Communication drives 54% of customer satisfaction during power outages

When a residential block loses power, two neighbors can end up with very different days. Within minutes, one gets a text explaining that a line is down on their feeder, that crews are on the way and that power should be back by 6pm. If that estimate changes, a follow-up arrives. The other neighbor gets silence, or a message that says only "we're aware of an outage in your area." Scott Perl and Stephanie Trottier describe this scenario in an Oct. 20 T&D World piece. Both homes get their lights back at the same moment. For one neighbor, the day is a mild inconvenience. The other calls three times and gets more frustrated with each call.

The outage is the same, and so is the restoration time. What customers hear during an outage shapes their experience as much as the service the utility actually delivers. According to McKinsey's analysis, 54 percent of outage-related satisfaction comes down to communication. T&D World reports that agentic AI gives utilities a new way to bring operational data to customers at scale.

The information that separates those two days sits on the operations side, in systems the T&D organization owns. That moves the central question away from the call center and onto the grid side: can what operations knows reach each affected customer quickly and accurately enough to make a difference?

More channels, lower scores

Utilities have put plenty of effort into the customer side. For the past decade they have added interactive voice response, mobile apps and chatbots, which by T&D World's account has done little to help. The authors cite McKinsey's 2025 survey of more than 10,000 utility customers, which showed an 11-point drop since 2018 in the share calling themselves very satisfied. Over the same period, the share reporting dissatisfaction climbed from 18 to 29 percent.

The authors acknowledge that these tools may make the utility easier for customers to reach. But getting through is one problem, and having something useful to tell them is a separate one. A call center, however advanced, cannot come up with a helpful outage message by itself.

A representative who picks up a call has no way of their own to check whether a specific feeder has a downed line or which homes have lost power. The T&D World authors note that those facts sit in operations: the outage management system (OMS) produces the restoration estimate, and smart meters confirm which premises are actually without service. If the call center has no live connection to those systems, it ends up relying on a generic update, the same "we're aware of an outage in your area" line the second neighbor got.

The facts that make an outage text useful already exist inside the utility; they just live in the OMS and the meters instead of the call center. For utilities that track digital progress by channel count or app downloads, the survey suggests a more telling benchmark: whether the customer-facing side can see the same restoration estimate the control room sees.

The gap is organizational as much as technical. In a sponsored Utility Dive piece, West Monroe partner Sandy Simon writes that years of investment in AMI, sensors, data platforms, distribution automation and AI pilots have widened visibility across operations, and that the challenge now is turning that data into faster, more coordinated decisions. Simon adds that many utilities still run customer programs and grid operations as separate initiatives, which limits their ability to optimize across the enterprise. Outage messaging sits right on that seam.

The storm-day loop an agent can run

Moving operations data to the right customer at the right time has always been a manual job, and it is difficult to keep up in a storm. A person must spot the outage, identify the affected customers, choose when an update goes out, draft the message and send it. That has to happen for thousands of customers simultaneously while the operations team is focused on restoring power. Depending on how an agent is designed and governed, it can:

1. Monitor the OMS feed for new outages
2. Identify the affected meters
3. Pull the restoration estimate
4. Write a message specific to that outage
5. Send it on each customer's preferred channel
6. Follow up when the estimate changes or service returns

The agent does this for every affected customer at the same moment. The T&D World authors call it a significant step up from a chatbot. The difference is where the work starts: a chatbot answers whoever shows up, while this loop starts from an OMS event and reaches the customer before they think to call.

Billing follows the same pattern. In a partner piece for Enlit, Bidgely president and CRO Gautam Aggarwal describes agents that answer a customer asking why their bill jumped €40 by finding the actual driver, such as a heat pump running longer during a cold snap or an unoptimized EV charging schedule, rather than sending a PDF of energy-saving tips. When a call does reach a person, the agent works as a co-pilot. It analyzes meter data, runs tariff simulations and hands the representative a recommendation as soon as the call connects.

Aggarwal stresses that chat, voice IVR and the human representative all draw on the same core analytical intelligence. Outage texts and bill explanations work the same way. In both cases the interface is the last step. The value comes from the meter data underneath it. For a utility deciding where to invest first, that suggests connecting to the OMS and meter feeds pays off on every channel at once, rather than being rebuilt for each new front end.

Who owns what the agent says

The West Monroe piece argues that governance has to come before scale, with AI held to the same tests as any other change: reliability, safety, affordability, regulatory compliance and public trust, and people staying responsible for oversight. Simon calls that governance a core leadership responsibility rather than a technology exercise.

People building agents for contact centers put it more bluntly. "The business process owner owns the outcome. Full stop," Michael Bevilacqua, vice president of AI product management at Adeptia, told No Jitter's Nathan Eddy, calling the agent an instrument and accountability something that cannot be delegated to software. Drata co-founder and CEO Adam Markowitz lists what an organization needs to know about any agent: its owner, identity, permissions, policy and decision trail.

Bevilacqua's prescription is a scoped identity, with an agent able to act on its own inside its scope and structurally unable to act outside it. An outage agent's scope could be drawn tightly: read the OMS and meter feeds, message affected customers. Ownership is harder, because the loop would span two process owners. Customer operations owns the message. Grid operations owns the restoration estimate inside it.

The weak spot in an outage text is the restoration estimate itself. An agent that faithfully relays a shaky estimate to thousands of homes could create the repeat calls the whole setup is meant to prevent. Speed of delivery only helps if the number being delivered holds up.

Why outage messaging suits today's agents

Grid planning offers a useful comparison. AWS says utilities are being asked to run more interconnection studies, faster, while engineering capacity has not kept pace. Its Agentic Grid Planning program, developed with Duke Energy and managed by AWS, puts agents to work on interconnection studies using the simulation software, grid models, scripts and engineering standards utilities already rely on.

AWS says the agents finish study tasks up to 4x as fast as manual analysis and cut engineer-active time by up to 92 percent. Engineers still make the final decisions, and each step goes into a versioned, reviewable record. AWS's pitch rests on the idea that much of a study, such as preparing data, moving it between tools and assembling reports, needs engineering oversight more than continuous engineering judgment.

Harder planning work is a different matter. Researchers at Oak Ridge National Laboratory and the University of Sarajevo replicated the current state of the art in agentic AI for power systems planning and, in an arXiv preprint, tested it on nodal planning problems spanning six levels of task complexity and four grid scales. Only the two lowest complexity levels were solvable, and only on some grid sizes. The authors call for stricter testing protocols and reproducible benchmarks before anyone judges these systems ready for operations.

Their motivation is the same interconnection crunch AWS is selling into. Citing S&P Global Market Intelligence, they note that data centers have added 43.8 GW of U.S. utility load over the past five years, with another 114.2 GW planned for 2026 through 2030. Demand for automation is real. Readiness for complex planning, by their tests, is not there yet.

Read together, these results suggest agents do best today on procedural, repeatable work wrapped around trusted systems of record. The outage loop fits that description closely: read the OMS, match the meters, write, send, repeat. The difficult engineering judgment, producing the restoration estimate, stays with existing systems and the people who run them.

So the question that decides whether this works sits with operations. Is the OMS restoration estimate good enough to send automatically to every affected customer, and does it refresh fast enough to keep up when crews learn more? A utility that can say yes already has most of what the agent needs.