AI in Permitting: OpenAI and PNNL Test AI Agents on Real NEPA Work
March 24, 2026
OpenAI and Pacific Northwest National Laboratory released DraftNEPABench on February 26 — a benchmark that tests how well AI coding agents perform on actual NEPA document drafting tasks. The results are useful, but the details matter more than the headline number.
What they tested
The benchmark includes 102 drafting tasks pulled from 18 federal agencies. Nineteen NEPA experts evaluated GPT-5 coding agents working through command-line interfaces to interact with technical files directly — not chatbot-style Q&A, but agents reading source material and producing draft language.
The finding: AI agents reduced drafting time by roughly 1-5 hours per subsection, about a 15% speedup. That's not a revolution. It's a useful productivity gain on the writing portion of a process where writing isn't the bottleneck.
What it doesn't tell you
NEPA review is slow for structural reasons, not because people type slowly. The time sink is in coordination between agencies, waiting for data from applicants, reconciling conflicting regulatory requirements, and managing the institutional risk of signing your name to a decision. A tool that makes the drafting 15% faster doesn't touch any of that.
The PNNL team was careful about this distinction. Expert oversight, validation, and sign-off remain required. The benchmark measures drafting support, not decision-making. That's an important line, and it's good to see it drawn clearly in the research itself rather than left for journalists to misinterpret.
The bigger pattern
DraftNEPABench dropped the same month that Denver approved a $4.6M AI permit review contract with ComplyAI and Deloitte's GovTech Trends report flagged agentic AI in government as a 2026 theme. Three data points in a month is a trend.
But look at what's actually happening. Denver's contract is for building permits — plan review, code compliance, intake automation. ComplyAI's CivCheck reads construction documents and checks them against building codes. That's a bounded, rules-based task with clear right-and-wrong answers.
NEPA is different. Environmental review involves judgment calls about significance, mitigation adequacy, cumulative impacts, and public concern — categories where "correct" depends on context, precedent, and who's asking. AI can help organize information and draft language. It can't make the call on whether a project's impacts are significant.
The companies and agencies that understand this distinction will build tools that last. The ones that don't will spend a lot of money on pilots that get walked back when the first controversial decision gets challenged in court.
What to watch
SearchNEPA, a related PNNL tool, is in beta with 200+ federal evaluators. Two more products (EngageNEPA for public comment analysis and CommentNEPA for comment generation) are in development. The question isn't whether AI will enter environmental permitting — it's entering right now. The question is which tasks it gets pointed at, and whether the people deploying it understand what it can and can't do.
OpenAI Announcement · TechInformed · DraftNEPABench Preprint (PDF)
AI in Permitting is a recurring column on Permitting Tech covering how artificial intelligence is entering permitting workflows. Written by Boon Sheridan.
