Decisions that cost revenue to keep the data honest

Walling off the early-support model from collections

In the winter of 2021 our customer-care team noticed what the collections reports missed. Many customers entering our hardship programme already owed more than A$2,500, and most had been struggling for months before their first missed bill. Our data could have told us sooner. Victorian interval data showed some of those households cutting their heating sharply in cold weeks, and payment records showed instalments shrinking. We were disconnecting about 6,400 residential customers a year for non-payment, and only 31 per cent of customers left the hardship programme with their debt cleared.

My team built a model that combined meter, billing, concession and payment data to flag households whose pattern had changed. The argument began once it worked. Our general manager of credit and collections wanted the score in his dialler, because early contact improves recovery; we estimated it would cut bad debt by a further A$9 million a year. Our CFO questioned the cost of a dedicated outreach team, about A$5.2 million a year. I recommended that the score go only to a new early-support team, that collections and credit systems be technically unable to read it, and that it be judged by help accepted rather than money recovered. At the bank I had watched good models turn early warnings into earlier pressure, and customers stop answering once a kind call proves to be a collections call.

The CEO supported the separation, and the Audit and Risk Committee asked for a yearly internal-audit test of whether the wall held. We gave up the extra recovery. Between 2022 and June 2026 the team contacted 61,000 households, and 23,000 accepted a payment plan, a cheaper offer, a concession they were missing or a referral to a financial counsellor. Disconnections for non-payment fell to 2,900 in 2025–26, and the share of customers leaving hardship with their debt cleared rose to 47 per cent. Bad debt fell from 1.9 to 1.4 per cent of residential revenue anyway, though by less than the dialler would have achieved.

I would change one thing. The first version looked mainly for rising debt. Financial counsellors, whom we consulted only after the pilot, told us that falling winter use in a concession household was the earlier and more important signal. I would invite them in before the first line of code.

Rebuilding the solar forecast after a spring of phantom demand

In the spring of 2023 our hedge book rested on a residential forecast bought from a specialist vendor. It estimated rooftop-solar output for each region from satellite irradiance and the capacity recorded by the distribution networks, then scaled the result to our share of customers. Between September and November our midday load in South-East Queensland and South Australia came in 9 to 13 per cent below forecast on most mild days. We had bought swaps covering demand that did not exist, and with spot prices often negative at midday, every surplus megawatt-hour cost us the swap price and the negative spot price together. The quarter's loss against plan was A$41 million.

The causes sat in our own book. Our solar and battery offers had attracted households with larger systems than the regional average, solar connections among our customers were growing by about 1,100 a month, and when we compared the network records with what meters were exporting, 9 per cent of our customers' systems were recorded at the wrong size.

Our head of trading wanted a better vendor product in place before summer. I proposed building the forecast ourselves from every solar customer's exports and running it alongside the vendor's for two full seasons before the desk relied on it. That meant another summer on the old forecast, with a manual midday adjustment that the desk disliked and I could not fully defend. Our chief risk officer was uneasy about an in-house model with no outside check, so we agreed to an independent validation every year. The build cost A$6.8 million and eleven permanent roles, and it took most of my pricing analysts away from their own programme for nine months.

By the spring of 2025 midday forecast error for the residential book had fallen from 11.4 to 4.1 per cent, and the trading team estimates hedging losses about A$19 million a year lower than under the old approach. We also passed the corrected capacity records to the networks concerned.

I would do two things differently. I would correct the capacity register first, since half the early error came from it. And I would give the trading desk a seat in the design rather than a demonstration at the end; their scepticism made the second season far better than the first.

Withdrawing 6,800 batteries until their records could be trusted

By early 2024 our virtual power plant had 36,000 enrolled batteries from three manufacturers, and about 18,000 of them were registered to provide contingency frequency control, a market that pays for fast responses when a large generator or line trips. The market operator can ask for the high-resolution measurements that show each response. In March 2024 my team reconciled a sample of events and found that for one manufacturer's fleet of 6,800 batteries, 14 per cent of the responses its cloud platform reported could not be matched to the measurements. Timestamps drifted by up to four seconds, and some devices logged nothing during the event at all.

Our general manager of distributed energy argued that the error was small at fleet level, that the manufacturer had promised a firmware fix within weeks and that withdrawal would cost revenue and goodwill. He was right about the goodwill. I recommended withdrawing that fleet from frequency markets until a full month of events reconciled, because we could not attest to responses we could not evidence. The CEO agreed, and the manufacturer objected in writing. We kept paying the affected customers their monthly participation credit, since the fault was not theirs.

The withdrawal lasted five months. We forwent about A$2.7 million of market revenue and paid A$1.1 million of credits with no matching return. The firmware fix took eleven weeks rather than two, and our testing found a second clock fault it had not addressed. When the fleet returned in August 2024, every event across all three makes was being reconciled automatically against independent measurements, and the match rate has stayed at or above 99.2 per cent since. We also rewrote our agreements with each manufacturer to cover clock synchronisation, retention of device logs and our right to audit them.

The larger change was the order of work. A new battery model now joins a market only after its data has passed reconciliation, rather than reconciliation following enrolment.

I would write those data terms into the manufacturer agreements before the first battery is enrolled, not after a fault. And I would tell the affected customers ourselves on the first day. Some learnt of the withdrawal from an owners' forum online, which cost us more trust than the withdrawal itself.

Building the emissions ledger with revenue-grade controls

As the legislation for mandatory climate-related financial disclosures passed through Parliament in 2024, it became clear that Wirrandel would be in the first group of reporters, so our first climate statement would cover the year to June 2026 and sit in the annual report. Our emissions figures then lived in three places: measured facility data for our two gas stations, reported for years under Australia's National Greenhouse and Energy Reporting scheme; spreadsheets kept by our sustainability team for scope 3; and a voluntary report whose methods had changed twice in four years. The largest figure by far, about 3.2 million tonnes a year, came from the gas our customers burn.

Our head of sustainability proposed specialist carbon-accounting software that her team knew and could have running in four months. Our CFO noted that the transitional arrangements allowed some disclosures to be phased in and suggested we report the minimum in the first year. I recommended a third course: calculating customer emissions from billed gas and electricity volumes, account by account, on our enterprise data platform, under the same reconciliation, change control and sign-off as revenue. Directors would sign these numbers, so their controls should resemble those behind the figures directors already signed. It cost A$4.3 million and eleven months, and it deferred an upgrade of our marketing analytics that had been promised for that year.

The Audit and Risk Committee endorsed the approach in August 2024. The sustainability team kept ownership of methods and emission factors, and my team took ownership of volumes and controls. Both teams found that split uncomfortable at first.

Reconciliation found a fault in our history. Customers who switched retailer part-way through a billing cycle, and accounts carrying both estimated and actual reads, had been counted twice under the spreadsheet method, overstating customer gas emissions in our 2023–24 voluntary report by 3.8 per cent. We corrected the figure publicly before the mandatory period began. The ledger now reconciles monthly to billed volumes within 0.2 per cent, and our auditor's first-year assurance work raised no matter about the underlying volumes.

I would bring the assurance team into the design in the first month rather than the seventh. Their questions about lineage would have shaped the ledger more cheaply early than they did late.