Pacific Northwest National Laboratory and OpenAI partner to accelerate federal permitting

Source: OpenAI Blog

Modernizing how the federal government permits critical infrastructure is essential to building a faster, safer, and more competitive U.S. economy. From energy projects and advanced manufacturing to transportation and water systems, permitting determines how quickly promising ideas become real‑world investments. Yet today, environmental and technical reviews often take years, slowing innovation, increasing costs, and delaying benefits to communities.

OpenAI and PNNL Collaboration

OpenAI has partnered with the U.S. Department of Energy’s Pacific Northwest National Laboratory (PNNL) and its PermitAI™ team to evaluate whether coding agents can help accelerate federal permitting work. PermitAI, an initiative funded by the DOE Office of Policy, worked with 19 subject‑matter experts on the National Environmental Policy Act (NEPA) review process to design a benchmark—DraftNEPABench—for assessing AI performance on tasks related to NEPA workflows such as drafting Environmental Impact Statements (EIS).

PermitAI Benchmark (DraftNEPABench)

The benchmark covers a representative set of drafting tasks spanning NEPA document sections from 18 federal agencies. Experts evaluated AI‑generated drafts on structure, clarity, accuracy, and references using a 1‑5 scale (1 = major deficiencies, 3 = partially correct, 5 = fully correct). Mean scores across 102 tasks were aggregated by lead agency.

Mean evaluation scores (1‑5 scale) across 102 tasks, grouped by lead agency. Scores aggregate assessments of structure, clarity, accuracy, and references.

Key Findings

• Generalized coding agents (e.g., Codex CLI) can reduce drafting time by 1 to 5 hours per subsection, roughly a 15 % reduction in overall drafting effort.
• The agents demonstrated the ability to:
  • Read and synthesize documents spanning hundreds of pages of technical and regulatory content.
  • Verify facts across multiple environmental, engineering, and regulatory sources.
  • Draft structured reports that meet highly specified legal and technical criteria.

These results suggest a meaningful step forward for AI support of complex government workflows.

Capabilities of Coding Agents

By giving models access to a command‑line interface (typically used for coding tasks), they can employ more general problem‑solving strategies than hand‑crafted heuristics. This approach enables agents to:

• Navigate file systems to retrieve and organize relevant data.
• Execute reasoning pipelines that combine natural‑language understanding with programmatic data manipulation.
• Generate dynamic outputs, such as web‑based reports and interactive visualizations, that go beyond static PDFs.

Implications for Government Workflows

• Agencies could review, refine, and approve proposals more efficiently, allowing human reviewers to focus on judgment, oversight, and complex decision‑making.
• AI‑augmented teams can handle time‑consuming portions of permitting work, accelerating project development and strengthening U.S. competitiveness.
• The benchmark highlights where current models can responsibly assist humans, while also clarifying their limitations.

Limitations and Future Directions

• The benchmark evaluates well‑specified drafting tasks with available context; it does not capture the full ambiguity and discretion of real‑world permitting decisions.
• Errors sometimes stemmed from outdated references or weak evaluation criteria, prompting rubric updates.
• Incomplete, inconsistent, or out‑of‑date source materials may go unnoticed without explicit instructions.
• Real‑world deployments are expected to incorporate expert feedback and iterative refinement, likely improving performance beyond the benchmark results.

Looking Ahead

OpenAI is supporting PNNL to further develop and refine solutions for PermitAI™. Over time, the average time to approval for federally reviewed infrastructure projects could fall from months to weeks, accelerating development, supporting long‑term economic growth, and helping the United States thrive in the Intelligence Age (Sam Altman’s essay).