Guide ยท AI in NERC compliance
Using AI in NERC compliance without creating a new finding.
AI is already in your compliance team's browser tabs. Here's where it genuinely helps, where it creates audit and CIP risk, and the line we draw in our own product.
Where AI helps
Explaining a requirement to a new engineer. Summarizing a 40-page technical rationale. Drafting a procedure. Organizing evidence. These are knowledge tasks where a good-enough answer that a person checks is genuinely useful, and the risk is low.
The determinism problem
A compliance check isn't a knowledge task. It's a calculation that has to be repeatable. General-purpose models generate answers probabilistically: the same prompt can produce different logic, different assumptions and different edge-case handling on different days. An auditor asking how you determined a setting complies needs a fixed method, documented inputs, and the same answer if you run it again.
Where it shows up: PRC tolerances interpreted differently between runs (PRC-024 boundaries, PRC-025 margins), missing transformer data filled with an inferred value in PRC-019, dataset filtering that changes silently.
The "looks right" problem
Generative models are built to produce confident, well-structured answers. That's a feature for writing and a hazard for validation. A clean table with a clear conclusion invites trust, even when an assumed margin, a misapplied threshold or a missing unit-specific exception sits underneath it. Traditional engineering tools show inputs and steps. A chat answer shows a conclusion.
The question an auditor asks isn't "does this look right?" It's "show me how you know."
The CIP problem
Protection settings, facility ratings, relay coordination data, EMS parameters, topology and event analyses can be BES Cyber System Information, and even when they aren't, they're operationally sensitive. Uploading them to a public AI service raises questions your CIP program has to answer:
- CIP-011: where is the information stored, retained and disposed of, and who can access it?
- CIP-013: is the AI provider an assessed vendor in your supply chain risk program?
- CIP-003: do your policies cover this data flow, or is it a shadow workflow?
- CIP-004: do staff know what they're allowed to share?
- CIP-002: do you know which information is tied to which BES Cyber Systems?
A policy checklist
Before people paste operational data into any AI tool, decide and write down:
- Which tools are approved, and in which configuration (enterprise tenant, retention off, no training)?
- Which data classes are never allowed in (BCSI, settings, topology), and which are fine (public standards, published guidance)?
- Whether AI output can ever be compliance evidence. Our view: not unless the method behind it is fixed and documented.
- Who reviews AI-assisted work products, and how that review is recorded.
- How the vendor is assessed under CIP-013.
Where we draw the line in our own product
data phleet uses machine learning to read documents, because that's where it's strong: pulling a pickup value off a scanned setting sheet or a rating off a nameplate. Every extracted value carries its source and a confidence, and anything uncertain goes to a person.
The compliance decision is deterministic code written per requirement with NERC experts. Same input, same answer, calculation shown, logic versioned. It runs inside the deployment you choose, up to fully on-premise, and your data never trains shared models.
More from us on this: non-deterministic compliance logic, the "looks right" problem, and sensitive grid data in public AI tools.
AI and compliance questions
Can I use ChatGPT or another general-purpose AI for NERC compliance checks?
For understanding standards, drafting and training, it can help. For the compliance determination itself, it's a poor fit. General-purpose models are probabilistic, so the same question can produce different answers with no fixed method to show an auditor. And pasting settings, ratings or topology into a public tool may put BES Cyber System Information outside the controls your CIP program requires.
Which CIP standards are relevant to using public AI tools?
Depending on your architecture and policies: CIP-011 (protecting BCSI: storage, retention, disposal, access), CIP-013 (supply chain risk from a new vendor and cloud processor), CIP-003 (policies governing how BCSI is handled), CIP-004 (training staff on what they can share), and CIP-002 (knowing which systems and information are in scope).
What does deterministic compliance logic mean?
The requirement's thresholds, tables, exclusions and clocks are written as explicit code, so the same inputs always produce the same output with the calculation shown. It can be re-run years later and give the identical answer, which is what an audit record needs.
See it on your own records.
Thirty minutes with one site's records: we surface the evidence live and show you the gaps. No slides.