Scaling AI pilots into utility-wide operations

# Scaling AI pilots into utility-wide operations

A large US investor-owned utility once ran an 18-month AI vegetation-management pilot across one service territory. The model used satellite and LiDAR (Light Detection and Ranging, a laser-based remote sensing method) imagery to predict which trees were most likely to fall into power lines and cause outages or wildfires. Trim crews, guided by the model instead of fixed schedules, cut vegetation-caused outages in the pilot area by a reported one-third. Leadership called it a success. Two years later, it was still running in that same single territory. It never scaled utility-wide.

This lesson uses that stall pattern to unpack why utilities struggle to move AI from pilot to production, and what "good" scale-up actually requires.

Why pilots succeed and rollouts stall

Pilots are built to succeed. They get a dedicated data science team, a cooperative regional manager, clean-ish data from one geography, and a willingness to tolerate manual workarounds. None of that survives contact with 10 or 20 operating regions.

Common failure points:

• Data fragmentation. The pilot region had recent LiDAR flights and a modern GIS (Geographic Information System) database. Other territories had paper records, outdated asset inventories, or vegetation data from different vendors in incompatible formats.
• Workforce and union agreements. Trim crew scheduling in many utilities is governed by collective bargaining agreements and established maintenance contracts. Shifting from calendar-based trimming to AI-prioritized trimming can require renegotiating work rules, not just retraining staff.
• Regulatory approval per jurisdiction. Vegetation management spending is typically recovered through rate cases reviewed by state public utility commissions (PUCs) in the US, or by national regulators in Europe (such as Ofgem in the UK). A pilot funded through an innovation budget faces a different approval bar than a permanent, capitalized, utility-wide program that regulators must approve for cost recovery.
• Liability and audit trail. If an AI model deprioritizes a tree that later falls and causes a wildfire, the utility needs a defensible, documented rationale. Pilots rarely have this rigor built in; production systems must.

This is not an AI modeling problem. The pilot's predictions were accurate. It is an organizational, contractual, and regulatory integration problem.

Where AI genuinely fits in utility operations

Before scaling anything, it helps to be clear-eyed about where AI adds real value across the utility value chain:

• Generation and grid operations: load forecasting, renewable output forecasting (wind and solar), predictive maintenance on turbines and transformers.
• Transmission and distribution: vegetation management (as above), fault location, dynamic line rating, outage prediction.
• Wildfire risk: combining weather, vegetation, and grid sensor data to trigger Public Safety Power Shutoffs (PSPS), a practice used by utilities like PG&E in California.
• Customer operations: demand response targeting, churn prediction in deregulated retail markets, chatbot-driven billing support.
• Trading and energy markets: short-term price forecasting for utilities that participate in wholesale markets like PJM or ERCOT in the US, or day-ahead markets under EU electricity market rules.

Not every use case deserves the same scale-up urgency. A chatbot pilot that underperforms costs little. A vegetation-management model that misprioritizes trimming in a wildfire-prone region carries safety and legal exposure. Scale-up planning should be proportional to risk, not just to pilot performance metrics.

A framework for evaluating scale-readiness

Before greenlighting utility-wide rollout, run the pilot through five questions:

1. Data portability. Can the model run on data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → found in your worst-instrumented region, not just your best?

2. Regulatory pathway. Has this been discussed with the relevant PUC or regulator, and is there a plan for cost recovery and safety justification?

3. Workforce integration. Does the union contract, dispatch software, and crew training plan account for the new workflow?

4. Vendor and infrastructure dependency. Is the pilot tied to one vendor's proprietary platform, or can it be re-procured and integrated with existing enterprise systems (GIS, outage management systems, SCADA, Supervisory Control and Data Acquisition)?

5. Governance and audit trail. Is there a documented process for explaining and defending model decisions if challenged by a regulator, insurer, or court?

A useful public reference for structuring AI risk and governance thinking, applicable well beyond utilities, is the NIST AI Risk Management Framework, a free framework from the US National Institute of Standards and Technology.

A simple worked example: rollout cost versus benefit

Assume (illustrative estimates, not vendor-quoted figures):

• Pilot territory: 5,000 miles of distribution line, cost of AI vegetation platform and integration: $2 million (estimate).
• Reported outage reduction: 30% fewer vegetation-caused outages, saving an estimated $3 million/year in avoided restoration costs and reliability penalties (estimate).
• Utility-wide footprint: 50,000 miles (10x the pilot).

Naive scale-up math: 10x the cost ($20 million) for 10x the benefit ($30 million/year) looks attractive. But this assumes uniform data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète →, uniform crew integration, and uniform regulatory approval, none of which hold. Realistic utility-wide rollouts commonly see:

• 30 to 50% cost overrun from data remediation and system integration (a pattern documented across large utility IT programs, treat as directional, not a fixed rule).
• Multi-year phased regulatory approval rather than one-time sign-off, since many state PUCs review capital programs incrementally across rate cases.

So the realistic year-1 rollout cost might be closer to $26 to $30 million, with benefits phased in over 3 to 5 years as territories come online one PUC filing at a time. The ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → (return on investmentreturn on investmentReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète →) story is still likely positive, but the timeline and cash flow profile look very different from the pilot's clean 12-month payback.

What "good" scale-up looks like

Utilities that have scaled AI vegetation and asset-inspection programs successfully (examples include elements of Xcel Energy's and Duke Energy's grid modernization programs, both publicly discussed in company sustainability and grid modernization reports) tend to share traits:

• They fund data standardization (common GIS schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.Voir la définition complète →, consistent LiDAR refresh cycles) as its own line item, before scale-up, not during it.
• They engage regulators early, framing the AI system as a safety and reliability investment, not just an efficiency tool.
• They pilot the workforce change (dispatch software, crew training) in parallel with the model, not after model validation.
• They pick 2 to 3 second-wave territories with deliberately worse data than the pilot, to stress-test before full rollout.

🎬 [VIDEO: "How AI Is Transforming the Electric Grid" - youtube.com - a utility-industry explainer on grid AI applications including predictive maintenance and vegetation management, useful for a non-technical overview]

A note on measurement discipline

One underrated scale-up killer: metrics that don't travel. A pilot might report "30% fewer outages" without specifying the baseline period, weather-normalization method, or whether savings are gross or net of the platform's cost. When rolling out utility-wide, insist on:

Reported metric checklist:
- Baseline period and weather normalization method stated?
- Gross savings vs. net of implementation/maintenance cost?
- Confidence interval or sample size for the outage reduction claim?
- Comparable metric available for non-pilot (control) territories?

Without this, executives approve rollout budgets based on numbers that can't survive scrutiny from a regulator's technical staff or an internal audit.

Key Takeaways

• A successful AI pilot proves the model works on good data in a friendly environment; it does not prove the organization can absorb it at scale.
• The main scale-up barriers in utilities are usually data fragmentation, union and workforce integration, and regulatory cost-recovery approval, not AI model performance.
• Evaluate scale-readiness with five questions: data portability, regulatory pathway, workforce integration, vendor lock-in, and governance/audit trail.
• Realistic rollout ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → () should assume cost overruns of 30 to 50% and multi-year phased regulatory approval, not linear scaling from pilot economics.