OpenAI and PNNL Partner to Accelerate Federal Permitting
OpenAI and the U.S. Department of Energy’s Pacific Northwest National Laboratory (PNNL) have partnered to evaluate how coding agents can accelerate federal permitting for critical infrastructure. This collaboration aims to modernize the review process for energy, transportation, and water systems to reduce the years-long delays typically associated with environmental and technical reviews.
Potential Efficiency Gains in NEPA Drafting
Generalized coding agents can reduce the drafting time for National Environmental Policy Act (NEPA) document subsections by 1 to 5 hours, representing up to a 15% reduction in overall drafting time. This finding is based on evaluations conducted by 19 subject matter experts across a representative set of drafting tasks spanning NEPA document sections from 18 different federal agencies.
The DraftNEPABench Benchmark
To assess AI performance in government workflows, OpenAI and PNNL developed DraftNEPABench, a benchmark designed to evaluate how AI models perform on tasks such as drafting environmental impact statements.
Technical Implementation
The partnership explored the use of generalized coding agents (specifically Codex CLI) to extract performance from reasoning models like GPT-5. By providing models with access to a command-line interface, the agents can employ general problem-solving strategies rather than relying on hand-crafted heuristics. These agents are tasked with:
- Synthesizing technical and regulatory content across documents spanning hundreds of pages.
- Verifying facts across multiple engineering, environmental, and regulatory sources.
- Drafting structured reports that adhere to specific legal and technical criteria.
Implications for Federal Infrastructure
Accelerating the permitting process is viewed as essential for U.S. economic competitiveness and the development of safer, more responsible infrastructure. The goal is to reduce the average time to approval for federally reviewed infrastructure projects from months to weeks.
Beyond drafting, the use of coding agents allows for the transition from static PDFs to dynamically generated web-based reports and interactive visualizations, which simplifies the validation process for human reviewers. This shift allows government workers to focus on high-level judgment, oversight, and complex decision-making while AI agents handle time-consuming drafting tasks.
Limitations and Scope
The benchmark evaluates model capabilities on well-specified drafting tasks with available context, rather than the full ambiguity and discretion required for real-world permitting decisions. Key limitations include:
- Reference Accuracy: Some identified errors were attributed to outdated references or weak evaluation criteria rather than model failure.
- Source Material Dependency: Models may not flag incomplete, inconsistent, or outdated source materials unless explicitly instructed to do so.
- Human Iteration: Real-world performance is expected to be higher than these self-contained benchmark tasks due to the expected integration of expert feedback and iterative refinement.
Future Development
OpenAI continues to support PNNL in refining solutions for PermitAI’s applications, which are designed to help federal agencies streamline the permitting processes for critical infrastructure.