Skip to content

AI in Utilities: Where It Actually Lands in 2026

AvanSaber Research 9 min read

Search for “AI in utilities” and you get two kinds of content: vendor promises and academic surveys. This page is neither. It is a map of where AI has concretely landed in utility operations as of July 2026, organized by function, with named deployments and sources, plus an honest flag on what is still a pilot. Each section links to a deeper article on this site, so treat this as the hub.

Key Takeaways

“The pattern I trust is AI next to the system of record, never in place of it. When a vendor demo cannot show you the audit trail for a bill adjustment, you are not looking at a production capability. You are looking at a roadmap slide.”

Nikhil Jathar, founder of AvanSaber

Billing and the CIS: Exception Handling and Bill-Shock Prediction

Billing is where AI meets the utility’s system of record, and it is where the most scaled deployments live. The pattern across every major vendor is the same: the CIS stays authoritative, and the AI works in front of it or beside it.

On the SAP side, the Utilities Customer Self-Service Agent went generally available in Q4 2025, handling inbound account and billing inquiries against IS-U and FI-CA data. We cover the wider agent roadmap in our review of SAP Joule for utilities. At the mid-market end, Cayenta’s Cayla agent went live at the City of Lakeland, Florida in 2026, which matters because a municipal utility running this in production is stronger evidence than any enterprise pilot.

Bill-shock prediction is further along than most buyers realize. Oracle’s Opower Proactive Alerts identify customers trending toward an unusually high bill and message them before it arrives; Oracle’s product documentation credits high bill alerts with 0.3% incremental energy savings on top of the 1.5% from Home Energy Reports, and a 9% reduction in high-bill calls. Fewer shocked customers means fewer angry calls, which is why this single capability pays for itself in contact-center load before anyone counts the energy savings.

Exception handling is the less glamorous half of billing AI and arguably the more valuable one. Every billing run produces exceptions, and the win is triaging them by revenue impact and clearing the routine ones automatically, with a person approving anything that changes a bill. Where that is heading, including collections and dunning, is the subject of our piece on agentic AI in the utility back office, and the broader billing picture is in AI in utility billing.

Customer Contact: Summarization Shipped, Dispute Handling Did Not

The production win in the contact center is agent-facing, not customer-facing. Oracle shipped AI call summarization and tagging for utility customer service in May 2025, cutting after-call work rather than replacing the agent. That is the right shape for the first deployment: the AI drafts, the human owns the interaction.

Customer-facing bots are a mixed record. A Verint survey of 1,500 consumers in 2024 found more than two-thirds reported a bad chatbot experience, with failure to understand the issue as the leading complaint. Bots do fine on outage status, balance checks, and payment arrangements with clear rules. They do badly on disputes, which are exactly the interactions that decide how a customer talks about you at a public utility commission hearing. Our deeper look at AI in utility customer service covers the agent-assist and chatbot architecture in detail.

Field Service: Inspection AI Is Real, Scheduling Claims Need References

The proven field-work AI is computer vision on inspection imagery. PG&E built its Sherlock application to let inspectors label defects in drone and helicopter photos, then trained computer-vision models on those labels to pre-screen imagery for equipment defects. Ahead of the 2025 fire season, PG&E also reported more than 630 high-definition cameras monitoring 90% of its high fire-risk territory, with AI detecting smoke and automating alerts. Those are named, dated, production programs.

AI-optimized crew scheduling and dispatch is a different story. The capability exists inside field-service suites, and the logic (skills, drive time, job urgency, weather) is sound, but public evidence of named utility deployments with measured results is thin as of July 2026. Treat scheduling optimization as a claim to verify with a reference customer of your size, not a fact to assume. The adjacent and better-evidenced area is failure prediction on assets, covered in AI predictive maintenance for utilities.

Grid Operations and Load Forecasting

Load forecasting is the oldest machine-learning problem in the industry, and it is getting formal infrastructure. In March 2025, EPRI, NVIDIA, and Articul8 launched the Open Power AI Consortium to build domain-specific AI models for the power sector, with an advisory committee drawn from more than 20 energy companies including Duke Energy, Exelon, and PG&E. That consortium exists largely because data-center load growth broke the old forecasting assumptions and utilities need better tools for it.

On the operations side, the strongest quantified reference is E.ON, which reported a 77% reduction in IT downtime over five years while deploying SAP S/4HANA-based AI for predictive maintenance and customer service, per June 2026 reporting. That is a sustained program result, not a pilot readout.

One honesty note buyers should hold onto: much of what gets marketed as grid AI is deterministic automation. Duke Energy’s self-healing technology in Florida avoided more than 950,000 extended outages and saved nearly 6.3 million hours of outage time since January 2024, which is a genuinely impressive result achieved with automated fault detection and rerouting, not machine learning. Both matter. They are different purchases with different risk profiles. For the demand-side analytics half of this topic, see AI for energy consumption analysis.

Outage Prediction: Machine Learning Before the Storm

Predicting outages before weather hits is one of the cleanest utility AI use cases, because the training data (past storms, past damage, infrastructure and vegetation data) already exists. The reference deployment is the UConn Outage Prediction Model built with Eversource, which runs ahead of storms to estimate outage counts across the service territory so crews and mutual aid are staged before the first call. The technology has since moved beyond one utility: Schneider Electric partnered with UConn and Eversource to fold the model into its commercial weather platform for other utilities.

The buyer question here is validation: any storm model should be backtested against your own territory’s outage history, not a national average, before its forecasts drive staging decisions that cost real money.

Meter Data: AMI Analytics and AI at the Grid Edge

The United States has about 119 million AMI meters installed, roughly 72% of all electric meters, per EIA’s most recent count. That interval data is the raw material for most of the analytics above, and two production developments stand out.

First, anomaly detection inside the meter-data platform: Oracle added AI anomaly detection to its Meter Data Management product in June 2025, aimed at energy-loss detection and faulty-meter identification. Second, AI is moving into the meter itself: Utilidata’s Karman platform, built on an NVIDIA edge module, raised $60.3 million in April 2025 to scale distributed AI across energy infrastructure, and Department of Energy grid awards are funding deployments of more than 150,000 Karman units at Portland General Electric, ComEd, and Duquesne Light. Meter-embedded AI is early, but it is funded, named, and shipping rather than conceptual.

What most utilities actually need first is duller: getting AMI, CIS, and GIS data clean and joined enough for any model to use. That groundwork is the subject of utility data management with AI.

Production Versus Pilot at a Glance

AreaIn production todayStill pilot-stageFirst question to ask
CIS and billingSelf-service agents (SAP GA Q4 2025, Cayla live), high-bill alerts (Opower)Autonomous bill correction without human approvalWhere does the human approve a bill-changing action?
Customer contactCall summarization (Oracle, May 2025), agent assistBots resolving billing disputesWhat happens on escalation, and how fast?
Field workInspection computer vision, AI smoke detection (PG&E)AI-optimized crew scheduling at utility scaleWhich utility my size runs this, and what changed?
Grid ops and forecastingML load forecasting, EPRI consortium models emergingAutonomous grid control, grid-specific LLMsWas the model validated on our load shape?
Outage predictionStorm outage models (UConn and Eversource)Real-time restoration optimizationBacktested against our outage history?
Meter data (AMI)MDM anomaly detection (Oracle, June 2025)Meter-embedded AI at full scale (Karman rollouts)Is our AMI data clean enough to train on?

What Is Still Pilot-Stage

Three things stay firmly in the pilot column as of July 2026. Fully agentic workflows that chain actions across CIS, finance, and field systems are being showcased, not routinely operated; SAP’s autonomous enterprise material names RWE in its offshore wind context, and that should be read as a directional signal rather than a deployment template. Grid-specific foundation models from the EPRI consortium are in early access, not in control rooms. And meter-embedded AI is at the start of its rollout curve, with the named deployments above still being installed.

“Utilities do not get credit for being early. They get credit for being right. Start where a wrong answer costs you a confused customer, and earn your way toward the places where a wrong answer costs you a feeder.”

Nikhil Jathar, founder of AvanSaber

What to Ask Before Buying

Every area above collapses into the same five questions. Is the capability generally available or a roadmap item, with a date either way? Can the vendor name a reference utility of your size running it in production? How was the model validated against your kind of data, your tariffs, your storms, your load shape? Where does a human approve any action that touches a bill or the grid? And what audit trail does it leave for the regulator who will eventually ask?

How the three major CIS vendors answer those questions today is laid out in our SAP vs Oracle vs Cayenta AI comparison, and the wider evaluation framework, AI included, is in the guide to choosing the right utilities software. For the full set of articles behind this map, the AI in utilities topic hub lists them in reading order.

Frequently asked questions

Where is AI in utilities already in production?

Customer engagement and billing lead. Oracle reported approximately 45 million households enrolled in Opower AI-driven programs as of April 2026, SAP's Utilities Customer Self-Service Agent reached general availability in Q4 2025, and Cayenta's Cayla agent went live at a Florida municipal utility in 2026. Machine-learning outage prediction (the UConn and Eversource model), AI meter-data anomaly detection (Oracle, June 2025), and computer-vision inspection programs (PG&E) are also operational.

What is the difference between AI and grid automation like self-healing networks?

Self-healing distribution systems, such as the one Duke Energy runs in Florida, detect a fault and reroute power using deterministic automation rules, not machine learning. They deliver large, measurable reliability gains, but they are automation. AI in the strict sense (models trained on data that make predictions) shows up in load forecasting, storm outage prediction, meter anomaly detection, and customer analytics. Buyers should keep the two categories separate when evaluating vendor claims.

Should a utility buy AI as a product or as a feature of its CIS and ERP?

As of 2026, almost all production AI for utilities ships as a layer on an existing system of record: Joule on SAP, Opower and CC&B AI on Oracle, Cayla on Cayenta. No AI-native utility CIS exists yet. The practical question is not which AI product to buy but how well your platform vendor's AI layer integrates with the data you already have.

What should a utility ask a vendor before buying an AI capability?

Five things: whether the capability is generally available or a roadmap item, a named reference utility of similar size running it in production, how the model was validated against your kind of data, where the human approval sits for any action that touches a bill or the grid, and what audit trail the system produces for regulators.

Related reading