Week 9 Session: Who Carries the Cost

A debt-recovery scheme acted on an estimate at scale, and the question that outlasts whether the estimate itself was any good — who pays when it is wrong

3 May 2027

Every session so far has asked how good your number was. This one asks what happens once a number stops being an exercise and starts being acted on — and it asks it about a case where one arithmetic shortcut was applied to hundreds of thousands of people, each of whom received a demand for money.

The scheme#

From 2015, a Commonwealth debt-recovery program compared a welfare recipient’s reported fortnightly income against their income for the whole year, taken from Australian Taxation Office records and divided evenly across the twenty-six fortnights. Where the fortnightly figure implied by that average did not match what the recipient had reported, a debt notice went out.

That averaging is a proxy problem in exactly week 5’s sense — a smoothed annual figure standing in for a real fortnightly one, and the department had been told as early as November 2014 that the substitution did not accord with the governing Act. Naming that is quick, and it is not what this session is about. What matters here survives even if you set the proxy question aside entirely: the estimate was wrong for a known, structural reason, it was used to act on people anyway, and it stayed in use for years after a tribunal said so in writing.

Two directions, one exposed to only one of them#

An average built from uneven income runs in both directions. Some fortnights it overstates what a person earned; others it understates it. What it destroys is the pattern: twenty-six different fortnights are replaced by one number repeated, and every question the Act actually asks is about a particular fortnight.

Be careful about what follows from that, because the obvious thing to say is wrong. The overstatements and the understatements cancel across the year — they have to, since they are deviations from the year’s own mean — but that is the only sense in which they balance. They are not the same size, and they are not equally frequent. Their shape depends entirely on the shape of the real income. Work the arithmetic on someone who earned the year’s income in a single busy quarter: in the nine quiet months the average overstates, fortnight after fortnight, and in the busy one it understates by a lot, a few times. Many overstatements, few understatements, same annual total. Income that arrives in bursts rather than evenly produces exactly that asymmetry.

Then look at what happened next, which did not depend on the arithmetic at all. Overstate a fortnight’s income and a debt notice appears in the mail, with a number attached and a date to respond by. Understate one and nothing happens — a smaller debt, or no debt, sits quietly inside a government ledger, discovered by no one, contested by no one, costing the recipient nothing.

That asymmetry is institutional, not mathematical. The rules decided that one direction of error would be pursued and the other would not. One lands on a person who has to do something about it. The other lands on an agency’s books, where it can wait indefinitely without hurting anyone.

Who had to prove what#

Ordinarily, a debt exists once Services Australia is satisfied that a particular payment, in a particular fortnight, was too high. The Royal Commission’s finding was that this program inverted that: a recipient who received a notice had to show the averaged figure was wrong, or the debt stood and was recovered — including, for a period, through external debt collectors.

Producing that evidence years later, for a fortnight the recipient had no reason to have kept records of at the time, fell on the person with the least capacity to produce it and the least power to argue about who should have to. The agency that built the estimate never had to prove it was right. The person it was used against had to prove it was wrong.

The Commission’s own words, from its final report:

“Robodebt was a crude and cruel mechanism, neither fair nor legal, and it made many people feel like criminals.”

What it cost, and when#

A March 2017 tribunal decision found the averaging methodology unlawful. The program continued substantially unchanged for more than two years after that.

The reckoning, when it came, landed on the Commonwealth, and it arrived in distinct pieces that are worth keeping apart:

• May 2020. The Commonwealth conceded that debts raised solely by income averaging had not been validly raised, and moved to zero those debts and refund what had already been repaid. This was a government decision, not a court order or a settlement.
• 11 June 2021. The Federal Court approved a settlement in the class action Prygodicz v Commonwealth (No 2) — $112 million, calculated in the nature of interest, alongside a declaration that debts based solely on income averaging were invalid. Compensation, not refunds: the refunds had already happened.
• 23 June 2026. The Court approved a further settlement on appeal, covering heads of loss the first settlement had not, and additional to both the 2021 amount and the debts already zeroed and refunded. Read its parts rather than its headline. A package of up to $548.5 million was approved, and it is three different things: $475 million in compensation for eligible group members who registered, $13.5 million for the applicants’ reasonable legal costs, and an amount up to $60 million for administering the scheme. Only the first of those is money paid to group members, and because the third is a ceiling rather than a fixed sum, so is the headline: up to $548.5 million is what was approved, not what will necessarily be spent. Quoting any of these figures as any of the others is the week 1 problem wearing a larger number.

Note what that sequence does with time. The first direction of error cost individual recipients immediately, personally, and for years. The second cost the Commonwealth eventually, collectively, and only after the scheme had run its course and been picked apart in public.

Being uncertain about a number and deciding who absorbs that uncertainty once it is used are two different questions. This case answers the second one for you, in public, at scale.

Build the ledger#

In pairs, build a who-bears ledger for this scheme with four cells: what happens to the recipient if the estimate runs high, what happens to the recipient if it runs low, what happens to the agency if the estimate runs high, what happens to the agency if it runs low. Fill all four before you decide which one actually happened — an asymmetry you only notice after filling in the cell that stayed empty is not one you have actually understood.

Compare ledgers. The scheme that was built is legible from which cells are full and which are not.

Your stakes note, for the Court#

Turn the same four cells on your own Client Estimate. Write, in your own words:

• who bears the cost if your City of Melbourne estimate comes in too high;
• who bears the cost if it comes in too low;
• whether those are the same party, or different ones;
• what that implies for how you state your range in the Estimate Court — a number whose two directions land on the same people can be defended flatly; a number whose two directions land on different people cannot.

This note is not an exercise. It is the stakes note the Client Estimate brief asks for, written now, while the case above is still specific in your head.

Leaving#

Nothing here goes in the Calibration Diary. There is no range to commit and nothing to resolve later — the point of today was the ledger, and the ledger is already done once you have written it.

Next week the target stops being something that already happened.