From Pilots to Practice: NRC AI Workshop #7 Recap

RTS Research · Workshop Recap · By Sam Rieck, Founder and Chief Nuclear Officer

Nuclear AI is entering the implementation phase. At the NRC’s seventh AI workshop on September 15, 2026, the question was no longer whether AI can add value. It was where AI belongs in a workflow, what evidence shows it is fit for purpose, and who keeps authority when the output matters.

Six key findings

• Trust is a system property. Trust depends on curated data, explicit task boundaries, traceability, validation, monitoring, and accountable human decisions, not model claims alone.
• Production value is measurable. Duke reduced an aging management review from about five months with ten engineers to about one month, while targeting 99 percent recall for critical records.
• The industry is converging on bounded autonomy. AI retrieves, ranks, optimizes, or recommends within qualified envelopes, while deterministic calculations or qualified professionals anchor decisions.
• Data access is the binding constraint. Legacy analog assets, fragmented document systems, inconsistent metadata, and security boundaries increasingly decide what can be deployed.
• Regulatory readiness is coming through use. NRC staff showed AI supporting inspection work and the AP1000 design certification review. International tabletop exercises exposed gaps in data quality, governance, and uncertainty treatment.
• Workflow design is the next frontier. Strong programs define the work first, set success metrics with subject matter experts, build evaluation sets, and only then select or train models.

A workshop about operational maturity

The hybrid meeting at NRC Headquarters in Rockville, Maryland connected three perspectives: the NRC’s internal AI maturity path, emerging regulatory readiness work, and industry applications moving into production. Speakers presented patterns matched to different consequence levels: retrieval over controlled documents, machine learning for classification, physics-informed surrogates for core design, knowledge graphs for traceability, non-invasive sensors for legacy assets, and localized compute for protected plant data.

The workshop consistently supported a graded approach. Discovery, summarization, and classification tolerate different controls than safety-related calculation or autonomous action. As consequence rises, expectations for provenance, validation, uncertainty treatment, independent confirmation, cybersecurity, and accountable approval rise with it.

Human in Command, bounded autonomy, and regulatory evidence

The most persistent debate concerned the “human in the loop.” A reviewer placed at the end of an AI process does not guarantee safety. A reviewer who sees the AI answer first can anchor on it, and tools that are usually right teach people that checking is rarely needed. Three control ideas converged:

• Human in Command®: a qualified person or organization defines the objective, authorizes use, owns the decision, and can stop or change the process.
• Bounded autonomy: a model may act without step-by-step review only within a tested task envelope, with defined inputs, outputs, limits, and escalation points.
• Deterministic anchors: AI steers, retrieves, or optimizes while established calculations or one-way physical controls supply the authoritative result.

“The agent assembles the structure and exposes the gaps, but the human makes the safety judgment.” Sam Rieck, RTS

What regulators signaled

NRC staff described AI assistance inside existing work: inspection preparation through SimplifAI, and a data-centric audit approach on the AP1000 Revision 20 review that generated advanced safety evaluation report templates within hours. The international RegLab, run by the OECD/NEA and CNSC, concluded that failure modes sit in the system rather than the model alone, and that autonomy does not move safety duties away from the licensee. It moves them forward into guardrails, verification, and governance evidence.

The implication: governance should name the accountable role, the decision, the evidence the reviewer must see, the failure modes to check, the time available to intervene, and the conditions that force escalation.

The real infrastructure is information

• TerraPower moved from semantic search over more than 100,000 project documents to a knowledge graph with about 6.5 million data points and 10 million relationships, with permission-aware access and links back to evidence.
• Cypress Envirosystems described non-invasive digital overlays on analog gauges across 40 nuclear plants, lifting readings from about two manual observations a day to as many as 96, without license amendment requests.
• InfraShield argued for keeping AI inside the plant boundary, on the Level 2 network behind a one-way data device, citing an estimated nine-month return on investment versus continuing cloud costs.

What deployments are teaching the industry

• Duke: define the error that matters. For subsequent license renewal, a classifier separated aging-related condition reports from other records. Because the consequential failure was a missed record, the team designed for recall, not generic accuracy.
• Southern Nuclear: make the output fit the work. With Everstar, AI screened procedures against a 111-criterion rubric built with procedure writers. The Hatch phase scanned 1,334 procedures. Users wanted familiar deliverables, including Word documents with tracked changes.
• INPO: organize performance data for intervention. Dashboards help evaluators find what changed between periods and focus resources, rather than labeling a station automatically.

The strongest deployments shared five traits: a narrow workflow, a consequential error defined in advance, an expert-built evaluation set, evidence visible to the user, and a feedback loop that treats deployment as the start of validation.

AI around physics, simulation, and control

• MLAPSE (Numerical Advisory Solutions) uses machine learning to choose the next simulation cases, while confirmatory simulation remains the analysis of record.
• CoreDesigner.ai (Blue Wave AI Labs) predicts core designs in seconds from millions of licensed-code simulations, with promising designs confirmed by licensed methods.
• CERVEROS (University of Illinois) remotely controlled the PUR-1 research reactor against a setpoint derived from real AI computing load, as a testbed for latency, load following, and cybersecurity.
• Cambrian Nuclear applied AI to site selection, exploring more alternatives while preserving deterministic outputs and documentation.

Part 53 as a machine-readable decision architecture

In the flash talk session, Sam Rieck presented RTS research developed with N. Prasad Kadambi, author of NUREG/BR-0303 and co-author of the RTS policy report Nuclear AI as the Key to Part 53. Part 53 provides a technology-inclusive, risk-informed, performance-based structure, and agentic AI may make its decomposition and traceability work practical on a real schedule.

• Anchor: §53.230(a) sets the primary safety function, limiting radioactive release, supported by control of reactivity, heat generation, heat removal, and chemical interactions.
• Decompose: the hierarchy runs from the primary safety function to supporting functions, systems, components, and the evidence that each is met.
• Assist: agentic AI can carry out preliminary decomposition, traceability, and iteration at project speed. The method has existed since at least 2002. What changed is feasibility.
• Command: people own the safety judgment, and every generated element stays traceable.

RTS recommended a topical report pathway to define the methodology, inputs, controls, traceability expectations, and Human in Command model before a first applicant has to invent them under schedule pressure.

“Part 53 created the regulatory opportunity, and agentic AI may now make the performance-based path practical enough to use.” Sam Rieck, RTS

A practical agenda for nuclear AI

For operators and developers

• Start with one bounded workflow, and write down the consequential error before choosing a model.
• Build the evaluation set with the people who perform and approve the work.
• Preserve evidence: source links, inputs, model and prompt versions, reviewer decisions, and change history.
• Match deployment architecture to information sensitivity and plant network level.
• Design human authority explicitly: who can authorize, reject, override, pause, and recover control.

For the NRC and standards community

• Publish worked examples of acceptable data characterization, validation, uncertainty treatment, monitoring, and change control.
• Use a graded, consequence-informed assurance model.
• Clarify expectations for continuously changing models and the qualified process around them.
• Develop reference datasets, benchmark tasks, and audit artifacts.

For AI vendors

• Expose the operating envelope and known failure modes in terms engineers can test.
• Integrate into established nuclear work products.
• Treat permissions, cybersecurity, traceability, and export controls as product requirements.
• Measure quality, uncertainty reduction, knowledge retention, and risk alongside hours saved.

The workshop’s strongest message was disciplined optimism. Nuclear organizations are showing real value today, but durable adoption will depend on evidence, architecture, and accountable command more than on model novelty.

About the author

Sam Rieck is Founder and Chief Nuclear Officer of RAISUN Technology Services (RTS), helping utilities, suppliers, and advanced reactor developers implement AI across the nuclear sector while holding to the industry’s standards for safety, quality, and compliance. Sam holds bachelor’s degrees in Mechanical and Nuclear Engineering and in Vocal Music Performance from Kansas State University, and master’s degrees in Nuclear Science and Engineering and in Technology and Policy from MIT.

Talk to RTS

One conversation. No pitch deck. Schedule a roadmap session or browse RTS research.