— Operational Risk Framework + Open-Source Tools
Trusted AI for the electric sector needs an operational risk layer.
SAFERai.power is an EPRI-led, Open Power AI Consortium (OPAI)-aligned initiative to build the risk framework, evidence templates, and open-source assessment tools needed before AI is deployed into utility workflows.
This is not a new general AI governance model but rather extends existing frameworks with power-sector evidence, testing, monitoring, and guardrails for the electric sector.
What it creates
A common way to evaluate AI by use case, consequence, autonomy, and operating context.
Built for utility operations, technology partner diligence, OPAI sandbox learning, and regulatory-facing evidence conversations.
Risk Framework
This framework includes risk scenarios, failure modes, autonomy levels, guardrails, and use-case tiers.
Evidence Packages
These will include AI system risk file templates for provider and deployer evidence.
Assessment Tool
This open-source toolkit will provide for tiered and scalable use-case risk assessment.
Safety Accountability Fairness Explainability Reliability
— Why Now?
AI adoption is accelerating faster than reusable risk evidence.
The electric sector is moving from experimentation and demonstrations to use across their ecosystem to include customer operations, planning, asset analytics, field work, restoration, DER integration, and emerging agentic workflows. Without a shared assessment framework, every company must invent its own approach.
How SAFERai.power fits with OPAI
EPRI’s Open Power AI Consortium convenes the ecosystem around shared datasets, open-source libraries, AI accelerators, sandbox evaluation, use-case groups, and implementation partners. SAFERai.power adds the complementary operational risk framework and assessment tools that help evaluate those AI assets before deployment.
Designed to enhance, not replace
The initiative maps to NIST AI RMF, ISO/IEC 42001, IEC 62443, NERC CIP concepts, the EU AI Act, and mature utility programs. The goal is to operationalize those frameworks for electric-sector use cases, not duplicate them.
Physical Consequences
AI failures can affect public and worker safety, asset integrity, system stability, and service continuity.
Interconnected Systems
Actions in one control area can propagate across interconnected operations and markets.
Data Constraints
Operational, customer, cyber-sensitive, and restricted data require disciplined provenance, access, and evidence handling.
Lifecycle Volatility
AI performance can drift with weather, DER behavior, topology, load shapes, and changing field reality.
AI Supply Chain
Utilities rely on technology partners, embedded platform features, foundation models, integrators, agents, and update streams.
— What We Are Building
Not just a document. A framework package that produces usable artifacts.
AI System Risk File
A minimum evidence package for material AI use cases: purpose, data provenance, capabilities, limitations, oversight, TEVV, deployment conditions, monitoring, incidents, and reassessment triggers.
U-SAFER Assessment Method
A stable question set that adapts by utility function, consequence, autonomy level, system access, write authority, and deployment context.
AI System Registry Schema
Minimum fields for owners, providers, deployers, intended purpose, dependencies, permissions, autonomy, risk tier, approval state, monitoring, and change triggers.
TEVV and Monitoring Protocols
Testing, evaluation, verification, validation, benchmarking, guardrails, safe fallback behavior, drift monitoring, anomaly monitoring, and revalidation triggers.
Provider and Deployer Evidence
Separate evidence tracks for technology partner product claims and local utility deployment reality: configuration, permissions, integrations, training, procedures, and monitoring.
Open-Source Toolkit Components
Reference components for LLM-assisted intake, registry-backed evidence management, assessment workflow, human review, and reassessment.
— FAQ
Your questions answered
Founding members make a two-year commitment to help co-develop the SAFERai risk framework before program expectations and standards are finalized. Participants work alongside EPRI and industry peers to shape the framework, evidence templates, OPAI-aligned pilot projects, and the open-source assessment toolkit.
Participation is offered at three levels based on organizational scale and capacity to contribute:
- Large Scale: $40,000 per year ($80,000 over two years) for large utilities, major IPPs, hyperscalers, AI platform providers, and infrastructure companies.
- Medium Scale: $25,000 per year ($50,000 over two years) for mid-size utilities, regional operators, established AI solution providers, implementation partners, and advisory firms.
- Small Scale: $15,000 per year ($30,000 over two years) for smaller utilities, electric cooperatives, municipal utilities, public power districts, and early-stage technology companies.
All participants receive equal substantive access to the initiative. Pricing reflects organizational size and capacity to contribute rather than level of influence, commercial control, or engagement.
The two-year roadmap is structured around four phases:
- Months 1-6 (Foundation): OPAI-aligned project charter, AI System Risk File schema, risk classification approach, autonomy guidance, onboarding, and stakeholder listening sessions.
- Months 7-12 (Build and Pilot): Test protocols, AI System Registry prototype, open-source assessment components, monitoring runbooks, and pilot evaluations.
- Months 13-18 (Scale and Evidence): Publication of SAFERai.power v1.0, safety and reliability case templates, evidence guidance, and an onboarding toolkit.
- 2028 Target (Toolkit Release): Release of the full open-source toolkit and transition to a long-term framework stewardship model.
Not in the broad, general sense. SAFERai.power is an operational risk assessment framework, evidence-package model, and open-source tool-building initiative for AI deployed in electric-sector contexts. It supports governance programs by supplying the sector-specific evidence layer they need.
Those frameworks and requirements are important reference points. SAFERai.power does not replace them. It translates broad lifecycle, safety, cybersecurity, reliability, and management-system concepts into electric-sector use cases, evidence templates, assessment methods, and operational guardrails.
OPAI convenes the ecosystem around shared datasets, open-source libraries, AI accelerators, sandbox evaluation, use-case work groups, and implementation partnering. SAFERai.power adds the complementary operational risk framework and open-source assessment tools for evaluating AI assets and use cases before deployment.
The AI System Risk File is the minimum common evidence package for a material AI use case. It covers intended purpose, data provenance, capabilities and limitations, autonomy and oversight, TEVV results, deployment conditions, monitoring, incidents, and reassessment triggers.
Founding members receive working group input, direct engagement with EPRI technical leadership, early access to framework artifacts, participation in pilot use-case development, access to reference implementations and open-source toolkit components ahead of public release, roadmap reviews, and founding member recognition at major milestones.
Yes. Tier differences reflect organizational scale and capacity to contribute. They do not buy commercial advantage over framework design, and they do not create a separate level of access to the substantive work.
Yes. Participation can be paid as a single upfront payment for the full two-year commitment or as two payments across the two-year project.
Yes. SAFERai.power qualifies for Self Direct Funds (SDF), and SDF funds can be used to pay for participation in the project.
Yes. Each participating organization is expected to contribute approximately 80 hours of skilled SME time. Contributions may come from operational domains such as grid operations, planning, asset management, customer operations, field work, vegetation management, and OT environments, or from digital and AI domains such as AI lifecycle, open-source tooling, testing, evaluation, agentic systems, and risk assessment automation.
A technology partner product can be appropriate in principle but still be deployed unsafely in a specific utility environment. Product evidence addresses the provider’s claims, product version, intended purpose, limitations, validation, and support commitments. Deployment evidence addresses local permissions, integrations, procedures, training, oversight, monitoring, and incident response.
No. The toolkit is intended to help collect evidence, route questions, draft documentation, manage registry-backed workflows, and trigger reassessment. Human owners, reviewers, operators, and domain experts retain authority for consequential decisions.
Autonomy is treated as a governed capability, not a feature toggle. The framework considers advisory systems, supervised action, guarded autonomy, and exceptional autonomous use, with stronger safeguards as consequence, system access, write authority, and operational reach increase.
Yes. The Small Scale tier is designed for smaller utilities, cooperatives, public power districts, municipal utilities, and early-stage technology companies. The framework is also intended to produce templates and tools usable by organizations without large internal AI programs.
Government and nonprofit regulatory, compliance, standards, and policy organizations that cannot otherwise join EPRI’s collaborative may be able to participate through EPRIs APEX pathway on a case-by-case basis.
Start with an initial briefing. The EPRI SAFERai.power team can discuss fit, participation tier, fee structure, potential use cases, SME contribution profile, and working group onboarding.
— Contact Us
Help define trusted AI adoption for the electric sector.
Schedule an initial briefing to discuss the initiative, OPAI alignment, participation tier, and the contribution profile that fits your organization.