OpenAI and PNNL launch DraftNEPABench for federal permitting workflows

What was announced

On February 26, 2026, OpenAI announced a partnership with the U.S. Department of Energy’s Pacific Northwest National Laboratory (PNNL) to study whether AI coding agents can help accelerate federal permitting work. The collaboration centers on DraftNEPABench, a benchmark designed around National Environmental Policy Act (NEPA) drafting workflows, including environmental impact statement sections and related technical documentation tasks.

The project was developed with PNNL’s PermitAI initiative and involved domain experts in environmental review. Instead of evaluating abstract prompt performance, the benchmark emphasizes document-heavy workflows where an agent must read large technical files, cross-check references, and produce structured drafts that match legal and policy expectations.

Why this matters

Federal permitting can delay infrastructure projects for years, especially in energy, transportation, manufacturing, and water systems. OpenAI and PNNL frame this work as an attempt to improve the drafting stage without replacing expert judgment. According to the announcement, 19 experts assessed tasks spanning sections used by 18 federal agencies and found that generalized coding agents may save 1 to 5 hours per subsection, representing up to about a 15% reduction in drafting time.

That signal is meaningful because permitting workflows are highly repetitive but still require precision. If drafting support improves while review quality remains high, agencies can redirect human effort toward adjudication, oversight, and edge cases rather than boilerplate composition and reference stitching.

Technical and policy implications

OpenAI highlighted that agent-style interfaces such as Codex CLI can unlock broader reasoning behaviors by letting models work across files and tools, not just in a single text box. In practice, this means AI systems can assemble citations, compare technical sections, and generate revision-ready outputs that humans can audit faster.

The company also noted limitations: DraftNEPABench covers well-specified tasks with available context and does not capture full real-world ambiguity, changing regulations, or incomplete source materials. Some apparent failures were linked to outdated references and rubric quality, which required updates during evaluation.

The next phase is continued support for PermitAI deployments and refinement. OpenAI and PNNL suggest the long-term goal is to move portions of federal review timelines from months to weeks, while keeping experts in control of final decisions.