Month: September 2025

Executive Summary

A new pattern is emerging in enterprise AI adoption: Shadow AI. Employees, frustrated with slow-moving official systems, are turning to consumer-grade AI tools like ChatGPT to get work done. MIT’s State of AI in Business 2025 study reports that over 90% of knowledge workers now use unsanctioned AI for drafting, research, and analysis.

In many industries, this trend raises governance and security concerns. In nuclear, it introduces regulatory, safety, and compliance risks that cannot be tolerated.

This paper examines Shadow AI as an enterprise trend, analyzes why it is incompatible with nuclear operations, and outlines how regulator-ready, domain-specific AI addresses the gap.

---

1. Shadow AI: A Growing Trend

MIT’s research identifies Shadow AI as one of the fastest-growing dynamics in AI adoption:

• Broad worker adoption. Employees bypass enterprise systems and use consumer AI tools directly.

• Enterprise lag. Corporate IT and compliance groups struggle to deploy secure alternatives at the same pace.

• Risk exposure. Sensitive information is copied into cloud-based tools with little oversight.

In industries like retail or marketing, the risks are financial. In nuclear, they are regulatory and existential.

---

2. Why Nuclear Cannot Tolerate Shadow AI

Nuclear operations rely on strict compliance regimes that consumer AI tools cannot meet:

• Part 810 Regulations. Export control prohibits the uncontrolled transfer of nuclear technical data. Shadow AI platforms, typically hosted on global cloud infrastructure, are non-compliant by default.

• Licensing-Basis Sensitivity. Technical Specifications, FSARs, and design-basis documents cannot be exposed to uncontrolled platforms. Even summaries must be regulator-ready.

• Audit Requirements. NRC oversight requires every evaluation and document to be traceable and verifiable. Shadow AI outputs are not.

Simply put: Shadow AI creates compliance gaps that nuclear regulators and operators cannot accept.

---

3. Case Study: Licensing Research

Observed Behavior:
Frustrated by slow search systems, engineers tested consumer AI tools to summarize licensing requirements. The AI produced text that appeared helpful but lacked citations and omitted key references.

Outcome:
Outputs could not be defended in NRC-facing documentation. The practice created risk, not efficiency.

Domain-Specific Alternative:
Nuclearn’s Gamma 2 model retrieves licensing basis documents from secure, on-premise repositories. Outputs include full citations, reasoning steps, and maintain alignment with IV&V. Engineers remain in control, but repetitive search is automated.

Result: regulator-ready documentation, compliant with Part 810, without Shadow AI risk.

---

4. Shadow AI vs. Secure AI

The Shadow AI trend reflects a workforce reality: employees want faster, more usable tools. Restrictive policies alone will not stop Shadow AI. Without secure alternatives, adoption will continue underground.

The solution is not prohibition, but replacement. Nuclear operators must provide systems that are:

• As usable as consumer AI. Engineers will only adopt what improves their daily work.

• As secure as required. On-premise, Part 810 compliant, and regulator-ready.

• Domain-specific. Trained on nuclear acronyms, licensing structures, and workflows.

---

5. Implications for Nuclear Operators

• Shadow AI is not theoretical. It is already happening across industries. Nuclear cannot assume immunity.

• Regulatory exposure is immediate. Even one instance of sensitive data entered into a consumer AI platform may trigger compliance investigations.

• Workforce demand must be addressed. Engineers will seek usable AI. If utilities don’t provide compliant systems, Shadow AI will fill the gap.

---

Conclusion

Shadow AI is the new trend shaping enterprise AI adoption. In nuclear, it is untenable. The compliance, regulatory, and safety demands of the industry mean that consumer-grade AI tools cannot be tolerated inside the plant.

The solution is not banning AI use — it is providing secure, domain-specific alternatives that meet the same standard as the industry itself: safety, compliance, and reviewability.

Nuclearn demonstrates that when AI is designed for nuclear — Part 810 compliant, regulator-ready, and embedded in real workflows — it delivers measurable value while eliminating the risks of Shadow AI.

Executive Summary

Enterprise adoption of generative AI is widespread, but measurable impact remains rare. The MIT State of AI in Business 2025 report found that only 5% of enterprise AI pilots advance into production. The remainder stall due to integration challenges, lack of compliance alignment, and outputs that do not withstand scrutiny.

In nuclear energy, this failure rate cannot be tolerated. Pilots that never scale waste engineering hours, introduce compliance risk, and erode workforce trust. This paper examines why most AI efforts fail to transition, why nuclear’s regulatory environment magnifies the risk, and what design principles are required for AI systems to succeed in production.

---

1. The Pilot Trap

Across industries, the “pilot trap” is common. Demos and small-scale trials show potential but collapse when scaled. Three recurring factors are identified in MIT’s research:

1. Workflow Misalignment – Pilots address isolated tasks but fail when integrated into enterprise systems.

2. Compliance Blind Spots – Outputs lack the transparency needed for audit or regulatory review.

3. Cultural Resistance – After repeated failures, workforces lose trust in AI initiatives.

For most industries, these failures represent opportunity costs. In nuclear, the consequences are higher. Every pilot requires engineering time, often from senior staff. If the pilot fails, scarce expertise has been diverted from safety and operational priorities.

---

2. Why Nuclear Is Different

Nuclear operations impose requirements that generic AI tools rarely meet:

• Independent Verification and Validation (IV&V): All calculations, evaluations, and analyses must be reviewable. Outputs that cannot be traced to source data are unusable.

• Part 810 Compliance: U.S. export control regulations prohibit uncontrolled data transfer. Cloud-hosted consumer AI platforms cannot meet this requirement.

• Licensing Basis Alignment: Documentation associated with plant licensing must withstand regulatory audit. Outputs that lack defensibility introduce unacceptable risk.

These conditions mean that nuclear cannot rely on general-purpose AI. Tools must be designed specifically for regulated, documentation-heavy workflows.

---

3. Case Study: Condition Report Screening

Nuclear plants generate thousands of Condition Reports annually. Each requires screening for safety significance, categorization, and assignment. Historically, this workload demands dedicated teams of experienced staff.

Pilot attempts with generic AI:

• Demonstrated short-term gains in categorization speed.

• Failed to provide traceable reasoning or regulatory-suitable documentation.

• Stalled at the pilot stage due to lack of reviewability.

Production deployment with nuclear-specific AI:

• Automated initial screening with embedded reasoning steps and citations.

• Retained IV&V by keeping engineers in the review loop.

• Scaled to full fleet use, saving tens of thousands of engineering hours annually.

This example illustrates the critical distinction: pilots demonstrate potential; production requires compliance-ready outputs.

---

4. Case Study: 50.59 Evaluations

The 50.59 process determines whether plant modifications require NRC approval. Evaluations typically require 8–40 hours of engineering time and extensive document research.

Pilot attempts with generic AI:

• Produced draft summaries of licensing documents.

• Lacked sufficient traceability for NRC acceptance.

• Failed to progress beyond trial use.

Production deployment with nuclear-specific AI:

• Retrieved relevant licensing basis documents with citations.

• Assembled draft evaluations in ~30 minutes.

• Enabled engineers to complete reviews in ~2 hours, maintaining full compliance.

The ability to produce regulator-ready outputs was the determining factor in moving from pilot to fleet deployment.

---

5. Lessons from MIT Applied to Nuclear

MIT’s research identifies three conditions for bridging the gap between pilots and production:

1. Domain Specificity: Tools must be trained on industry-specific data sets.

2. Workflow Integration: Systems must embed within existing processes rather than operate in isolation.

3. Adaptive Learning: AI must improve with use and align with regulatory context.

Applied to nuclear, these principles translate to:

• Training models on NRC filings, license renewals, and utility documents.

• Embedding tools into CAP, 50.59, and outage workflows.

• Designing outputs for traceability, citation, and regulatory review.

Without these conditions, AI pilots in nuclear will remain demonstrations with no lasting impact.

---

6. Implications for Nuclear Operators

The findings have clear implications:

• Evaluate vendors beyond demos. Demand evidence of regulator-ready outputs, not just functional prototypes.

• Prioritize compliance from the start. Systems must be Part 810 compliant and built for IV&V.

• Focus on critical workflows. Target documentation-heavy processes where measurable impact can be achieved without compromising safety.

• Guard against cultural fatigue. Each failed pilot increases resistance. Operators should commit only to systems designed for production.

---

Conclusion

The majority of enterprise AI pilots fail to transition into production. In nuclear, this failure rate is not sustainable. Documentation is safety-critical, compliance is non-negotiable, and workforce trust is essential.

To bridge the gap from pilot to production, AI systems must be domain-specific, workflow-integrated, and regulator-ready. Evidence from early deployments shows that when these conditions are met, nuclear plants can save thousands of engineering hours annually while maintaining safety and compliance.

The lesson is clear: nuclear must move beyond pilots. Production-ready AI, designed for nuclear, is not optional — it is required.

Introduction

Across industries, generative AI is being tested in pilots, proof-of-concepts, and trials. The promise is simple: automate routine work, generate documentation faster, and let knowledge workers focus on higher-value tasks.

But the data tell a different story. In its State of AI in Business 2025 report, MIT found that 95% of enterprise GenAI pilots fail to deliver measurable value. Most never move beyond a demonstration. They stall because they don’t integrate into workflows, they forget context, or they produce outputs that can’t be trusted in regulated environments.

For nuclear, this failure rate isn’t just disappointing — it’s unacceptable. Documentation in nuclear isn’t optional; it is the backbone of safety, compliance, and regulatory oversight. If an AI tool cannot produce outputs that are traceable, reviewable, and regulator-ready, it has no place inside the plant.

This is the GenAI Divide. Most industries are struggling to cross it. Nuclear requires a different approach.

What MIT Found

MIT researchers analyzed more than 300 AI initiatives and interviewed senior leaders across industries. Their conclusions highlight why adoption is high but impact is low:

• High pilot activity, low production: More than 80% of organizations have tested tools like ChatGPT or Copilot. Fewer than 5% of custom AI solutions made it to production.

• Generic adoption, limited disruption: Consumer tools help with quick drafting, but enterprise-grade deployments stall.

• The learning gap: Most tools don’t retain context, adapt to workflows, or improve over time. This brittleness means they can’t handle complex processes.

In short, pilots succeed at showing potential. They fail at delivering operational transformation.

Why Nuclear Can’t Afford the Divide

In many industries, failed pilots mean lost time or missed efficiency. In nuclear, they can undermine safety and compliance.

1. Documentation is not peripheral.
   Every Condition Report, Corrective Action Program entry, or 50.59 evaluation is required by regulation. These aren’t internal notes; they are part of the permanent regulatory record.

2. Traceability is essential.
   Every calculation, every engineering judgment, every modification review must be linked back to source material. If outputs cannot be cited and verified, they cannot be used.

3. Workforce turnover magnifies the need.
   With a quarter of the nuclear workforce set to retire within five years, plants need tools that help new engineers become productive quickly. AI that generates unreviewable or inaccurate documentation wastes scarce expertise instead of preserving it.

The conclusion is clear: nuclear cannot tolerate the 95% failure rate seen in other industries. AI must meet the same standards as the industry itself — safety, transparency, and compliance.

Nuclearn’s Approach

Nuclearn was founded by nuclear professionals who saw these challenges firsthand at Palo Verde. Our approach is fundamentally different from generic AI deployments:

• Nuclear-specific data sets: Our Gamma 2 model is trained on NRC filings, license renewals, technical specifications, and utility-provided documentation. It understands the acronyms, licensing basis requirements, and processes unique to nuclear.

• Reviewable outputs: Every output includes citations back to source material and exposes the AI’s reasoning steps. Engineers can perform independent verification and validation (IV&V) just as they would for junior engineer work.

• Workflow integration: Nuclearn doesn’t sit on the side as a chatbot. It is embedded into CAP screening, 50.59 evaluations, outage planning, and licensing research — the real processes that consume plant resources.

• On-premise, secure deployment: Data never leaves plant control. Our systems are Part 810 compliant and designed to meet U.S. export control regulations.

Case Example: CAP Screening

At a typical reactor, thousands of Condition Reports are filed every year. By regulation, every CR must be screened and categorized: is it adverse to quality? Does it require corrective action? Which group is responsible?

Historically, this requires full-time teams of experienced staff. It is repetitive, manual, and essential.

With Nuclearn:

• AI automates the screening and categorization process.

• Experienced engineers remain in the loop, reviewing and verifying.

• Plants save tens of thousands of hours annually, freeing highly skilled staff for higher-value work.

The process is faster and more consistent — but still compliant with regulatory expectations for reviewability.

Case Example: 50.59 Evaluations

The 50.59 process requires engineers to determine whether a proposed modification changes the plant’s licensing basis and whether NRC notification is required. It is one of the most documentation-intensive processes in the industry.

Traditionally:

• Each evaluation takes between 8 and 40 hours.

• Engineers must search thousands of pages of licensing documents.

• Work often involves multiple layers of review and verification.

With Nuclearn’s agent-based workflows:

• Relevant licensing basis documents are retrieved automatically.

• Key requirements and citations are assembled.

• Engineers receive a draft evaluation in about 30 minutes.

The final review still takes human expertise, but the process now takes ~2 hours instead of several days. Outputs remain fully traceable, with citations back to source material for regulatory confidence.

Aligning with Industry Findings

Where most AI pilots fail, Nuclearn succeeds because our approach directly addresses the barriers highlighted by the industry reports:

• Process-specific customization: We don’t try to solve everything. We focus on CAP, 50.59, outage planning, and licensing.

• Workflow integration: Our tools are embedded in actual plant processes, not running in isolation.

• Learning and adaptation: Our models are trained on nuclear-specific data and tuned for each utility.

• Compliance and traceability: Outputs are regulator-ready, built for IV&V.

This is exactly what MIT identifies as the path across the GenAI Divide: adaptive, embedded, domain-specific systems

Closing Thought

The MIT study is a warning. Most enterprises will spend money and time on AI tools that never scale. They will produce demos, not durable solutions.

Nuclear does not have that luxury. Our industry requires AI that can withstand NRC oversight, peer review, and decades of operational scrutiny. That is what Nuclearn delivers: solutions that are reviewable, verifiable, and regulator-ready.

If AI can meet nuclear’s bar, it can meet any bar.