A Model Ranked the Disconnection Queue. The Tariff Did Not Move.
Arrears scoring, deposit sizing and disconnect prioritisation are ordinary analytics problems everywhere except here, where the output turns off someone's heat. The protections that govern that decision predate the model by decades, and they bind the utility regardless of what the ranking recommended.
Four Decisions, Not One
"Using AI in collections" is usually four separate automated decisions with four different legal characters, and treating them as one programme is where the compliance work goes wrong. The deposit decision is credit-like. The payment arrangement offer is a benefit allocation. The disconnection order is an essential-service action governed by tariff. The outreach targeting decides who even learns that assistance exists.
The last one is the quietest and often the most consequential. A model that predicts which accounts will self-cure, used to decide where to spend limited outreach capacity, allocates the assistance programme. If the prediction correlates with characteristics that track protected groups — and arrears prediction built on payment history usually does — you have distributed a benefit unevenly without anyone making a decision they would recognise as one.
The Failure Modes, Ranked by How Often They Ship
Your outcomes are already public. Utilities report disconnection volumes to commissions, and service territories map cleanly onto census geography. Anyone can join those two datasets and produce a distribution chart without ever seeing your model. Assume that analysis will be run by someone who is not you, and run it first — the difference between a finding and a defence is usually whether the utility measured the same thing a year earlier and acted on it.
Why "We Never Used Race" Is Not the Answer
It is the answer to a question about intent, and the frameworks that apply here mostly ask about effects. Civil rights duties attached to federal assistance, state utility non-discrimination provisions, and credit discrimination rules all contemplate a facially neutral practice that produces a disparate outcome. Under that structure the relevant defence is that the practice serves a substantial legitimate objective and that no less discriminatory alternative achieves it comparably — which is a factual question about model alternatives you tested, not a statement about inputs.
Practically, that means the artefact that helps you is a record of alternatives considered: a simpler model, a version without the proxy-heavy features, a different threshold, a policy of contacting before ranking. If a less discriminatory alternative performed nearly as well and you did not adopt it, that is the finding. If you tested and documented that it did not, you have the beginnings of a defence.
A Deployment That Survives a Commission Complaint
- Encode protections as hard upstream gates. Medical certificates, moratoria, active arrangements, pending assistance and dispute status exclude an account from the candidate pool before scoring, not after.
- Separate the four decisions. Deposit, arrangement, outreach and disconnection have different rules and should have different models, thresholds and approvals.
- Audit features for proxies. Test each candidate feature for correlation with protected characteristics at the geography you serve, and record what you removed and why.
- Keep a human decision on the final order. A named person authorises the disconnection list, with the ability and the standing instruction to remove accounts. Automation ranks; a person acts.
- Retain scores with feature snapshots. The score alone is not reconstructable. Store the inputs as they stood at decision time, for at least your longest complaint window.
- Give specific, disputable reasons. On deposits and denials, the principal reasons and the route to challenge them — written for a customer, not for an auditor.
- Run outcome testing on a schedule. Quarterly, disaggregated, compared to the prior period, with a documented response when a gap appears. The schedule is what turns a chart into a programme.
Frequently Asked Questions
We are a municipal utility, not investor-owned. Do these rules reach us?
Different rules, similar exposure. Municipal and cooperative utilities often sit outside state commission jurisdiction, which removes the tariff layer but adds others: as public bodies they carry constitutional due process obligations around termination of an essential service, local ordinances frequently impose their own notice and moratorium duties, and federal assistance programmes bring civil rights conditions with them. The practical effect is that the disconnection process has to be procedurally regular and explainable, which is the same requirement the tariff imposes on an investor-owned utility by another name.
Is a model that only predicts payment likelihood a regulated decision?
It becomes one the moment its output is used to allocate something. A prediction sitting in a dashboard is analytics. The same prediction determining who receives a payment arrangement offer, what deposit is demanded, or where an account sits in a disconnection queue is an automated decision producing a significant effect, and it is that use rather than the model that the rules attach to. This is why the inventory should be organised by decision rather than by model — one score often drives three decisions with three different regulatory characters, and the strictest one governs.
How do smart meters and remote disconnection change the analysis?
They remove the friction that used to function as an informal safeguard. A field visit created a last human checkpoint — someone at the premises, occasionally noticing circumstances that stopped the action. Remote disconnection executes the list exactly as computed, at scale, with no such checkpoint. If your pipeline ends in a remote command, the human authorisation step and the upstream protection gates carry the entire load, and they should be tested as safety-critical controls rather than as business logic. Several commissions have added specific requirements around remote disconnection for this reason.
Can customers opt out of automated collections decisions?
In several states, in substance, yes. Comprehensive privacy statutes give consumers a right to opt out of profiling that produces legal or similarly significant effects, and access to an essential utility service is close to the paradigm case. Some regimes also require a risk assessment before such profiling and offer a route to human review of the outcome. The operational implication is that you need a mechanism for a customer to request human review of an automated collections outcome and a record that the review actually occurred — a rubber-stamp review is worse than no mechanism, because it is documented.
What is the single highest-value control to add first?
The hard gate, enforced upstream and tested with seeded accounts. It prevents the specific harm that produces the worst outcomes — a wrongful disconnection of a protected household — and it is the one control that is easy to verify and easy to demonstrate. Fairness testing matters and takes months to move; the gate can be built and proven this quarter. Seed the pipeline with synthetic accounts carrying every protected status you recognise, run the whole flow, and confirm none of them reaches an order. Then run that test on every release, because the way the gate fails is a change nobody connected to it.
Rank the Queue, Do Not Author the Decision
Everything workable in this area follows from one separation: the model orders work, and the rules decide who is eligible to be worked. Protections upstream, a named human on the final list, and retained inputs so any single account can be reconstructed.
Build the gate and the seeded test first. The fairness analysis is the longer project and it matters, but the gate is what stops the disconnection that turns an analytics programme into a docket entry.