A settlement analyst at a mid-size RTO stared at a reconciliation run one afternoon and realized she had been manually matching 40,000 metering records against market schedules every month, a task so routine that a junior engineer did it in silence and never mentioned the error rate. That quiet corner of utility operations, invisible to the executive suite, is exactly where the next generation of high-leverage grid AI is hiding.

Where to Look for Novel Applications

The first wave of grid AI (load forecasting, outage prediction, vegetation triage) got there because those problems had large, clean datasets, measurable outcomes, and an executive willing to fund a pilot. But the grid runs on dozens of processes that are just as expensive and just as error-prone, where no one has yet asked the fundamental question: could a model do the repetitive part of this faster and more accurately than a human?

The discipline for finding these hidden use cases starts with mapping labor. Walk through a single month of work with a settlement team, a market bidding desk, and an EV fleet coordinator. Count the hours spent on tasks that are (1) high-volume and repetitive, (2) rule-bound enough to be teachable, and (3) currently caught by a downstream error rather than an upstream check. Those three conditions together are the fingerprint of a tractable AI application.

The interconnection queue is already the most-discussed AI use case in the program. But three adjacent domains are just as ripe and considerably less crowded in the vendor market: energy market settlement, day-ahead and real-time market bidding optimization, and electric vehicle fleet orchestration.

Settlement: The Hidden Audit Trail

Energy market settlement (also called billing reconciliation or imbalance settlement, depending on your market structure) is the process of reconciling metered delivery against scheduled and dispatched quantities, applying tariff rates, and issuing invoices. At an ISO or RTO, this happens daily for thousands of market participants. At a large IOU with retail and wholesale exposure, a settlement team may spend 30 to 50 hours per month chasing discrepancies in data that should have been machine-readable.

The AI opportunity here is not forecasting or optimization but classification and anomaly detection. A well-trained model can flag metering records that are statistically inconsistent with load shapes, weather, and prior-period actuals before a human analyst ever sees them. MAPE (mean absolute percentage error, the forecaster's metric for accuracy) has a close cousin in settlement: the exception rate, which is the share of records that require manual intervention. Reducing exception rate from 3 percent to under 0.5 percent on a portfolio of 40,000 metering points can save hundreds of engineer-hours per month and materially reduce invoice dispute resolution time.

The data for this problem already exists inside the EMS (energy management system) and the metering data management system (MDMS). What typically does not exist is a labelled training dataset of "exceptions and their resolutions." Building that dataset is the first real work of a settlement AI project, and it requires a human expert who understands why records fail, not just that they fail.

The Compliance Caution in Settlement AI

Settlement figures feed directly into revenue. An AI that misclassifies a metering exception in the wrong direction can either overcharge a market participant (a tariff compliance issue) or undercharge and erode revenue. FERC has been clear that tariff compliance is a utility obligation, not a software vendor's. Any settlement AI must operate in an advisory capacity: flagging, sorting, and drafting corrections for human review and signature. The audit trail requirement is stricter here than in load forecasting because the output has financial and regulatory consequences within hours, not days.

Market Bidding Optimization: The Quiet Competitive Edge

In organized electricity markets (PJM, MISO, SPP, CAISO, NYISO, ISONE, and the bilateral markets connected to them), generators and large loads submit day-ahead and real-time bids that reflect their marginal cost of production, their operational constraints, and their strategic view of where prices will clear. For a utility or independent power producer with a mixed fleet (gas peakers, pumped hydro, storage, wind contracts), optimizing these bids is a problem that combines short-term price forecasting, generation cost modeling, and operational constraint tracking.

AI-assisted bidding optimization is not a new category: algorithmic trading desks at large merchant generators have used predictive models for years. What is new is the accessibility of the tooling for utilities that have traditionally treated bidding as a spreadsheet exercise. A mid-size co-op participating in a day-ahead market may submit bids for a handful of resources but still leave meaningful revenue on the table by not accounting for the correlation between wind output forecast uncertainty and next-hour price spikes.

The opportunity is in the integration layer. A forecasting model that already runs day-ahead load predictions can, with the right market-data feed and operational constraint inputs, generate a draft bid schedule. A human trader reviews, adjusts for market intelligence the model cannot see (a major transmission outage announced that morning, a competing generator known to be running hot), and submits. The AI's contribution is compressing the preparation time from two hours to twenty minutes and reducing the number of "did we account for that?" moments that lead to suboptimal bids.

Numbers to verify: published studies from organized markets suggest that AI-assisted bidding can improve revenue capture by 1 to 4 percent for resources with significant flexibility, though results are highly dependent on market design, fleet composition, and the quality of price forecasting inputs. Treat any vendor claim in this range as a starting hypothesis to test on your own portfolio, not a guaranteed outcome.

EV Orchestration: The Fastest-Growing Coordination Problem

Electric vehicle fleet orchestration is emerging as one of the fastest-growing AI use cases in distribution operations, and it sits at the intersection of three trends that are all accelerating simultaneously: the adoption of medium-duty and heavy-duty electric fleets (municipal buses, utility service vehicles, delivery vans), the deployment of managed charging infrastructure at depots and commercial sites, and the growing availability of vehicle-to-grid (V2G) capable hardware.

The coordination problem is this: a depot charging 50 buses overnight has a base load of perhaps 2 MW. If all buses begin charging at the same time on return (say, 6 PM), the demand spike can trip a substation protection relay, require emergency demand response, or simply blow through a demand charge tier that erases the fuel savings from electrification. Naive charging (plug in, charge as fast as possible) is operationally simple but grid-hostile. Smart charging (stagger charge start times, honor vehicle-ready deadlines, respond to real-time price signals) requires optimization over a constraint set that changes every day.

AI adds value here at two levels. At the depot level, a reinforcement learning or constraint optimization model can manage individual charger schedules to minimize cost and flatten the load profile while guaranteeing that every vehicle departs with sufficient state of charge. At the distribution planning level, an AI that aggregates forecast charging demand across dozens of depots can help planners identify feeders that will need upgrades 18 to 24 months earlier than traditional methods would flag them.

The DER (distributed energy resource) management system (DERMS) is the system-of-record for this work. Fleet orchestration AI either integrates natively with the DERMS or operates as a parallel optimization layer that passes setpoints to charger controllers. Reliability professionals should be aware that V2G dispatch, when used as a grid resource, brings the fleet operator into a quasi-demand-response relationship with the utility, with all the settlement, telemetry, and curtailment-notification obligations that entails.

Finding the Next Use Case: A Structured Search

Beyond settlement, bidding, and EV orchestration, the method for identifying novel applications is repeatable. Start with data inventory: ask every operational team what data they generate that no one outside the team consumes. This orphan data (substation temperature logs, protection relay event files, service request metadata) is often the raw material for a useful model. Then ask the team where their work gets stuck: what review steps take the longest, where do errors accumulate, what decisions get deferred because the analysis is too slow?

Map each candidate against three filters. First, is there a measurable outcome that a model could optimize for? Second, is there enough historical data to train and validate? Third, is the outcome consequential enough to justify the governance overhead of deploying AI in a regulated environment? A promising candidate passes all three. A candidate that fails the third filter is not a bad idea; it may simply need a lower-stakes entry point, such as an advisory dashboard rather than an automated action.

The table below illustrates how several emerging use cases score on this framework:

The Load Surge Context: Why Novel Use Cases Matter Now

The 166 GW of projected five-year peak demand growth, roughly 90 GW of it from data centers alone, is not just a forecasting challenge. It is a systems challenge. Every process in the utility (settlement, market operations, distribution planning, interconnection study) will face more volume, more complexity, and faster turnaround requirements as the load surge matures. The settlement team that took comfort in a stable 40,000-record monthly run is about to see that number grow as new large loads come online, each with their own metering configurations, tariff riders, and demand-response obligations.

AI applications that are novel today (settlement anomaly detection, EV orchestration at scale) will be operational necessities within three to five years. The utilities that identify and pilot them now will have the institutional knowledge, the trained models, and the validated governance frameworks when the volume surge arrives. Those that wait will be building under pressure, with less time to run the holdout tests and reliability reviews that a safety-critical deployment requires.

The NERC Computational Load Entity (CLE) registry, committed for delivery by December 31, 2026, will add a new class of registered grid actors: large compute loads with reliability obligations. Every CLE that connects to your system is a new source of settlement complexity, dispatch telemetry, and potential demand-response capacity. The utility that has already piloted AI-assisted settlement and EV orchestration will onboard CLEs faster and with fewer errors than one that is still reconciling records by hand.

Worked Example: Spotting a Settlement Candidate

Consider a fictional but representative scenario at a 500,000-customer IOU with wholesale market participation through a regional transmission organization. The settlement team of four analysts runs a monthly reconciliation of metered load against day-ahead schedules and real-time deviations. In a typical month, roughly 1,200 records (about 3 percent of total) require manual review: the analyst pulls the raw meter file, checks the interval data against the EMS historian, identifies whether the discrepancy is a communication gap, a meter calibration drift, or a legitimate load deviation, and either corrects the record or escalates to the metering team.

The process works, but it consumes roughly 80 to 100 hours per month of skilled analyst time, and it catches errors after the fact rather than flagging them in real time. An AI pilot would start by labelling three years of closed exception records with their resolution type. That labelled dataset becomes the training corpus for a classification model that runs nightly against incoming metering data and flags records that match known exception patterns. The model's output is a prioritized exception queue, not a corrected record: the analyst still reviews and resolves, but she starts the day knowing which 20 records need immediate attention rather than scanning all 40,000.

The pilot success metric is reduction in analyst time per exception, not elimination of analyst review. A good first result is 30 to 40 percent time reduction on exception handling. A great result is a secondary benefit: because the model flags exceptions earlier, some can be resolved before the monthly cycle closes, reducing invoice disputes. Both benefits are measurable and defensible in a rate case as operational efficiency gains.

The Innovation Governance Gate

One reason novel use cases stall at the idea stage in utilities is that the governance process for deploying AI is designed around the compliance and reliability risks of the use case, not the upside. A settlement anomaly detector that never autonomously corrects a record still needs to pass through an IT security review (it reads production metering data), a data governance review (it stores historical exception records), and potentially a NERC CIP review if the metering data originates from a substation within the electronic security perimeter.

The practical response is to front-load the governance conversation before writing a single line of model code. Present the use case to your IT/OT security team, your compliance lead, and your data governance officer as a concept paper, not a deployment proposal. The concept paper describes the data sources, the model's output (advisory only, no autonomous action), the human review step, and the audit trail. This early engagement does two things: it surfaces objections that would kill the project later at lower cost, and it builds the internal champions who will accelerate the governance review when the pilot produces results.

Utilities that have successfully deployed novel AI applications consistently report that the governance timeline, not the model development timeline, is the binding constraint. A settlement anomaly model can be trained and validated in six to eight weeks on a reasonable dataset. Getting it through IT security review, data governance, and executive sign-off often takes three to six months. Budget that time in your pilot plan from day one.

Key Takeaways

• The highest-leverage novel AI applications in utilities are often hiding in back-office operations (settlement, market bidding, fleet coordination) rather than in control-room AI, because the data exists, the outcomes are measurable, and the labor cost is significant but invisible to leadership.
• Settlement anomaly detection, market bid optimization, and EV fleet orchestration are the three emerging use cases with the strongest combination of data availability, measurable outcomes, and near-term volume pressure from the load surge.
• Any AI application that touches settlement or market bidding operates under FERC tariff compliance obligations; the model advises and flags, the human reviews and signs off, every time.
• EV orchestration AI integrates with the DERMS and must account for V2G vehicles entering quasi-DR relationships with the utility, including telemetry, settlement, and curtailment notification obligations.
• The structured search method for novel use cases starts with orphan data (what data does your team generate that no one else consumes?) and routes through three filters: measurable outcome, data sufficiency, and regulatory consequence level.
• The NERC CLE registry and the load surge forecasts, which utility-reported data puts at roughly 166 GW of five-year peak growth (with significant analyst uncertainty around the actual realized figure), mean that settlement complexity and fleet coordination volumes will grow substantially by 2028; utilities that pilot these AI capabilities now will have validated governance frameworks when scale is mandatory.
• Numbers from vendor case studies on bidding optimization (1 to 4 percent revenue improvement) and settlement efficiency (30 to 50 percent exception-handling time reduction) are starting hypotheses to test on your own portfolio, not guarantees to carry into a rate case without verification.