OpenAI wants coding agents to speed up federal paperwork

Building a bridge, power line, or semiconductor plant in the U.S. rarely stalls on construction. It stalls on paperwork. For decades, federal environmental reviews have taken years to clear, burying projects under thousands of pages of reports.

Now, OpenAI wants to use AI to slash that bureaucratic backlog. The company just teamed up with the U.S. Department of Energy's Pacific Northwest National Laboratory (PNNL) and its PermitAI initiative to launch DraftNEPABench, a new benchmark designed to test how well AI coding agents can handle federal environmental permitting documents.

Giving code agents government paperwork

The National Environmental Policy Act (NEPA) requires agencies to evaluate the environmental footprint of major infrastructure before shovels hit the ground. Reviewers must digest hundreds of pages of engineering specs, cross-reference multiple legal codes, and draft dense, legally vulnerable analyses.

To see if machines could help, PNNL and OpenAI enlisted 19 domain experts across 18 federal agencies to test 102 specific drafting tasks. Instead of standard chat models, they deployed GPT-5 reasoning models paired with Codex CLI, a coding agent that interacts directly with file systems.

It turns out a tool built to inspect code repositories is remarkably well-suited for federal regulations. The agent was able to parse long technical documents, cross-check facts across environmental sources, and generate structured drafts. It could even spin up interactive web dashboards rather than flat PDFs, making verification simpler for human reviewers.

The benchmark found that the agent cut drafting time by 1 to 5 hours per subsection, amounting to roughly a 15% speedup. Over time, OpenAI says it hopes this work helps shrink federal approval timelines from months down to weeks.

The limits of automated red tape

Drafting text, however, is not the same as granting approval.

OpenAI explicitly noted that DraftNEPABench evaluates well-scoped tasks where all reference materials are neatly provided. It does not measure how an agent handles the messy, ambiguous judgment calls that human regulators make every day.

There is also the garbage-in, garbage-out risk. When source documents were incomplete or outdated, the models rarely flagged discrepancies unless explicitly instructed to do so. In fact, several initial failure cases in the benchmark turned out to be caused by outdated agency rubrics rather than model hallucinations.

In real-world offices, experts will still need to review and iterate heavily on every paragraph. The agent remains an eager research assistant, not an autonomous bureaucrat.

Why it matters

For builders and enterprise developers, this project offers a clear lesson in agent architecture. We think applying coding agents like Codex CLI to non-coding domains is a brilliant move. Any workflow that requires navigating local files, cross-referencing multiple disparate sources, and enforcing strict structural rules can benefit from a command-line agent far more than a basic chat window.

For the tech industry at large, the strategic play here is impossible to miss. Advanced AI models require massive amounts of energy, which in turn demands fast deployment of transmission lines, clean energy projects, and data centers. Environmental permitting has long been the primary bottleneck strangling U.S. infrastructure. By partnering with the Department of Energy through its OpenAI for Government initiative, OpenAI is quietly building the digital machinery needed to speed up the physical world its models depend on.

Still, we see serious friction ahead. A 15% reduction in drafting time is useful, but environmental reviews face relentless legal scrutiny from third parties. If an agent overlooks an outdated habitat survey, the ensuing lawsuit could delay a project for years. The true bottleneck in Washington has never been typing speed, but political liability and fear of litigation. Until federal staff feel legally protected signing off on AI-assisted drafts, the timeline gains may remain modest.