Drafts

OpenAI and PNNL Build Benchmark for AI-Assisted Federal Permitting Drafts
Image: OpenAI Blog

What Happened

On February 26, 2026, OpenAI and Pacific Northwest National Laboratory (PNNL) announced a new benchmark called DraftNEPABench. The benchmark evaluates how well AI coding agents can assist in drafting documents for federal environmental review under the National Environmental Policy Act (NEPA).

Results show that AI assistance can reduce NEPA drafting time by up to 15%. The goal is to modernize a permitting process that currently takes years and represents a significant bottleneck for infrastructure development.

Why It Matters

Federal permitting is a genuine problem. NEPA reviews for major infrastructure projects - transmission lines, roads, energy facilities - can take 5 to 10 years. This is a well-documented constraint on U.S. infrastructure development. A 15% reduction in drafting time is meaningful, though it addresses only part of the overall permitting timeline.

The introduction of DraftNEPABench as a public benchmark is also significant. Benchmarks create accountability and enable comparison across different AI systems. If PNNL and OpenAI publish this openly, other labs can test their models against the same tasks, which drives improvement across the field.

For AI practitioners in government or consulting contexts, this represents a concrete, domain-specific application of AI agents that goes beyond generic productivity claims.

Our Take

A 15% drafting time reduction is real but modest. The bigger opportunity is in the review, feedback, and iteration cycles that follow initial drafts. If AI can also assist with responding to public comments, coordinating across agencies, and maintaining consistency across document versions, the cumulative time savings could be substantially larger.

This is the kind of unglamorous, domain-specific AI application that tends to generate actual value - applying models to tedious structured tasks rather than open-ended generation.