OpenAI and Pacific Northwest National Laboratory (PNNL), a US Department of Energy research facility, are testing whether artificial intelligence coding agents can accelerate the federal permitting process for infrastructure projects.
The partners developed DraftNEPABench, a benchmark designed with 19 subject matter experts to evaluate how well AI models perform on drafting tasks under the National Environmental Policy Act (NEPA), the legislation governing environmental review of major federal projects.
The benchmark draws on drafting tasks from 18 federal agencies and scores outputs on a one-to-five scale across 102 tasks, measuring structure, clarity, accuracy and references.
Experts found that generalised coding agents saved between one and five hours per subsection, amounting to roughly a 15% reduction in overall drafting time.
The tested workflow used Codex CLI, OpenAI's command-line coding tool, alongside reasoning models including GPT-5, which read lengthy technical files, verified facts across engineering and regulatory sources, and drafted reports to legal and technical specifications.
OpenAI said some errors identified during evaluation stemmed from outdated references and insufficiently precise scoring rubrics, both of which were revised during the process.
Related reading
- OpenAI appoints Arvind KC as chief people officer to lead hiring and workforce strategy
- OpenAI signs four of the world's biggest consultancies to sell its AI coworker platform into enterprises
- OpenAI: the state of enterprise AI
OpenAI said it is supporting PNNL in developing PermitAI, a suite of applications intended to help federal agencies process approvals more efficiently.
The partners said the work could reduce average approval times for federally reviewed infrastructure projects from months to weeks.
The recap
- OpenAI and Pacific Northwest National Laboratory launched a permitting benchmark.
- Experts assessed 102 tasks across 18 federal agencies.
- Aim to reduce approval times from months to weeks.