LLM.co · Private AI & Custom AI DevelopmentPrivate AI for energy & utilities

AI for utilities that respects the electronic security perimeter

Utilities hold decades of operating knowledge in switching procedures, equipment manuals, relay settings files, inspection records, outage logs and crew notes. Much of it lives on networks that are deliberately cut off from the internet. AI for utilities has to work inside those boundaries, or it cannot touch the information that matters most.

LLM.co builds private AI on open-weight models that run on hardware the utility controls. A system can sit in the corporate data center, in a control center environment or on an air-gapped network. It opens no new paths to the internet and makes no calls to an outside model provider.

What NERC CIP AI has to account for

Any system that reaches BES Cyber Systems or their information falls under the NERC CIP standards. That brings requirements for electronic security perimeters, access management, configuration change management, information protection and logging. NERC CIP AI is a short name for building AI so it fits those requirements from day one.

In practice, we classify each use case by the networks and data it touches before any build starts. Corporate use cases run on the IT side. Use cases that need BES Cyber System Information run on hardware inside the right zone, with access tied to your existing authorization records. Model and software updates arrive as signed bundles that go through your change and baseline process. The system is built to support your CIP program, and your compliance team decides how it is classified.

Workflows for operations, field crews and compliance

The strongest first projects sit close to work crews and compliance staff already do by hand.

• Procedure search. Operators and field crews ask plain-English questions and get the relevant switching or maintenance procedure, with the revision and section cited.
• Outage and event reports. Narratives are drafted from SCADA event logs, OMS records and crew notes for an engineer to review.
• Compliance evidence. Records for CIP and other audits are collected, matched to requirements and assembled into evidence packages with gaps listed.
• Asset knowledge. Equipment histories, nameplate data, test results and manuals become searchable by asset ID.
• Work package preparation. Job briefs and safety checklists are drafted from the work order, the asset history and the applicable procedures.

Deployment at the edge and in the field

Many utility use cases need answers where connectivity is poor. Smaller models can run on a ruggedized server in a substation or on a laptop in a truck, with a local copy of the approved procedures. Larger models run in the control center or data center on the OT network side. On-premises AI in each location follows the same identity, logging and update rules.

Systems in these environments read data and draft documents. They do not issue control commands to grid equipment. That boundary is set during scoping and enforced in the architecture.

What to pilot first and what to avoid

For most AI for utilities projects, procedure search for one operating area is a sound first pilot. The documents are controlled, operators can judge answers quickly, and the evaluation set comes from questions crews have already asked. Compliance evidence assembly is another good start because the output goes to people who check it carefully.

Avoid any tool that requires uploading CIP-protected information to a vendor cloud. Avoid connecting a model directly to SCADA or EMS write paths. Avoid starting with use cases where the source documents are out of date, since the model will repeat what it reads.

Why custom AI development fits utilities

Each utility has its own network zones, procedure formats and change process. Custom AI development lets the system match them. Work starts with a two-week discovery sprint, and a focused first system usually reaches production in eight to twelve weeks. You own the code, prompts, evaluation sets and any fine-tuned weights.