A nuclear-industry AI assistant called NIVA is out of beta and in use at power plants across North America. Its practical promise is to make specialist operating and regulatory knowledge easier to find, while keeping access tied to the permissions users already have.
Watch Desk analysis
What happened
NIVA was developed through a collaboration between the Electric Power Research Institute (EPRI), the Institute of Nuclear Power Operations (INPO) and the Nuclear Energy Institute (NEI), with Atomic Canyon selected as its technology partner. Nuclear News reports that the tool uses retrieval-augmented generation to search documents held across separate databases rather than training a model on the organisations’ closely held data.
That federated setup matters: users can retrieve only material they are already entitled to access. NIVA’s nuclear-focused FERMI models were trained on more than 53 million pages of public Nuclear Regulatory Commission data, while the system also connects users to other permitted industry documents. New reports and operating experience can be added to its searchable databases; answers provide citations and links to source documents. Read the Nuclear News report.
Why it matters
Nuclear utilities rely on a large body of technical, operational and regulatory material, and finding the right document across separate systems can be laborious. NIVA is designed to bring that search into one place, including helping users find current regulatory guidance while still locating older versions relevant to a plant’s licence commitments.
The system also has an API using Model Context Protocol, allowing utilities to connect NIVA as a data source to enterprise AI tools. The report says a later test was under way at EPRI for combining NIVA resources with local, station-specific information. That is a distinct stage from the current access model: confidential station-specific information is not held in NIVA’s databases.
Our read
This is a more interesting AI deployment than another general-purpose chatbot with a hard hat on. The useful design choices are specific: retrieval across separate databases, existing access controls, cited answers and a route to connect the tool to other workplace software. Those features make the system’s information boundaries part of the product, not a footnote.
The important test is whether nuclear staff can find and check relevant material faster without treating an AI-generated answer as a substitute for the cited source. In a sector where the correct revision of a document matters, the citations are doing real work.
What to watch
- Whether utilities report measurable improvements in finding and using operating experience.
- How the EPRI test of station-specific information through NIVA’s MCP connection develops.
- Whether additional industry groups contribute data while preserving their access agreements.
- How NIVA performs as new reports and regulatory guidance are added.
Discussion spark: Should a specialised industry AI assistant be allowed to search across organisations’ databases when it preserves existing access permissions, or is that still too much shared infrastructure for sensitive operational knowledge?
Sources and evidence
- Source update (13 November 2026, 00:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.