AI for Nuclear Energy
Nuclear is the sector where we are most explicit about what we will not build. Everything safety-classified stays with your qualified engineers and your regulator; the useful work is substantial and sits entirely elsewhere.
A nuclear licensee operates under a safety case that assigns responsibility to qualified people and demands demonstrable justification for everything within it. That is not an obstacle to work around; it is the framework, and the sensible applications sit outside it.
AI for nuclear energy covers non-destructive testing and inspection review support, technical documentation and configuration management, outage planning and scheduling, decommissioning characterisation, supply chain and obsolescence analytics, and non-safety operational analytics.
What we will and will not build
| Application | Position | Reasoning |
|---|---|---|
| Technical document retrieval | Build | Enormous controlled corpora; the strongest fit in the sector |
| Configuration and modification records | Build | Structured, high-consequence information work |
| NDT and inspection review prioritisation | Build, carefully | Orders the queue; the qualified inspector decides |
| Outage planning and scheduling | Build | Complex constrained scheduling with real cost impact |
| Decommissioning characterisation | Build | Survey data volumes exceed manual analysis capacity |
| Supply chain obsolescence | Build | Long-lived plant, discontinued components, real risk |
| Anything within the safety case | Decline | Qualified people and a regulator own this, correctly |
| Safety-classified control functions | Decline | Not a boundary we will approach |
We do not work inside the safety case, and we say so first
Systems that are safety-classified, or that inform safety-classified decisions in a way the safety case relies on, belong to qualified nuclear engineers working with the regulator. We are not that, and a supplier who suggests otherwise is misunderstanding the framework or hoping you do. Everything on this page sits deliberately outside — and it is still a substantial body of valuable work, which is the point.
Documentation and configuration, where the value concentrates
A nuclear site holds decades of technical documentation under strict revision control: design records, modification histories, operating instructions, maintenance procedures, safety case references. Finding the right document at the right revision is a persistent operational cost.
Grounded retrieval with revision fidelity
Every answer traceable to a specific document at a specific revision, with the system declining rather than inferring where the corpus is silent. In this sector that is the defining requirement, not a refinement. See RAG development.
Configuration questions are high-consequence information work
Which modification applies to which system, which drawings are current, which components are interchangeable, what a change affected downstream. This is structured and answerable, and answering it faster is genuinely valuable.
Deployment is inside your boundary
Technical data here is frequently export controlled and always sensitive, which means models run in your environment rather than a hosted service. That is a first-week architectural decision. See private LLM deployment.
The output supports a person; it does not replace a check
A retrieval system that helps an engineer find the right document faster is valuable. One presented as authoritative on a technical question is a different thing, and we build the first.
Outage planning and decommissioning
- Outage planning is a large constrained scheduling problem. Tasks, resources, permits, radiological constraints, sequencing dependencies and a critical path where every day has a large value.
- Historical outage data supports duration estimation. Which tasks consistently overrun, and under what conditions, is learnable from your own history.
- Decommissioning generates survey data at scale. Radiological characterisation across large volumes produces more data than manual analysis absorbs.
- Waste categorisation has real cost consequences. Better characterisation reduces conservative over-classification, which is expensive across a decommissioning programme.
- Obsolescence is a genuine risk on long-lived plant. Predicting component availability problems years ahead is ordinary supply chain analytics with unusual stakes.
Decommissioning is where the data volumes are
An operating station is engineered to be stable and produces relatively little variation to learn from. A decommissioning programme produces enormous survey, characterisation and waste data across years, with real cost consequences attached to classification decisions. It is the part of the nuclear lifecycle where data analysis has the most to work with, and it is considerably less explored than operations.
How an engagement runs
The safety case boundary is established in week one, and everything is built outside it.
Boundary, export control and scope
What is safety-classified and therefore out of scope, and where technical data may be processed.
Corpus and data foundation
Technical documentation with revision control preserved, or outage and survey data structured.
Build
Grounded retrieval that cites document and revision, or outage scheduling with radiological and permit constraints.
Expert evaluation
Assessed by the engineers who will use it, on real questions, with error categories recorded.
Controlled operation
Version-pinned inside your environment, with change control and periodic re-evaluation.
What you receive
Faster access to what your engineers need, without going anywhere near the safety case.
Scope boundary
What is safety-classified and out of scope, agreed and documented before any build.
Grounded document retrieval
Citations to document and revision, declining where the corpus is silent.
Configuration data structuring
Modifications, drawings and component effectivity linked and queryable.
Outage scheduling support
Constrained optimisation over tasks, resources, permits and radiological limits.
Decommissioning characterisation
Survey data analysed at volumes manual review cannot absorb.
Obsolescence analytics
Component availability risk on long-lived plant, years ahead.
Is this the right starting point?
Worth being direct. There are situations in nuclear where custom AI work is the wrong spend, and those are listed rather than buried.
Worth doing if
- Engineers spend significant time locating documents at the correct revision.
- Configuration questions require manual searching across decades of records.
- Outage planning is done in spreadsheets against a critical path worth millions per day.
- Decommissioning survey data exceeds manual analysis capacity.
- Component obsolescence is discovered reactively on long-lived plant.
Do something else if
- You want AI inside the safety case. We decline, and would question a supplier who accepts.
- You want safety-classified control or protection functions. Not a boundary we approach.
- Technical data cannot be processed in an environment you control.
- Document revision control is unreliable, so grounding cannot be trusted.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Will you build AI for safety-classified systems?
No. Safety-classified functions, and systems the safety case relies on, belong to qualified nuclear engineers working with your regulator. We are not that and we say so at the first meeting. A supplier who offers to work inside the safety case is either misunderstanding the framework or hoping you do — and the work outside it is substantial enough that there is no need to go near the boundary.
What is the strongest application in a nuclear context?
Technical documentation and configuration retrieval. A site holds decades of records under strict revision control, and finding the right document at the right revision is a persistent daily cost. Grounded retrieval that cites document and revision, and declines where the corpus is silent, addresses that directly — and the grounding requirement is the defining design constraint rather than a refinement.
Can AI support NDT and inspection?
By prioritising the review queue and directing an inspector's attention within a scan, which addresses a genuine bottleneck. The determination stays with the qualified inspector, and the system should be tuned heavily toward sensitivity because a missed indication and a false positive are not comparable costs. Any tool used across a programme also needs its own qualification evidence, which is a programme-level control.
Where do the data volumes actually exist?
In decommissioning, more than in operations. An operating station is engineered to be stable and produces relatively little variation to learn from. A decommissioning programme generates enormous radiological survey and characterisation data with direct cost consequences attached to classification — conservative over-classification is expensive across a programme, and better characterisation reduces it.
How do you handle export control and sensitivity?
As a first-week architectural decision. Nuclear technical data is frequently export controlled and always sensitive, which rules out hosted services and points to models running inside your own environment. Settling that early is considerably cheaper than discovering it during a security review, and it constrains model choice in ways worth knowing before anyone gets attached to an approach.
Related verticals
Organisations in nuclear usually share data, buyers or regulators with these. All fourteen are listed on the Energy & Utilities page.
AI for Power Generation and Utilities
Demand forecasting, network asset health, inspection analytics and outage response — advisory, not actuating.
Read more →AI for Oil and Gas
Production optimisation, condition monitoring, subsurface analytics and integrity — inside the safety case.
Read more →AI for Water and Wastewater Utilities
Leakage detection, treatment optimisation and overflow prediction — where public accountability is the constraint.
Read more →Tell us what the problem looks like.
Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.