Building Trust in AI on the Grid

Deck: 

From Experimentation to Deployment

Fortnightly Magazine - November 2026

Most utilities are still early in their AI journey. A 2025 EPRI survey of more than 50 organizations found that while 91 percent have experimented with generative AI, only seven percent reported advanced or expert-level maturity. Nearly half remain in the early stages of adoption.

That gap between experimentation and deployment tells an important story. The biggest challenge facing the industry today is not access to AI. Rather, it is a combination of data readiness, trust, experience, and capability. However, without trust, none of the other factors matter.

Utilities already use AI in forecasting, asset management, wildfire mitigation, and knowledge retrieval, among many other applications. A key question that remains unanswered is what happens as these systems gain greater autonomy.

As AI systems move from information retrieval and decision support toward greater operational independence, utilities, regulators, and the public need confidence not only that models can perform a task, but also that they will operate within defined boundaries and identify when to make decisions or flag specific items for human review.

Using the latest model does not necessarily lead to the best outcomes. Trust must be earned through training, testing/benchmarking, governance, and ongoing evaluation and retraining.

AI models are trained on historical data, so they generally perform well in circumstances they have seen before; however, the grid is highly complex and constantly changing. New generation sources, loads, equipment, outages, and maintenance frequently create scenarios a model may never have seen. This challenge, often described as model drift, can degrade performance as conditions move outside the training data.

This may help explain why organizations can be slow to adopt AI projects, especially those with significant autonomous components. Autonomous grid operations, advanced control room applications, and AI-enabled field technologies carry significant operational and safety risks and impacts. Utilities operate critical infrastructure and need confidence that the systems perform accurately and consistently before they deploy these technologies at scale.

Trust is central to AI deployment, which is supported by demonstrated performance and appropriate governance. Performance can be evaluated and documented through testing/benchmarking, operational evaluation, ongoing validation, and retraining as part of an overall governance program. Building confidence requires evidence that AI systems behave appropriately in known and new situations and respond appropriately to uncertainty or conflicting objectives.

Effective governance includes providing models and agents with appropriate levels of access, monitoring actions through independent systems, and ensuring that human operators can intervene when necessary. Placing a human in the process is not enough if that person lacks the visibility, authority, or time needed to exercise meaningful oversight.

From Pilots to Solutions: Practical Guidance

The important question to address is how utilities can effectively turn successful pilots into trusted, scalable solutions. Resources such as EPRI’s AI Playbook provide practical guidance for deployment, helping organizations take the next step in hosting and deploying models to translate early successes into successful deployments.

Plus, initiatives such as EPRI’s SAFERai.power, a recently launched collaborative initiative that will equip utilities and technology providers with the tools, evidence, and guardrails needed to scale AI responsibly, are helping the industry tackle this challenge. The goal is to ensure that AI systems perform as expected, have appropriate controls and oversight that are proportional to the level of risk, and support safe and reliable power for society.

What’s more, EPRI joined NVIDIA and more than 100 organizations to advance the new Open Agent Safety Platform for safer, more accountable AI agents. The platform provides continuous oversight to help organizations monitor AI behavior, enforce policies, and reduce risk.

Ultimately, future adoption of AI on the grid will accelerate as utilities and technology providers establish evidence that the systems are safe, controllable, and reliable in operation.