DOE AI Tool Cuts Nuclear Licensing Document Drafting From Weeks to One Day

A 208-page nuclear licensing document that typically consumes four to six weeks of expert labor took Everstar's Gordian AI exactly one day to produce. That benchmark, demonstrated by a team spanning the Department of Energy, Idaho National Laboratory, Argonne National Laboratory, and Microsoft, is the most concrete speed signal the nuclear licensing world has seen from an AI tool, and it raises a direct question for anyone tracking advanced reactor deployment: does this change licensing timelines, or just drafting timelines?

The team used Gordian, a domain-specific AI platform built on Microsoft Azure and developed by New York City-based Everstar, to convert a Preliminary Documented Safety Analysis for the National Reactor Innovation Center's generic high-temperature gas reactor into sections formatted to NRC license application standards. The methodology was documented in INL technical report INL/RPT-25-87369, published March 18, and announced publicly by DOE shortly after. The work was conducted under DOE's Genesis Mission, a broader initiative to bring AI-enabled tools into the nuclear energy value chain.

Here is what the demonstration actually established, and what it did not.

Gordian converted source material from DOE's authorization pathway into NRC-formatted licensing sections while simultaneously flagging gaps where information was missing or incomplete. That second function matters nearly as much as the speed gain. Incomplete submissions are one of the most consistent sources of NRC review delays, and a tool that surfaces those holes before filing could prevent rounds of deficiency letters that stretch timelines by months. DOE also confirmed that human experts remained in the validation loop throughout, a design choice that is not incidental but central to any regulatory acceptability argument.

What the demonstration did not do: Gordian's 208-page output has not been submitted to the NRC as an actual license application. No NRC reviewer has evaluated it against regulatory requirements in a live proceeding, and the agency has not issued any public statement on AI-assisted submission methodology. The team is still developing a benchmarking rubric to assign confidence grades to Gordian's outputs, and a reviewing agent that will evaluate AI-generated sections against NRC guidance before submittal is in progress, not yet deployed.

"Nuclear is poised to solve today's critical energy challenges," said Kevin Kong, CEO and founder of Everstar. "We're excited to partner with INL to meet the moment, working together to accelerate regulatory review and commercialization."

For this demonstration to function as a licensing policy precedent rather than a lab benchmark, four artifacts are worth tracking. A full public release of INL/RPT-25-87369 with the complete V&V methodology will reveal whether the confidence-grading system has regulatory-grade rigor. Any NRC staff guidance or public comment acknowledging AI-assisted drafting as an acceptable methodology would represent the actual green light the industry needs. A live license filing that explicitly cites AI-assisted drafting in its cover documentation would signal that a developer is willing to stake regulatory capital on the output. Finally, follow-on Genesis Mission funding awards tied to Gordian's next deployment phase would indicate DOE treats this as a program, not a one-time proof-of-concept.

The physics-aware, domain-specific architecture of Gordian is the credibility argument underpinning the whole exercise. General-purpose language models do not carry weight in NRC-adjacent work; tools built specifically for nuclear-grade technical content do. That specialization is why industry analysts have consistently flagged domain design as the prerequisite for regulatory acceptability, and it is why this demonstration is worth watching.

The one-day figure is verified. Whether it compresses an actual licensing schedule depends entirely on what happens next at 11555 Rockville Pike.