Let’s be honest: most nuclear teams don’t have the time (or sometimes the interest) to think deeply about AI. Between outage prep, staffing gaps, and the daily grind of designing or operating a plant, nobody’s running prompt experiments for fun. So they watch demos quietly, buy tools to “stay current,” and end up with unused seats gathering digital dust while the talent drain keeps getting worse.

The good news: you don’t need to evaluate everything. You just need to know what signals to look for. In this article, I’ll share our team’s thoughts drawn from years of business development, software engineering, and guiding nuclear customers through the sales cycle. They fall into two categories: (1) how to pick a vendor and (2) how to stress-test that vendor effectively.

Deciding Which Vendors to Bother With

A “free trial” might make sense for consumer software, but every nuclear deployment involves cybersecurity reviews, IT integration, and a fair bit of paperwork. You have to do your homework and believe it’s worth the effort.

1. Nuclear-Smart AI Starts with Nuclear-Smart Teams

Every week brings another vendor claiming they have the best AI for your needs. Some may even be right. Start with this mindset: Can they show, then tell?

A credible vendor should be able to discuss and demonstrate specific nuclear use cases. If they can’t pull up a real example and talk about how it would have made their job easier in a past life, they’re probably not ready for your work.

At Everstar, our nuclear team is built for quality. We’ve written, approved, inspected, or audited every kind of analysis and workflow we now automate. We aren’t just counting years in the chair—we’re seeking proven impact.

I test every feature myself. When answers fall short (they rarely do), I work directly with our engineers on the fix. Some days I’m stress-testing it like an NRC inspector; other days I’m probing something completely new. The feedback loop is fast, and I’m usually pleasantly surprised.

2. You Aren’t Supposed to Be an Expert (Yet)

AI in nuclear isn’t one thing. It’s a range of capabilities: document generation, vision models, design iteration with on machine learning, training modules, and incident screening or prioritization. The best vendors teach you what’s possible and debunk what’s just marketing.

If a vendor can’t clearly articulate three use cases that matter to your plant, engineering team, or consulting shop, they’re probably selling you a generalized productivity tool with a nuclear sticker on it.

Nuclear teams don’t have patience for hand-waving. If a vendor can’t speak clearly about real tasks, they’re not ready for real work.

3. Momentum Matters

Screenshots and short videos on a vendor’s website are important proof points. They show direction and momentum. Even if they’re outdated the day after they’re posted, they prove that a vendor understands your domain.

AI is changing every day. Any demo you get is just a snapshot in time. When you evaluate vendors, look for rate of improvement, not whether it exactly matches what you’re looking for. How quickly are they integrating feedback? Shipping updates? Refining nuclear context?

Vendors standing still will soon be irrelevant. So will teams waiting for “final” versions of tools that never arrive. Look for companies that will work with you to get ROI out of what is already on offer while adjusting to maximize your benefits.

4. Custom Without Reinventing the Wheel

Don’t ask for bespoke systems built from scratch. You’ll spend a year waiting for results while other companies lap you. Ask for adaptable tools that build on what already works. The best vendors can combine proven AI infrastructure with the context you need to solve your problems—quickly. Also, listen for a promise that they won’t train on your sensitive data without explicit, written approval.

(Stay tuned for a deeper dive soon on why we use retrieval, not retraining, to build explainable, secure systems that scale.)

5. Nothing Under the Rug

Some important topics can get skipped in a flashy overview. Be ready to ask harder questions:

• Tell me why you’re showing these specific workflows?
• Are there limitations you aren’t showing me?
• How do you choose your source materials?
• What types of nuclear documents has your system actually handled?
• How do you balance accuracy, traceability, and speed?

Also listen for what comes unprompted. The best vendors talk about risk management before you ask. They acknowledge that applying AI to nuclear work isn’t zero-risk. It’s about managing risk responsibly, just like every other engineering decision.

6. A Practical Framework for Choosing Well

When you’re in the early stages of a vendor evaluation, start with these five questions:

1. Can they show real examples using nuclear-specific documents?
2. Do they explain their decision logic or only show polished results?
3. What happens when the system is wrong—can they trace it?
4. How fast can they adapt to your use case without a rewrite?
5. How do they handle feedback and continuous improvement?

Seeing What Really Works

So when you're ready to get started, what are you actually supposed to do? Here are a few tips for maximizing your first interaction with an AI vendor. No digital dust for you!

1. Date Before You Marry

Skip the idea of a “free trial.” You don’t want the vendor who tosses you a login and disappears. You want the one who invests time to train your users, tailor the experience, and make sure the tool earns its place in the workflow.

To choose the best tools for your business, you need a real evaluation period**—**roughly 60 to 90 days—where your team tests tools side by side while doing their actual work. Keep scope and number of users tight, with a weekly feedback cadence.

Set clear expectations. What use cases will you test? What documents or data will you share? What light customizations will they provide at low or no cost? The best partners will help you define that scope so both sides learn something meaningful.

Keep the pilot small and affordable: a focused group, measurable outcomes, and a fair price. If you need full 810-compliant configurations, expect higher cost. That’s the price of secure compartmentalization.

2. Make it Concrete and Specific

You’re spending money on this evaluation. Make sure you know what you’re getting.

I recommend that you develop a structured test plan with your AI vendor. Prioritize getting answers and outputs that actually matter in your world. You should be testing the kind of tasks your engineers, trainers, and analysts handle every day. Make sure you get specific advice on how to run all of the scenarios you are curious about, and be ready to do the same for other tools you have available.

Your goal is to map strengths and limits against your stack. If they can fix gaps during the pilot, great—if not, you just learned cheaply.

3. Use KPIs to Make Your Decision Obvious

Decision rule: If ≥4 of 5 metrics hit target and two real workflows are in active use by day 60, proceed. If not, cut bait.

• Adoption: weekly active users / invited users (target ≥70% by week 4)
• Throughput: % reduction in time to produce a given artifact (e.g., 50.59 screen, RAI response outline, training module)
• Citation fidelity: audited sample of answers with correct, resolvable sources (target ≥95%)
• Review burden: reduction in reviewer edits/comments per document
• Change requests closed: # of vendor-delivered improvements during pilot

4. Try Our Questions to Get Started

Opening an AI tool for the first time can feel like walking into a machine shop with no training, or trying to make dinner from whatever’s in your pantry. Too many options, too much power, and you don’t know how to start. Context and examples are what make technology usable. You need a cookbook or a manual.

Don’t waste your pilot period asking generic questions like “Summarize this PDF.” I’ve assembled some of my favorite trial questions to help you get started. Are you happy with the answers you’re getting today?

👉 Download our sample questions here.

No surprise: Gordian returns source-cited answers that stand up to expert review. Contact me at theresa@everstar.ai if you want to see OUR answers to your favorite questions.

The Bottom Line

AI adoption in nuclear isn’t optional. It’s the only scalable way to manage talent drain while maintaining performance. In competitive markets, the AI-native players will outpace the rest. Even in stable, regulated settings, slow adopters will feel the drag of declining knowledge and rising costs.

Start small, but start now. Reach out to me today at theresa@everstar.ai to start this journey.