For years, the conversation around artificial intelligence in utilities has largely focused on what AI can help us understand. Can it predict asset failures earlier, improve load forecasting, identify anomalies faster, or help operators navigate an increasing volume of alarms, events and system information?
These capabilities are becoming increasingly relevant as utilities modernize their operations. But as AI evolves from analyzing information to recommending and potentially executing actions, I believe utility leaders will face a more consequential question:
How much decision-making authority should we delegate to AI and under what conditions?
Utilities have relied on automation for decades. What makes this next phase different is the increasing ability of AI systems to interpret complex information, evaluate alternatives and influence decisions that were traditionally dependent on human judgment. The challenge is no longer simply about what AI can do. It is about deciding what AI should be allowed to do. In a reliability-critical industry, that distinction matters.
We Are Moving Beyond Decision Support
Utilities have used analytics, optimization and automation for decades. AI itself is not entirely new to the industry either. What is changing is the breadth of information AI can process, the speed at which it can evaluate alternatives and, increasingly, its ability to plan and execute multiple steps across connected systems. AI's emerging role in electricity networks spans forecasting, detection, diagnosis, prioritization, simulation and optimization. Many of these capabilities are already being explored or adopted in planning, asset management and maintenance, where people remain firmly within the decision process.
Real-time operational control is different.
The closer AI moves toward actions that can directly affect grid operations, the higher the expectations become around explainability, validation, cybersecurity, fallback mechanisms and accountability. This suggests we should stop thinking about AI adoption as simply a choice between human or machine. Instead, there is a spectrum of AI involvement and delegated authority:
Observe → Recommend → Assist → Act with Approval → Act Autonomously
This is not necessarily a maturity ladder that every AI use case must climb. Rather, it represents different levels of involvement that may be appropriate depending on the operational context.
The leadership challenge is deciding where different utility decisions should sit on that spectrum.
Not Every Decision Carries the Same Risk
Consider three examples.
Using AI to prioritize thousands of inspection images so engineers can focus on assets most likely to require attention generally carries relatively limited operational risk, particularly when recommendations are reviewed before action.
Using AI to recommend which transformer should receive maintenance priority introduces greater operational and financial consequences, but an engineer can still evaluate the recommendation before deciding.
Using AI to initiate a switching action on the power system is fundamentally different. An incorrect action could have immediate consequences for reliability, equipment and personnel safety.
The consequence of being wrong increases significantly, as does the importance of understanding why the system reached its conclusion. It is also important to recognize that autonomy itself is not new to utility operations. Protection relays, automatic generation control and other established control systems already execute actions without waiting for individual human approval. However, these functions operate within engineered control schemes, defined operating requirements and established validation processes.
Introducing AI-driven decision-making creates additional considerations, particularly where models are probabilistic, adaptive or dependent on changing operational context.
The question is therefore not whether utilities should embrace automation, but how the authority and assurance mechanisms developed over decades should evolve to accommodate a different form of intelligence. This is why I don't believe the goal for utilities should be maximum autonomy. It should be appropriate autonomy. The level of autonomy should reflect the consequences, reversibility, uncertainty and operational criticality associated with each decision.
We Can Already See the Direction of Travel
Public research and industry initiatives suggest that AI is progressing most readily in applications where it augments rather than replaces engineering judgment.
The International Energy Agency has highlighted AI applications across electricity networks, including forecasting, maintenance, inspection, planning and operational decision support. These use cases illustrate how AI can improve efficiency and analytical capability without necessarily taking direct control of grid operations.
Similarly, research initiatives at the U.S. Department of Energy's national laboratories have explored AI-enabled grid analysis, situational awareness and human-AI collaboration to help operators evaluate complex operating conditions and potential actions.
These developments point toward an important distinction - The technology may increasingly be capable of recommending an action, but operational acceptance depends on more than capability. It depends on trust, validation and confidence that the system will behave appropriately under both expected and unexpected conditions.
AI should initially expand the operator's field of view, not eliminate the operator from the decision. That approach creates opportunities to build experience and confidence before considering greater levels of delegated authority.
Speed Changes the Equation
There is another reason this conversation matters.
Grid complexity is increasing. The growth of distributed energy resources, battery storage, electrification, variable generation and increasingly interconnected digital systems means operators must manage a wider range of operating conditions and dependencies. In some situations, the number of possible scenarios can exceed what people can reasonably evaluate within the available time.
AI-assisted analysis and optimization can help evaluate alternatives more quickly, identify potential risks and support faster decisions.
Grid topology optimization offers one example. Computational approaches can evaluate alternative network configurations and identify options that engineers and operators can assess against operating constraints. That creates an interesting tension.
Humans bring judgment, experience, situational awareness and accountability. Machines bring computational speed, scale and the ability to evaluate large numbers of alternatives. The future operating model may therefore be less about choosing one over the other and more about designing how they work together.
The challenge is ensuring that greater speed does not come at the expense of understanding, control or operational confidence.
Trust Must Be Designed, Not Assumed
In my previous article, Context Is the Missing Layer, I argued that AI needs more than data. It needs the relationships and operational context that give that data meaning.
Decision autonomy raises the bar further.
Before allowing AI systems to participate in higher-consequence decisions, utilities will need confidence in more than model accuracy.
Can recommendations be explained and independently validated? Can decisions be reconstructed later?
Are operating limits and safety constraints explicitly enforced? What happens when data becomes unavailable, contradictory or outdated?
Can an operator intervene or override the system?
And how does the system respond when it encounters conditions outside those against which it was validated?
Perhaps most importantly, who remains accountable when a machine participates in the decision?
These questions cannot be left until after deployment. They need to influence architecture, governance, testing and operational design from the beginning.
Trust should not be treated as an outcome of successful deployment. It must be designed into the system before deployment.
Human-in-the-Loop May Also Need to Evolve
We often use the phrase human-in-the-loop as reassurance when discussing AI adoption. But simply placing a person somewhere in the workflow does not automatically create meaningful oversight.
Consider an operator receiving an AI-generated recommendation during a rapidly evolving grid event. If the operator has only seconds to review and approve it, is that genuinely human decision-making, or has the person effectively become a confirmation step?
As AI capabilities mature, utilities may need to distinguish more deliberately between different forms of human oversight. Some decisions should remain directly under human control. Others may allow AI to act within clearly defined boundaries while people supervise outcomes and retain the ability to intervene. Certain fast-response functions will continue to require deterministic automation operating at speeds beyond human reaction time.
The important question is not whether a human touches every decision. It is whether decision authority, operating boundaries, escalation mechanisms and accountability are deliberately designed.
Human oversight must be meaningful, not merely procedural.
A Different AI Roadmap
This suggests that utilities may need to evaluate AI use cases not simply by technical feasibility, potential ROI or productivity improvement, but also by the level of decision authority being considered.
A useful framework could begin with three questions:
What can AI know?
Is the underlying data accurate, timely and sufficiently contextualized? Does the system understand the relevant assets, relationships, constraints and operating conditions?
What can AI recommend?
Can its recommendations be validated, explained and trusted? Are the assumptions and uncertainties visible to the people responsible for acting on them?
What can AI do?
Which actions, if any, should it be permitted to execute? Within what boundaries, under whose authority and with what mechanisms for monitoring, intervention and recovery?
As organizations move from the first question toward the third, the required level of assurance and governance should increase according to the risk involved. This also requires a shift in how AI initiatives are governed. Technology teams may develop and validate the capability, but operational leaders must help define where decision authority begins and ends. Cybersecurity, engineering, risk and compliance functions have equally important roles in establishing the conditions under which AI can be trusted.
A successful pilot may demonstrate that AI can recommend the right action under tested conditions. It does not necessarily demonstrate that the organization is ready to delegate authority to that system in production.
Technical readiness and operational readiness are not the same thing.
Utilities should therefore approach decision autonomy as a carefully governed progression, supported by validation, operational experience and clearly defined responsibilities.
Autonomy should be earned, not assumed.
Closing Thought
As AI becomes more capable, utilities will have increasing opportunities to move beyond decision support toward greater decision autonomy.
But capability alone should never determine authority.
The electricity industry has spent generations engineering reliability, protection, redundancy and accountability into the physical grid. As AI becomes part of operational decision-making, the same discipline must extend to how authority is delegated between people and machines.
Some decisions will continue to require human judgment. Others may benefit from AI recommendations, while certain well-defined activities could eventually operate autonomously within carefully engineered boundaries.
The goal should not be to eliminate human involvement wherever technology makes it possible. It should be to design an operating model where human judgment and machine intelligence complement each other, with clear accountability for the outcomes.
The future of the intelligent grid will not be defined by how many decisions AI can make, but by how wisely utilities decide which decisions to entrust to it, because in critical infrastructure, autonomy is not simply a technological capability, It is an operational responsibility.