Read the meter for hardship before you read it for debt

Every energy retailer in Australia now holds a record of how its customers live. In Victoria almost every home has had a smart meter for more than a decade, and the rollout in the other eastern states is moving towards near-universal coverage before the decade ends. The meter reports what a home draws from the grid, and what its panels send back, at intervals of thirty minutes or less. A retailer that wants to know whether a family is home during the day, whether a heater runs overnight or whether the dryer has stopped being used can find out without asking anyone.

Most public debate about that record concerns privacy, and rightly so. I want to argue about something narrower and, I think, more urgent: what a retailer does when its data suggests that a customer is beginning to struggle. My position is that the signal should be used only to offer help, that the separation should be built into systems rather than written into policy, and that a retailer should be willing to say openly what the separation costs.

What a cold week in a quiet house tells us

Payment difficulty rarely begins with a missed bill. By the time a bill goes unpaid, a household has usually tried other things first. In our data the earliest sign is often not debt at all but behaviour. A pensioner couple in Geelong whose usage falls by a third during a July cold snap has not suddenly become more efficient. They have switched off the heater. A family whose payments shrink from the full amount to a round fifty dollars, then arrive a little later each month, is negotiating with itself before it negotiates with us. A small business whose daytime load disappears for a fortnight may have closed its doors.

Financial counsellors have a name for the first pattern: self-rationing. It is the quietest form of hardship and, for health, among the most dangerous, because a cold house in winter carries real risk for older people and young children. It also produces no debt. A collections model built to predict non-payment will never see a household that has stopped heating its home in order to keep paying us.

At Wirrandel we built an early-support model in 2022 to look for these changes: falling winter use in a home that holds a concession card, payments shrinking or slipping, bills rising faster than usage would explain, a sudden collapse in daytime load. The first version, I should admit, looked mainly at the trajectory of debt. Financial counsellors whom we consulted after the pilot told us we had the order wrong, and they were right. Self-rationing is now among the model's strongest inputs.

A wall between the help score and the collections queue

The model worked, and that was when the argument began. A score that identifies customers drifting into difficulty is also, unavoidably, a good predictor of future bad debt. Our collections team wanted it in their dialler, because an early call tends to recover more, and they estimated, fairly, that it would cut bad debt by a further nine million dollars a year. I would not agree, and I want to explain why in terms that go beyond kindness.

I spent five years building collections models at a bank. I know how efficiently a well-meant early warning becomes earlier pressure. The incentives in a collections function are honest ones, set by the business, and they point towards recovering money. Put a hardship signal into that function and it will be used to recover money sooner, whatever the policy says. The customer then learns that an early call from us is a collections call and stops answering. The next time we ring to offer a concession they are missing, nobody picks up.

So the separation at Wirrandel is technical, not merely procedural. The early-support score is produced in an environment that collections and credit systems cannot read. It goes to one team of 38 people whose objectives are help accepted and disconnections avoided, and whose calls are reviewed against those objectives. It is excluded from marketing and from any credit assessment. Our internal auditors test the separation every year and report to the board's Audit and Risk Committee on whether it held.

The cost is real and should be stated. We gave up an estimated nine million dollars a year of additional recovery, and we spend about five million a year on the team. Against that, between 2022 and June 2026 the team contacted 61,000 households, and 23,000 accepted something: a payment plan, a move to a cheaper offer, a concession they were entitled to but not receiving, or a referral to a financial counsellor. Disconnections for non-payment fell from 6,400 in 2021–22 to 2,900 in 2025–26. Bad debt fell too, though by less than the dialler would have achieved.

The strongest objection to this approach deserves a hearing. A retailer that calls because its data suggests a family is struggling is acting on an intimate inference, and some customers find the call intrusive. We have had complaints. Our answer has been to make the first contact an offer rather than a diagnosis. Our staff never tell a customer what the data says about their heating. They explain that we review accounts each winter, ask whether we may check that the customer is on the right plan and receiving every concession, and leave the door open. About one in six declines, and the decline is recorded so that we do not ask again that season.

The obligations retailers already carry point in this direction without quite arriving. Hardship policies approved by the AER, and Victoria's payment difficulty framework, describe what we must offer once a customer is in difficulty. They say much less about noticing. Privacy reforms passed in 2024 will soon require organisations to explain in their privacy policies how automated decisions that significantly affect people are made. I welcome that, and I would go further for our own sector: a retailer should be able to show any customer which uses of their meter data are permitted and which are walled off.

Four questions for a retailer's board

Directors do not need to inspect a model to hold management to account for it. I would put four questions to any energy retailer's board, and I ask them of myself.

First, does the company use meter data to identify customers who may be in difficulty, and if so, where does that signal go? If the honest answer includes collections, the board should hear it plainly.

Second, is the separation technical or only written down? A policy that relies on people not looking is weaker than a system that does not let them.

Third, what is being measured: help accepted, disconnections avoided and hardship debts cleared, or money recovered? The measure will decide how the model is used more surely than any policy.

Fourth, who outside the company has examined the approach? Financial counsellors see the consequences of our decisions in kitchens we never visit. We pay one to sit on our data ethics panel, and she has changed more of our models than any regulator has.

Smart meters will soon be close to universal along the east coast, and every retailer will hold the evidence we hold. The question is whether the industry reads it first for hardship or for debt. That choice will be made model by model, in rooms like mine, long before any rule requires it.