Like many energy and utility companies today, Xcel Energy is managing an aging grid alongside unprecedented demand growth and more frequent and severe weather events. Area engineers were left piecing together reliability signals (KPIs, outage history, work orders, and weather data) across disconnected systems, making it difficult to know where to focus limited infrastructure resilience spend.

Lovelytics built 3RE (a version of their Gridlytics accelerator), Xcel Energy’s first production Databricks Application to natively embed agentic AI. 3RE centralizes operational, outage, and weather data, then layers in machine learning risk scoring and an AI agent, powered by Anthropic’s Claude models, that synthesizes root causes, historical work, and engineer notes into automated repair recommendations.

Every day, Xcel Energy’s distribution operations team keeps the lights on for customers across 8 states and 3 interconnections. Their work spans everything from initial outage response to system expansion and reliability enhancements, all measured against core operational metrics like Customer Minutes Out (CMO) and Customers Experiencing Multiple Interruptions (CEMI).

But turning those metrics into a clear plan for where to invest a fixed annual resilience budget meant manually cross-referencing SAP work orders, outage logs, and weather data, one system at a time.

That’s where Lovelytics came in.

One Platform for Outage, Weather, and Operational Data

Rather than asking Xcel’s engineers to work harder across the same scattered tools, we set out to give them one place to work from. In partnership with Xcel data scientists and area engineers, we built 3RE on Databricks, an application that centralizes operational, outage, and weather data in a single governed environment.

For the first time, area engineers had one place to investigate CEMI & DEMI events, spot emerging hot spots, and generate the service notifications that can be fed directly into Xcel’s SAP work order pipeline.

With that foundation in place, Xcel didn’t stop at descriptive reporting. The team layered in a machine learning model that produces a 14-day predictive risk score by combining historical performance with weather forecasts, including known stress points like equipment failure risk above 85°F, when conductors and insulators are most prone to expansion and contraction damage.

Now, instead of simply flagging known temperature issues, 3RE helps engineers pinpoint the specific areas of the grid that are disproportionately at risk before problems occur.
The newest chapter of the story is agentic AI.

Rather than requiring engineers to write complex queries, 3RE’s AI assistant, powered by Anthropic’s Claude models, answers natural-language questions like “Why does this feeder have recurring spikes?” grounded in governed, verified data.

A recommendation engine, similar in spirit to how a streaming service suggests what to watch next, performs causal analysis across CEMI, DEMI, and CELI events. Its newest capability reads unstructured work order and field-report text to generate automated repair recommendations, so engineers spend less time digging and more time acting.

None of this would matter without trust, so Lovelytics built 3RE on a governed foundation from day one. Unity Catalog ensures the AI only ever touches tables an engineer is already authorized to see. Databricks Genie powers natural-language querying, Databricks Apps handles production hosting, and MLflow logs every AI response for full auditability.

The Databricks AI Gateway provides secure, governed access to foundation models (including Anthropic’s Claude) all without a single byte of operational data leaving Xcel’s own environment. The result is a tool Xcel’s engineers can rely on every day.

Xcel Energy needed a partner who could move at the pace of its distribution engineers while still building on a governed, enterprise-grade foundation. Lovelytics had already earned that trust through the TD Migration, ISP, and CETS engagements, working daily alongside Xcel’s IT and AI teams as an extension of their own staff rather than an outside vendor.

For 3RE, that trust extended to Lovelytics’ hands-on expertise pairing Databricks with Anthropic’s Claude models — expertise Xcel had already seen pay off elsewhere across its AI portfolio. Lovelytics combined that platform depth with deep utility domain knowledge, translating complex reliability metrics like CEMI and DEMI into an agentic experience that area engineers could trust and adopt quickly.

Delivering an accurate risk score turned out to be only half the challenge — getting engineers to act on it was the other half. Early versions of the predictive model produced solid rankings, but area engineers were hesitant to reprioritize field work without understanding why a given feeder was flagged.

Lovelytics addressed this by exposing model feature importance directly in the 3RE interface, pairing every risk score with the specific weather, asset, and outage-history factors driving it, then layering the agentic assistant on top so engineers could simply ask why a feeder was flagged rather than dig for the answer themselves.

That shift, from a black-box score to a reasoning-forward recommendation, is what took 3RE from a reporting tool to something engineers actually incorporate into daily prioritization. It’s also the lesson Lovelytics carried into scoping the next phases of 3RE: governed, agentic AI gets adopted fastest when it shows its reasoning, not just its output.

3RE replaced hours of manual, cross-system investigation with an AI-accelerated workflow that gives Xcel’s area engineers a single, governed view of grid reliability, from descriptive KPIs to predictive risk to automated repair recommendations.

• $2.5M–$3.4M in estimated annual value from AI-assisted, risk-based capital prioritization.
• Centralized reliability view spanning 8 states and 3 interconnections.
• 14-day predictive risk scoring to get ahead of weather-driven equipment stress.
• Automated repair recommendations generated from unstructured work order and field-report data.
• Fully governed and auditable AI, powered by Anthropic’s Claude — every response traceable through MLflow, scoped by Unity Catalog permissions.

Xcel and Lovelytics are already scoping Phase 4 of 3RE (Gridlytics), set to kick off later this year. The next phase will extend predictive risk scoring to additional asset classes and data sources, including wildfire and vegetation signals, and deepen the agentic recommendation engine’s ability to generate work-order-ready repair guidance directly from field data. Lovelytics remains embedded with Xcel’s distribution operations team to support this next chapter of the platform.