OpenAI and PNNL say coding agents could cut federal permitting draft time by up to 15%

OpenAI and the U.S. Department of Energy’s Pacific Northwest National Laboratory announced a new partnership on February 26, 2026 focused on one of government’s most document-heavy bottlenecks: federal permitting. The project pairs OpenAI with PNNL’s PermitAI team to test whether coding agents can help accelerate National Environmental Policy Act workflows, especially the drafting of environmental impact statements and related review documents.

The centerpiece is DraftNEPABench, a benchmark designed with 19 subject matter experts around the NEPA review process. OpenAI said the evaluation spans drafting tasks drawn from 18 federal agencies. Across that representative set, the experts found that generalized coding agents could reduce drafting work by 1 to 5 hours per subsection, which OpenAI translates into up to roughly 15% lower drafting time. That does not mean the full permitting process becomes autonomous, but it does suggest a measurable productivity gain in a workflow that can stretch for months or years.

What makes the announcement notable is the tool choice. OpenAI said PNNL explored coding agents, specifically Codex CLI working with reasoning models like GPT-5, as a general interface for research, technical analysis, and structured report writing. Instead of relying only on fixed prompts, the agents were asked to read and synthesize hundreds of pages of technical and regulatory material, verify facts across environmental and engineering sources, and draft reports that meet tightly specified legal and technical requirements. In other words, the benchmark tests whether agent-style systems can manage real file-system workflows, not only chat interactions.

OpenAI also took care to define the limits. The company said DraftNEPABench measures well-specified drafting tasks where the relevant context is available, not the full ambiguity and discretion of real-world permitting decisions. It also noted that some failure cases were tied to outdated references or weak evaluation criteria, and that real deployments would still require expert feedback, iteration, and human oversight. Those caveats matter because permitting decisions carry environmental, legal, and public-interest consequences that cannot be delegated casually.

Even with those limits, the policy significance is clear. Federal agencies review bridges, transmission lines, power plants, manufacturing facilities, and other infrastructure projects under timelines that directly affect investment and competitiveness. OpenAI said it expects better PermitAI tools could eventually move average approval timelines from months to weeks. That claim will need validation in live settings, but the partnership is an important signal that AI deployment is moving beyond office productivity into regulated, infrastructure-facing government workflows.