AI in energy is distinctly different from AI in marketing or consumer apps.
When it comes to a utility’s AI system, it’s all about the real-world aspects: physical assets, regulated processes, and operations that affect safety. This technology can transform how electricity is managed, from the flow of power and crew dispatch to how customers stay informed and how markets are adjusted. So, there are a couple of things to keep in mind. First, the potential benefits are huge—improving reliability, cutting costs, and accelerating decarbonization can really add up when you’re working at a grid scale. On the flip side, the standards for accuracy, traceability, and cyber safety are much higher than what most software teams anticipate.
This article aims to give product leaders a straightforward view—focused on reality, not hype or fear. I want to provide a clear overview of the most impactful AI solutions that utilities are actually implementing (or can implement), detailing the necessary data, what the models do in practice, common production challenges, and how to develop this with either an in-house team or an external engineering partner.
Throughout the article, I’ll look at “energy and utilities” as a broad spectrum that includes electric distribution utilities, transmission operators, integrated utilities, independent system operators, retail energy providers, and grid-adjacent services like microgrids and EV charging stations. While the leading solutions may overlap, the specific constraints and integration points differ across the board.
Before diving into the “Top 10” solutions, let’s share a common understanding of what “production” means in the context of utility AI.
In grid environments, the most challenging aspect isn’t really about creating the neural network itself. Instead, it involves navigating the complex interplay of systems: connecting SCADA, AMI, OMS, DMS/ADMS, EMS, GIS, work management, customer systems, DER platforms, and weather/market feeds to present a unified operational vision.
Additionally, you’ll have to consider the existing interoperability and security standards that will affect your architectural decisions. If you’re developing systems that interact with substations and protection/control mechanisms, you’ll encounter IEC 61850 concepts and data models. It’s crucial to understand the goals of this standard focused on interoperability and structured device models, even if you don’t directly implement the “raw IEC 61850” elements.
Whenever you exchange grid models across systems (like planning, operations, and market tools), you’ll run across references to “CIM” (Common Information Model) along with related IEC standards. For instance, European Transmission System Operators (TSOs) actively conduct interoperability tests based on CIM-related standards.
When orchestrating flexible loads and DER programs on a larger scale, you’ll engage with OpenADR (Open Automated Demand Response), which is designed to be a non-proprietary standardized interface for signal management in demand response.
If you’re operating in North America, where bulk electric system regulations apply, you’ll come across the NERC CIP, which encompasses a family of cybersecurity standards that affect auditability, access control, monitoring, and change management for systems adjacent to operational technology.
Thus, production-grade AI in utilities moves beyond a simple “model → API → app” relationship. Instead, think of it as “model → controls and decision support → integration → safety constraints → monitoring → audit trail.”
It’s essential to be clear about where the AI fits into the decision-making process.
Some AI serves an advisory role—like identifying the 20 feeders that are most at risk during an upcoming storm. Other AI systems are semi-automated, like making switching recommendations that require operator approval. Then there’s closed-loop AI that automates tasks, such as adjusting inverter settings within pre-defined limits. The closer you are to an automated process, the more critical it becomes to establish formal constraints and ensure safe fallbacks.
Remember this key takeaway: utilities don’t just deploy “AI.” They deploy “AI + guardrails + integration + accountability.”
Now, let’s explore the top solutions that can effectively manage this degree of complexity.
Load forecasting remains the highest ROI AI application in the power sector because it influences unit commitment, dispatch, hedging, congestion planning, staffing, demand response triggers, and DER scheduling. When forecasting accuracy improves, it makes all subsequent optimizations more straightforward.
However, “load” isn’t simply one curve anymore.
Modern utilities require several forecasts: system-level, substation-level, feeder-level, and sometimes even at the distribution-transformer level. They need forecasts that cover short-term (minutes to hours) to long-term (daily, weekly, and seasonal) horizons. Both “gross load” and “net load” (load minus embedded PV and other generation sources) are necessary. Furthermore, utilities require probabilistic forecasts, not just single-point estimates, to effectively manage risks.
A helpful framework is “hierarchical forecasting,” where you generate forecasts at different levels while ensuring consistency: feeder forecasts should contribute to substation forecasts, which in turn support system-wide forecasts. Research and utility programs emphasize building accurate models at transformer, feeder, and substation levels by using high-resolution data, because operational decisions are made at these levels.
In reality, the quality of load forecasting hinges more on data alignment than algorithm choice.
You usually require historical load (either AMI interval data or SCADA measurements), geographic context (like GIS data, feeder boundaries, and switching history), calendar features (holidays, school schedules), and weather data (temperature, humidity, wind, cloud cover, heat index). For net load, access to PV capacity estimates, inverter telemetry when available, and indirect signals like irradiance or cloud cover are also essential.
One tricky issue is “topology drift.” Changes in feeder configurations, tie switches, and load transfers can alter the relationship between meters and feeders. If your dataset assumes a static topology, the model may learn patterns that cease to be accurate after operational changes occur. Successful forecasting programs either (a) utilize topology-aware features and log switching events or (b) segment training windows based on stable topology periods.
Simple models often excel initially, using approaches like gradient-boosted trees for day-ahead and week-ahead forecasts with engineered features; linear models can work for baselines and weather, while deep learning can be introduced for multi-horizon and probabilistic forecasting as data quality improves. Deep learning shines when handling multiple correlated signals (like load, PV, and weather data) and providing multi-step forecasts with uncertainty. However, it alone will not resolve topology drift.
A practical production approach is to construct an understandable baseline model for operators, then layer in advanced models over it, continuously comparing their performance. Your monitoring efforts should focus on “forecast error by feeder class” rather than just an overall system-level percentage error.
Once in operation, the forecast shouldn’t be a single number. It should function as a service: emitting a forecast distribution for multiple horizons every 15 minutes, complete with confidence intervals and “feature attribution” summaries for auditing purposes (like, “temperature anomaly contributed to the forecast variance”). Plus, it’s important to establish a feedback loop—when actual data comes in, compute the error by segment and retrain regularly.
When engaging external teams, the emphasis should not merely be on whether “they can build an LSTM model.” The focus should be on “their ability to construct a data lineage pipeline capable of managing your complex meter topology and switching realities.”
As renewable energy sources become more prevalent, system operators are shifting their focus from just “total load” to “net load variability.” This aspect drives the need for ramping, reserve requirements, and managing congestion in the grid.
Forecasting for solar and wind isn’t just valuable—it’s essential for modern grid operations. The National Renewable Energy Laboratory (NREL) highlights machine learning and data analytics as critical for short-term forecasting, using production-grade simulation tools like PLEXOS to assess the system’s value of improved forecasts.
You will need to forecast generation output from both utility-scale renewable assets (such as wind and solar farms) and behind-the-meter embedded generation. Ramping (i.e., quick changes) also needs a forecast since these ramp-ups can stress operations. Different spatial scales are necessary: asset-level, zone-level, and balancing authority levels.
Moreover, forecasting “forecast error” is vital too. It may sound meta, but it’s incredibly useful as operators reserve capacity not simply for average outcomes but to account for extremes. Including calibrated uncertainty in your forecasts, it allows for reduced reserves without amplifying risk.
For wind forecasting, use data from turbine SCADA, numerical weather prediction outputs, plus terrain features and roughness. In solar forecasting, needed data can include irradiance proxies, satellite cloud movement projections, plant telemetry, and albedo effects in some areas.
For behind-the-meter PV generation, you might not have direct telemetry. This is where inferred generation methods using feeder net load patterns, PV installation databases, and irradiance proxies come into play. This overlaps with probabilistic methods and robust state estimation in forecasting.
Forecasting truly adds value when it directly impacts decision-making: whether that’s in dispatch schedules, reserve procurement, curtailment decisions, or managing DER dispatch. If you create a forecast but your scheduling tools can’t incorporate uncertainty bands, you miss out on valuable opportunities.
An effective design would involve a forecast service that provides both point and quantile forecasts along with scenario ensembles. Downstream optimizers can work with these scenarios, enabling a computation of “expected costs under uncertainty” instead of just “cost based on a single estimate.”
Distribution networks weren’t designed to manage millions of small generators and flexible loads, but that’s where we’re headed with rooftop solar, batteries, EV charging, smart thermostats, industrial flexibility, and microgrids. Utilities and aggregators need tools to monitor, predict, and coordinate these resources while maintaining voltage limits and protection measures.
This challenge falls under DERMS (Distributed Energy Resource Management Systems) and VPP (Virtual Power Plant) intelligence. Research and industry evaluations highlight DERMS as a response to the increasing adoption of distributed energy resources and the necessity of managing their value while ensuring safety and reliability.
A typical misstep among founders is viewing DER orchestration purely as a cloud scheduling issue. It’s much more than that. The distribution grid imposes local constraints: voltage limits, transformer thermal limits, phase imbalances, and the complexity of operating without complete observability at lower voltage levels.
A DERMS system must translate high-level objectives (like reducing peak demand by 20 MW) into achievable setpoints at the grid’s edge. For instance, it should guide actions like “curtail these inverters by X%,” “charge these batteries now,” or “delay these EV chargers” while respecting local operational limits.
DER telemetry can come from various sources, such as inverter APIs, aggregators, smart meters, EV charging networks, or edge gateways. Insights on grid constraints can be sourced from DMS/ADMS, GIS, power flow models, transformer ratings, and sometimes line sensors.
Interoperability is key here. You will likely be bringing together many vendors. Frameworks like OpenFMB exist primarily to tackle interoperability challenges at the grid edge as DER integration grows, making device-to-device and system-to-system communications essential.
The AI aspect here typically isn’t a singular model; rather, it often involves multiple models that forecast availability, anticipate customer responses, identify anomalous DER behavior, and carry out optimization under uncertainty.
While reinforcement learning may be suggested, real-world applications tend to favor optimization with clearly articulated constraints. That said, reinforcement learning could prove beneficial for adapting to unpredictable human responses or learning control policies using DER as flexible devices, but this would come after thorough simulation and ensuring safety is prioritized.
For this space, it’s important to demand a fully “closed-loop simulation environment” prior to field forces being automated. That means incorporating a power flow simulator, a DER behavior simulator, and a customer response model. Testing without this setup can lead to unsafe policy implementations.
You’ll also want to ensure there’s a clear auditability process. If an optimizer reduces customer DER, you must be able to explain the rationale and demonstrate adherence to program guidelines. This level of accountability is crucial in heavily regulated markets.
Solution 4: Voltage Optimization and Control Techniques
Voltage optimization is often seen as a less glamorous topic among utilities, yet when executed well, it leads to significant financial benefits.
Historically, distribution utilities maintained higher voltage levels to ensure that customer endpoints remained within acceptable limits. However, with improvements in sensing and control systems (like capacitor banks and smart inverters), it’s now feasible to operate voltage closer tooptimal levels, which helps reduce losses and peak demands while still meeting service standards. Conservation Voltage Reduction (CVR) is a proven strategy to cut demand or energy usage by adjusting feeder voltage setpoints; extensive research exists detailing its potential benefits, along with the complexities tied to quantifying the financial returns amid real-world variations.
Volt/VAR control is fundamentally rooted in power systems optimization. But AI can assist in three areas.
First is forecasting: predicting upcoming voltages and loads so utilities can act proactively. The NREL has published studies integrating state forecasting with optimal power flow techniques specifically for voltage regulation using forecasted system states.
Secondly, involves estimation. Since distribution networks often lack visibility, AI can enhance state estimation by inferring unmeasured voltages, identifying phase imbalances, and calculating losses.
Thirdly, there’s the potential for learning control policies: research is underway exploring reinforcement learning for voltage control in distribution grids with PV and storage as flex elements, particularly in uncertain contexts.
The real challenge lies in accurately measuring the actual CVR effect without falling into the trap of misleading yourself.
Load composition fluctuates with seasons and times of the day. Weather influences demand. Customer behaviors may shift. If you conduct a CVR pilot by merely comparing results before and after, you might erroneously attribute savings entirely to CVR actions. A robust analysis of CVR performance will address factors like temperature and timing patterns, utilizing statistical methods to gauge sensitivity.
In production, also ensure your systems tightly integrate with DMS/ADMS and device control systems. It’s wise to have a secure fallback policy; if sensors are unavailable or inconsistent, reverting to safe static settings is necessary.
Engineers will focus on metrics such as voltage violations, operations of regulator taps (related wear), frequency of capacitor switching, losses, and customer complaint data. Finance will be interested in energy savings and peak reduction figures, while regulators will monitor compliance. Your AI system should encapsulate all these facets, not just indicate “loss reduced.”
Outage prediction is where AI starts becoming very tangible for utilities, customers, and regulators.
The aim isn’t to magically prevent outages but to enhance preparedness: positioning crews, pre-arranging resources, prioritizing vegetation maintenance, adjusting switching procedures, and streamlining customer communication processes. Furthermore, accurately estimating outage duration (ETR) is vital, as the credibility of these estimates builds trust among customers.
Recent research showcases practical machine learning strategies in this area, including models that identify potential outage locations during severe weather based on a mix of static (infrastructure, vegetation, soil) and dynamic features (storm characteristics).
There is also a published work on real-world utility data to predict outage duration classes to enhance customer notification, which involves feature selection as well as gradient-based features and gradient boosting on historical outage datasets.
When a storm approaches, your AI system can analyze forecasts, historical outage data, asset condition indicators, and vegetation information, producing a risk map that ranks feeders and branches based on failure probabilities. It will also generate an estimate of “expected damage modes” like tree contact, pole failure, flooding risks, conductor slap, and so forth.
The operations strategy can adapt accordingly. Crews can be stationed near high-risk areas, necessary materials can be pre-loaded (such as transformers, poles, and insulators), mutual assistance can be coordinated, and switching strategies can be modified for likely sectionalization.
If you have DER and microgrid assets, the planning should also include strategies for islanding vital loads.
Often, utilities have historical outage data feasible for training models. Still, they struggle with the structured asset attributes that significantly enhance outage prediction accuracy. The “static features,” including details on equipment type, age, maintenance history, vegetation proximity, and geographical context, play a crucial role.
That’s why many outage prediction initiatives end up focusing on creating a new “asset and vegetation feature store” as their primary output.
It’s important to tread carefully: Taking proactive steps based on predicted failures can lead to consequences. If you decide to de-energize lines or shift loads following risk assessments, justifying those decisions becomes vital. Most utilities opt to keep outage prediction in an advisory capacity, rather than fully autonomous, and that is generally a sensible choice.
A mature solution will include model explanations (such as feature contributions) not to impress executives, but to help engineers build trust and debug the system.
Outage prediction focuses on “where failures might happen,” whereas fault location addresses “where the failure occurred” and “how to restore service safely.”
Distribution grids can be quite noisy, with partial sensor data and ambiguous fault currents. Traditional FLISR relies on protection devices, fault indicators, and engineering heuristics, while AI offers an enhancement by learning from sequences of events, waveform signatures (when high-resolution data is available), and correlated meter outage signals.
In reality, the best application of AI in this context often supports operator decision-making rather than full automation. The AI can provide a ranked list of probable fault locations with confidence scores and suggest switching sequences that limit customer interruptions while adhering to constraints, requiring operator approval.
You’ll combine SCADA events, feeder relay logs, fault indicator signals, AMI last-gasp messages, OMS outage reports, and GIS connectivity data. If µPMUs or line sensors are present, AI can facilitate much more precise fault location determinations by using impedance-based methods.
The integration process may be complex, but it is standard: OMS and DMS connectivity models must align accurately with GIS, and historical switching information must be accurate, too. Mismatches can cause AI to inaccurately recommend sectionalization, which is why utilities often invest heavily in consistency frameworks and interoperability testing in certain regions; ensuring correctness in data exchange is foundational to success and not just in academia.
Any switching recommendation engine must integrate hard constraints such as backfeed limitations, transformer capacities, voltage drop parameters, and safety protocols for crews. AI can rank switching options, but feasibility checks should be deterministic and auditable.
This is a category that most executives favor because it translates directly into avoiding failures, reducing such truck operations, and streamlining capital expenditure planning.
Predictive maintenance for utilities encompasses more than just vibration sensors and sophisticated models. It spans a wide variety of methods such as dissolved gas analysis (DGA) for transformer oil, circuit breaker operational counters, partial discharge monitoring, temperature and load cycling assessments, and even inspection imagery generated by field teams.
The Electric Power Research Institute (EPRI) runs research programs focusing on “substation asset analytics,” which incorporates both transformer and circuit breaker analysis. It highlights the role of AI in data mining and pattern recognition to derive actionable insights based on performance data gathered across the industry.
On the transformer side, DGA is still a fundamental diagnostic procedure, and studies continue to use machine learning techniques to analyze DGA patterns more uniformly under different environments.
Predictive maintenance efforts falter when teams struggle to define what constitutes a “failure” and how to align labels over time.
Transformers degrade over time rather than fail instantly, unlike web servers. A “failure” could indicate an unplanned outage, a maintenance action, a protective trip, or crossing a DGA limit. If labeling is inconsistent, the model merely learns noise.
A more effective approach is to define “health states” and predict transitions. For instance: normal → watch → recommended intervention → urgent. Each state should connect to concrete actions, rather than vague scores.
The output of predictive maintenance needs to integrate seamlessly with enterprise asset management (EAM) and workflow systems. A health model existing only in a dashboard doesn’t hold much value. Conversely, a health model that triggers inspections, influences crew schedules, and updates asset management strategies is much more effective.
There should also be a feedback loop: when maintenance occurs, it’s important to assess outcomes and build on those insights (e.g., did the intervention deter a failure? Did DGA patterns stabilize?).
Predictive maintenance initiatives frequently stall in “Proof of Concept purgatory” because vendors deliver only a model and not the accompanying workflow. You need a partner who can construct the complete pathway: ingesting signals, aligning them with asset identifiers (which is often complex), producing understandable risk metrics, and integrating them into work orders.
Demand response has evolved from being a niche program to a critical piece of grid operations as electrification (think EVs, heat pumps, industrial loads) and variability in supply increase.
OpenADR has been instituted expressly to standardize and automate demand response and DER communications, allowing utilities and aggregators to manage demand and decentralized production cost-effectively; the Department of Energy describes OpenADR as a non-proprietary interface that enables electricity providers to communicate Demand Response signals using a familiar vernacular while leveraging existing communication technologies such as the Internet.
AI plays a role in predicting customer responses, identifying opt-out trends, customizing incentives, and scheduling events to minimize consumer inconvenience while maximizing grid benefits.
For instance, instead of a broad message saying “curtail all loads,” AI can classify loads—targeting first the flexible thermal storage customers, industrial operations with some operational leeway, and EV chargers that are plugged in for an extended period. This tailored approach drives participation rates up while reducing impacts on these systems.
Additionally, AI can assist in forecasting “availability of flexibility.” A battery may only be accessible at a certain time if it was charged earlier; an EV charger provides flexibility only if the associated vehicle is parked; and an industrial load may have flexibility dictated by process constraints. These availability models are what make orchestration realistic.
Demand response directly influences customers, meaning you’ll need to outline policy constraints around fairness (e.g., not always targeting the same customers), respect contractual terms, and handle opt-out procedures. This requires a governance framework as much as a modeling one.
As a founder developing demand response tools, your key differentiator tends to be less about “better ML” and more about having an efficient program compliance engine, robust measurement, verification processes, and solid interoperability.
In many areas, non-technical loss (like theft, tampering, and billing fraud) poses a significant financial challenge. Even within “well-managed” markets, issues like tampering and unusual consumption patterns persist.
Smart meters simplify theft detection since they provide higher-frequency consumption data, allowing comparisons between household patterns and transformer-level measurements, alongside temperature influences. The research includes various machine learning techniques for identifying electricity theft using meter data and related features, which often involves ensemble learning and deep learning methods, since the telling pattern is usually subtle shifts in behavior instead of obvious spikes, but a mild behavioral change over time
Effective Implementation: Anomaly Detection Plus Investigation Workflows
In practice, electricity theft detection does not work merely by the model flagging theft for prosecution. It functions as: “the model highlights anomalies, and then you perform an investigation.”
The practical pipeline includes feature extraction from meter intervals (like shape features, deviations around seasons, and weekday/weekend differences), assessing transformer balance (comparing total meter data to upstream measurements), peer-group evaluations (considering similar households), and tamper signals from meters (such as open covers, reverse energy flows, or magnetic interference).
Ultimately, the model outputs a prioritized list of suspicious cases with supportive evidence because investigators need compelling reasons behind flagged anomalies, such as “consumption dropped by 40% following a meter change,” “transformer losses increased while neighboring households remained stable,” or “recorded reverse energy events.”
False positives skyrocket if you don’t consider true behavioral changes that might occur, like changes in occupancy, solar installation, appliance upgrades, business closures, or seasonal migrations. This underscores the importance of feature design and leveraging metadata. Ingesting events related to PV installations or meter replacements can significantly reduce false positives.
Moreover, you must remain aware that adversaries will change tactics over time. Hence, it’s crucial to maintain continuous learning and implement drift detection protocols rather than relying on a static model.
Grid congestion is becoming a significant hurdle to electrification and integrating renewables. Building new infrastructure can take years. Therefore, utilities and system operators are on the lookout for solutions that maximize the utility of existing infrastructure safely.
Dynamic Line Rating (DLR) stands out as a prime example. The Department of Energy describes DLR as technologies and methodologies for establishing conductor thermal ratings dynamically using enhanced, more granular, or real-time data, allowing operators to loosen static assumptions when conditions allow—while also maintaining caution to prevent new hazards and considering potential limits and unintended consequences.
DLR’s foundation is actually rooted in physics (through heat-balance calculations), bolstered by improved sensing. AI shines when it comes to generating forecasts—like day-ahead ratings for optimizing scheduling and market operations, not just immediate ratings.
This encompasses predictive analytics for ampacity across anticipated weather conditions and uncertainties. Some studies focus on applying machine learning techniques to anticipate overhead line ampacity to support day-ahead adjustments and optimization efforts.
Congestion management also involves optimizing topology, managing phase-shifting transformers, and conducting co-optimization with storage and demand response strategies. AI Here can help predict the risk of congestion and recommend preemptive actions.
Anything that alters line loading limits carries safety implications. Therefore, the product must encompass detailed fallback policies, confidence intervals in forecasting, sensor validation, and clearly defined operator authority.
If you’re developing DLR or congestion-related AI systems, your credibility fortifies itself through well-defined safety protocols and ample validation processes, not just high-end model designs.
While the ten solutions listed appear different on the surface, successful production hinges on similar underlying principles. If you’re working on these systems—especially with a distributed engineering team—this is where victories can be achieved or lost.
For every use case, clearly articulate the “data contract.” Identify which systems provide reliable data, the latency involved, relevant identifiers, and any access restrictions.
In utility contexts, mapping identities presents its own set of challenges. Asset IDs in GIS might misalign with IDs in SCADA, meter locations can change, and feeder boundaries might shift. Without a solid mapping layer, every model can develop unpredictably degraded performance over time.
This underlines the importance of establishing interoperability standards and consistent information models. CIM-related standards exist to facilitate integration and data-sharing across utility systems, enhancing distribution extensions, particularly as real-world utilities employ multi-vendor frameworks and require a shared semantic model.
Advisory-only frameworks can progress quickly, whereas semi-automated systems need approval and audit trails. Closed-loop controls demand well-defined constraints, simulations, staged rollouts, and more thorough cybersecurity assessments.
If your roadmap lacks clarity on the control-loop role of each feature, you risk unintentionally promising more automation than a utility is prepared to accept.
For grid control and DER orchestration, having a simulation framework isn’t optional. It’s vital to create a digital test environment that includes power flow studies, device models, and realistic noise elements.
This necessity stems not only from a technical standpoint but as a trust-building exercise. Operators are far more likely to trust automated solutions that have been evaluated under diverse stress tests.
Utility setups are prime targets for attacks. If your system touches operational tech (OT) networks, you’ll need to focus on segmentation, access controls, logging, and change management from the outset. NERC CIP requirements shape security measures and auditing in bulk electric contexts across North America.
Even in regions outside North America, utilities will typically insist on adherence to similar principles, including least privilege access, rigorous identity verification, unalterable logs, vendor risk management, and preparation for incident responses.
In utilities, drift can stem from changing weather patterns, evolving customer behaviors, fluctuations in DER adoption, tariff modifications, or topology reconfigurations. Therefore, monitoring must be comprehensive: implementing data quality checks, assessing model performance by segment, and tracking operational outcome metrics.
For outage prediction, tracking metrics of not just AUC or accuracy but elements like “crew staging effectiveness,” “ETR error distributions,” and “customer complaint rates” becomes essential to understand the impact.
A standalone dashboard rarely suffices. Utilities operate through OMS tickets, work orders, switching plans, and dispatch functionalities. Therefore, AI outputs should be seamlessly embedded within these operations.
Predictive maintenance initiatives that connect to work order management systems often outperform those that stop at simply presenting a health score. Substation analytic programs also gain efficiency when they focus on delivering actionable insights for maintenance program decisions and benchmarking rather than just forecasts alone.
If you’re a product leader in the West collaborating with an external team, the most frequent shortfall stems from ambiguity rather than capability. Utilities are intricate, and unspoken knowledge can abound. If you don’t translate domain-specific nuances into explicit acceptance criteria, the vendor may end up delivering something that showcases well but isn’t deployable.
A more effective methodology involves contracting for “system outcomes” instead of simple “model delivery.” This way, the vendor becomes responsible for integration, monitoring, security measures, and adoption processes, not just model accuracy metrics and scores.
You should also aim for a phased approach to production: commence with advisory decision support, validate value through shadow mode, and then incrementally incorporate semi-automation with necessary approvals to reduce risk and enhance trust.
Data residency and access considerations should be prioritized from the get-go. Many utilities will restrict raw operational technology data from leaving controlled environments. While this doesn’t impede outsourcing, it does demand structural adjustments: forge secure deployment packages, run models within the utility environment, and regulate external team access. This is standard procedure in regulated fields.
Lastly, ensure that operational metrics are part of your contractual obligations. If the vendor can’t specify how “success” will be evaluated (in terms of metrics like SAIDI/SAIFI impacts, truck roll reductions, peak demand reductions, loss reductions, forecast error improvements), you’re not procuring a product—you’re onboarding an experiment.
The energy transition is driving the grid to function as a dynamic, data-centric system. This shift makes AI a necessity. However, those who excel will not be the teams boasting the most advanced models. Instead, they will be the ones who successfully integrate AI with interoperability, power system limitations, cybersecurity measures, and operational workflows.
For founders or product leaders, the most crucial question isn’t “Should we employ AI?” but rather “Which of these ten solutions can we successfully implement first, delivering measurable outcomes, while adhering to the realities of the actual grid?”
Because that’s how we turn utility AI into a tangible reality.
