A demand response program manager pulls the enrollment data after the first automated targeting run and notices something: the algorithm enrolled almost no customers in the eastern zip codes. Those zip codes are predominantly lower-income neighborhoods with older housing stock. The program offered bill credits, faster outage response, and load flexibility value to the grid. The customers who were not enrolled got none of that. And the grid got less flexible load from the part of the feeder that most needed it.

The Equity Problem in AI-Driven Enrollment and Targeting

Demand response (DR) programs, distributed energy resource (DER) incentive programs, and flexible load platforms all depend on targeting: choosing which customers to contact, which to enroll, which to prioritize for event dispatch, and which to curtail when the grid needs relief. When these targeting decisions are made by human program managers working through spreadsheets, biases can exist but they are visible and auditable: you can look at the eligibility criteria, the outreach lists, and the enrollment records and ask whether the program is reaching the customers it is supposed to reach.

When these targeting decisions are made or mediated by an AI model, the bias can become invisible. The model optimizes for the objective it was given, which is usually something like "maximize curtailable load" or "minimize per-unit program cost" or "enroll customers most likely to respond." These are reasonable operational objectives. But each of them, without explicit equity constraints, can produce systematically skewed enrollment that disadvantages specific customer populations.

The populations most likely to be disadvantaged are those who have historically been underserved by grid investment: lower-income customers in older distribution infrastructure; renters with limited control over their HVAC and appliances; customers in non-English-speaking households who may not have responded to prior outreach; and customers without smart meters or smart devices whose flexibility is invisible to the data the model trains on. An AI model trained on historical program participation data will learn that these customers are "low probability enrollees" and will rank them lower in every future targeting round. The historical pattern becomes a self-reinforcing prediction.

An AI model that optimizes enrollment efficiency is not neutral. It inherits the inequities of the programs it was trained on and replicates them at scale.

How Targeting Bias Appears in DER and DR Data

Before you can detect inequitable targeting, you need to know what the data signatures of bias look like. Three patterns are most common in DER and DR programs using AI targeting.

Enrollment skew by geography. Map the enrolled customers against census data for income, race, and housing tenure. If enrolled customers are concentrated in specific geographies that correlate with higher income, newer housing, or owner-occupancy, and non-enrolled customers are concentrated in other geographies that correlate with lower income, older housing, or renter status, you have a geographically proxied bias. The AI may not have used race or income directly as targeting variables, but if the proxy variables it used (home age, smart meter presence, historical bill size, prior response rate) correlate with protected characteristics, the discriminatory effect is the same.

Disparate participation in curtailment events. Some DR programs, particularly those targeting large commercial and industrial customers, route curtailment requests preferentially to customers with the most flexible load. If the algorithm consistently routes more curtailment events to smaller residential customers or to specific neighborhoods, while larger commercial customers in wealthier areas receive fewer events, the burden of grid flexibility is inequitably distributed. This pattern does not appear in enrollment data; it appears in event dispatch logs.

Skipping in DER incentive programs. Utility-sponsored DER programs (rooftop solar incentives, battery storage rebates, EV charging infrastructure) often use AI to prioritize outreach or to assess application completeness. If the outreach algorithm under-prioritizes customers in certain geographies, or if the application completeness AI consistently applies more conservative standards to certain customer types, those customers do not receive equal access to the program benefits. This is "skipping": the algorithm bypasses a customer without an explicit exclusion, through differential prioritization.

Detection Techniques: Running an Equity Audit

An equity audit for a DER or DR targeting AI does not require a data science background. It requires access to the program data, knowledge of what fair participation looks like by geography and customer type, and a structured comparison. Here is a practical framework.

Step 1: Disaggregate enrollment data by geography. Export the enrollment file and join it to census tract or zip code data. Calculate the enrollment rate (enrolled customers divided by eligible customers) for each geographic unit. Rank the geographic units by enrollment rate. Now overlay demographic data: are the low-enrollment areas also the low-income areas, the minority-majority areas, the high-renter areas? If yes, you have identified a geographic proxy correlation that warrants deeper investigation.

Step 2: Check the eligibility criteria for proxy discrimination. Look at every variable in the AI's targeting model that influences enrollment prioritization. Does the model use smart meter data as a criterion? Smart meter deployment may correlate with income or neighborhood. Does it use prior program participation? Prior participation may correlate with receiving outreach. Does it use housing age or housing type? These are property characteristics that correlate strongly with income and race. Each proxy variable requires a disparate impact assessment: does using this variable produce enrollment rates that are lower for protected populations?

Step 3: Audit the curtailment dispatch log. For active DR programs, extract the curtailment event log: which customers received event dispatches, at what frequency, and for what duration. Calculate the average annual curtailment burden per customer by geography. Compare the distribution. Are specific geographies receiving a disproportionately high curtailment burden? Are commercial customers in certain areas consistently exempt while residential customers in others consistently bear the load?

Step 4: Test the decision model directly. If you have access to the AI's targeting model (or to the vendor documentation of how it works), construct test cases representing similar customers in different demographics and run them through the model. A customer with identical load flexibility, bill payment history, and equipment characteristics but different geographic location or housing type should receive similar targeting scores. If they do not, the model is using geographic or housing-type proxies that produce disparate outcomes.

Step 5: Compare against program design intent. The program was designed to achieve specific grid and customer objectives. Pull the original program design document and compare the intended beneficiary population against the actual enrolled population. If the program was designed to be available to all residential customers but 80 percent of enrollees are in the top two income quintiles, the program is not performing as designed.

Remediation and Equity by Design

Detecting inequitable targeting is necessary but not sufficient. The organizational question is what to do about it. Three intervention strategies apply at different levels of the AI deployment pipeline.

Pre-processing interventions. Before the AI model runs, ensure the input data reflects the full eligible customer population rather than the historically reached population. If the training data for the enrollment model came from prior program enrollment, it reflects who was historically enrolled, not who is eligible. Re-weighting the training data to represent the full eligible population, or adding equity constraints to the objective function, can shift the model's behavior before it reaches production. This is a data science intervention that requires collaboration between the program manager and the AI development team.

In-processing interventions. Modify the model's objective function to include explicit equity constraints. Instead of "maximize total enrolled flexibility," the objective becomes "maximize total enrolled flexibility subject to achieving minimum enrollment rates by geographic equity tier." This forces the model to satisfy the equity constraint as a binding requirement rather than an optimization target it can trade off against efficiency. Equity constraints can be defined geographically (achieve at least X percent enrollment in every census tract), demographically (no more than Y percent enrollment concentration in any income quintile), or by customer type (residential enrollment must be proportional to residential eligible population across each service territory sub-area).

Post-processing interventions. After the model produces targeting scores, apply a structured human review for the lowest-scoring geographic areas before those scores drive outreach decisions. A program manager who reviews the bottom-scoring zip codes before the outreach list is generated can identify patterns that warrant investigation, add targeted outreach resources to under-enrolled areas, and document the review process for regulatory reporting. This is the most auditable approach because it creates a record of human judgment applied before a potentially discriminatory list becomes operational.

The choice of intervention strategy depends on the utility's capabilities, the program's timeline, and the regulatory environment. In many state PUC jurisdictions, equity in DER and DR program access is an explicit regulatory requirement. The California PUC, New York PSC, Illinois ICC, and several other commissions have issued orders requiring utilities to demonstrate equitable program access as part of their program approval and reporting obligations. This means the equity audit is not just good practice; in many jurisdictions it is a regulatory deliverable.

The Regulatory and Reporting Framework

A state commission reviewing a DER or DR program filing in 2026 is likely to ask explicit equity questions. What is the enrollment rate by income tier? What is the curtailment burden by geographic area? What steps has the utility taken to ensure the program does not systematically under-serve specific customer populations? These questions are not hypothetical; they appear in commission data requests, program approval orders, and annual reporting requirements in multiple jurisdictions.

The reporting obligation creates a documentation requirement. The utility must be able to show: the demographic composition of the enrolled population; the methodology used for targeting and any equity constraints applied; the results of any equity audits performed; and any corrective actions taken when bias was identified. An AI-assisted targeting program that produces inequitable outcomes and has no documentation of equity review is a regulatory exposure in addition to a social equity failure.

The NERC reliability framework does not directly address equity in DR program design, but the flexibility value of demand response depends on broad participation. A DR program that enrolls only the highest-income quartile of residential customers, with newer homes, smart thermostats, and internet connectivity, is drawing its flexibility from a small slice of the load. A program that achieves equitable enrollment across the residential population draws from a much larger, more diverse flexibility base, improving program performance at the same time as it satisfies regulatory equity requirements.

This is the business case for equity by design: broader, more equitable enrollment is not just a regulatory requirement; it produces more total flexibility, lowers per-unit program cost, and creates a more resilient demand response base. The compliance benefit and the operational benefit point in the same direction.

Worked Example: An Automated Enrollment Program's Equity Review

A mid-size IOU launched an AI-assisted residential DR program with an enrollment target of 25,000 customers in the first year. The targeting model used smart meter data, prior program participation, housing age (from public records), and estimated home cooling load to rank customers for outreach. After six months, enrollment reached 18,000 customers.

The program manager ran a standard equity review. She joined the enrollment file to census tract data and calculated enrollment rates by income quintile. The results: top income quintile, 38 percent enrollment rate. Bottom income quintile, 9 percent enrollment rate. The fourth quintile (second-lowest income) showed 12 percent enrollment rate. The third quintile showed 21 percent.

She then audited the targeting model's input variables. Smart meter presence: 94 percent among enrolled customers, compared to 71 percent of the eligible population. Housing age under 20 years: 61 percent of enrolled customers, compared to 34 percent of the eligible population. Prior program participation: 52 percent of enrolled customers had participated in at least one prior utility program, compared to 18 percent of the eligible population.

Each of these three variables correlated with income and housing tenure in her service territory. The model was efficiently targeting the easy-to-reach population: customers who already had smart meters, newer homes, and prior program experience. That population was also disproportionately higher-income and owner-occupied.

The corrective action had three parts. First, the program team added a geographic equity tier to the outreach list: any census tract with an enrollment rate below the 25th percentile of the service territory got prioritized outreach, including direct mail and community organization partnerships. Second, the model's scoring was adjusted to give additional weight to customers in equity-priority geographies, reducing the penalty for older housing and absence of prior program history. Third, the program team worked with the commission staff to report the equity review findings and the corrective actions in the next program status report. The findings did not surprise the commission; what it appreciated was that the utility had identified and acted on the issue proactively.

Equity Drift: When the Model Becomes More Biased Over Time

Equity bias in AI targeting is not a one-time condition that you audit and correct. It can worsen over time through a feedback loop that is easy to miss if you are only monitoring enrollment counts rather than the distribution of those counts. The mechanism is straightforward: if the targeting model under-enrolls certain geographies in year one, those geographies have lower participation rates in the training data for year two's model update. The updated model learns that those geographies are even lower-probability enrollees, tightens its targeting away from them further, and the under-enrollment deepens. Each retraining cycle amplifies the bias rather than correcting it, because the training data is the product of the biased program rather than the full eligible population.

This drift dynamic is analogous to the load forecasting drift problem: the model is training on data that reflects its own prior outputs, not on data that reflects the true underlying pattern. The corrective discipline is the same: periodic recalibration against an independent baseline that is not contaminated by the model's prior decisions. For equity in targeting models, the independent baseline is the full eligible population by geography and customer type, drawn from the utility's customer information system (CIS) rather than from enrollment history. When the model is retrained, the training data must represent the eligible population, not the historically reached population.

Program managers should track equity metrics on the same rolling monitoring schedule they use for model accuracy: enrollment rate by geographic equity tier compared against its baseline, curtailment burden by geography compared against prior periods, and a delta between the current model's targeting distribution and the eligible population distribution. When any equity metric drifts beyond a defined threshold (for example, when the enrollment rate gap between the highest and lowest income quintile widens by more than 10 percentage points quarter over quarter), the threshold triggers a review, just as a MAPE alert triggers a forecasting review.

The Grid Reliability Case for Equitable DR: More Than a Social Good

Energy professionals sometimes treat equity in DER and DR programs as a social policy objective that exists alongside, but separate from, the reliability objectives. That framing is wrong and it is worth correcting explicitly because it shapes how seriously operations and planning staff take equity governance.

A demand response program that achieves broad, equitable enrollment across the residential customer population has a fundamentally different reliability profile than one concentrated in the top income quintile. The diversity of the participating load matters. High-income households with newer HVAC systems, smart thermostats, and home automation have more responsive and more predictable flexibility. But a DR program that relies entirely on this demographic concentrates its flexibility in a narrow geographic band: the newer, wealthier, single-family suburbs. The lower-density, lower-income corridors served by aging distribution infrastructure are typically the circuits where reactive power support, voltage management, and peak load relief are most needed.

Equitable DR enrollment spreads the flexibility across the circuits that most need it. This is a grid operations argument, not just a regulatory compliance argument. A transmission and distribution planner evaluating whether a DR program can substitute for a distribution upgrade in a load-constrained corridor needs to know how deep the program's enrollment is in that specific corridor, not just in the service territory as a whole. If the AI targeting model has systematically skipped that corridor, the planner has no DR capacity to dispatch there and the infrastructure investment goes forward. The equity failure becomes a capital planning failure.

The NERC reliability framework does not require equitable DR design explicitly, but the load model accuracy obligations embedded in NERC MOD-032 are relevant: if a utility relies on DR program performance in its load modeling for reliability analysis, that performance assumption must reflect the actual enrolled population and its demonstrated response rates, not an aspirational enrollment level. A DR program with systematic enrollment gaps in specific geographies that does not reflect those gaps in its reliability modeling has a potential compliance issue with the accuracy requirements of MOD-032, independent of any equity regulatory requirement.

The bottom line for the working professional: equity in AI targeting is simultaneously a social equity obligation, a regulatory requirement in an increasing number of jurisdictions, a grid reliability argument, and a program efficiency argument. It is not a trade-off against operational objectives; it is aligned with them when the analysis is done correctly.

Key Takeaways

• AI targeting for DER and DR programs inherits the inequities of historical participation data: models trained on who was previously enrolled predict that the same populations will enroll again, creating self-reinforcing under-enrollment in historically underserved communities.
• Three data patterns signal inequitable targeting: enrollment skew by geography correlated with income or race proxies, disparate curtailment event burden across geographies, and systematic skipping of certain customer types in DER incentive program outreach.
• An equity audit requires disaggregating enrollment data by geography, checking eligibility criteria for proxy discrimination, auditing the curtailment dispatch log, testing the decision model with comparable customer profiles, and comparing outcomes against program design intent.
• Three remediation strategies operate at different pipeline stages: pre-processing (re-weight training data to represent the full eligible population), in-processing (add explicit equity constraints to the objective function), and post-processing (structured human review of low-scoring geographies before outreach lists are generated).
• Equity bias can worsen with each retraining cycle if the training data reflects the model's own prior enrollment decisions rather than the full eligible population; equity drift monitoring requires tracking enrollment distribution against the eligible population baseline, not just against prior enrollment.
• In many state PUC jurisdictions, equity in DER and DR program access is a regulatory deliverable, not just a best practice; commission filings require documentation of equity methodology, audit results, and corrective actions.
• Equitable enrollment is a grid reliability argument as well as a regulatory one: broad participation across geographically diverse load pockets gives planners DR capacity precisely on the circuits where distribution infrastructure is most stressed, turning equity governance into a capital deferral enabler.
• A utility that relies on DR performance in its NERC MOD-032 load modeling must ensure the enrolled population and response rates reflected in that model match the actual program, including any geographic gaps created by AI targeting bias.