Western Australia's government is asking Deloitte for a quick refund on a feasibility study about merging the University of Western Australia, Curtin University and Murdoch University. The study had references that do not exist. When journalists then checked the appendix, they found cost totals that were wrong by hundreds of millions of dollars. Deloitte will correct the report and reissue it.
Wrong sums are a different kind of failure from fake references
A fake reference can only be caught by looking it up, one source at a time. A wrong total needs no outside knowledge, only the report's own numbers. In the high-cost merger of Curtin and UWA, the report lists A$514m to get to the first day and A$145m over the five years after. Those add up to A$659m, and the report prints A$399m. In the low-cost three-way merger, A$396m plus A$116m is A$512m, and the report prints A$823m.
So one total is too low by A$260m and the other is too high by A$311m. The errors do not lean toward one side of the argument. They look like a table that nobody added up before it went out.
The government has been advised that the findings are unaffected. That may be true, but a decision-maker holding a table that disagrees with itself cannot tell which number drove the conclusion. That is why the table has to be recomputed, not just trusted.
The refund is smaller than the risk
The state paid A$166,000 for the study. The error in one scenario is more than 1,500 times that. A merger of three universities is a decision worth hundreds of millions, so the fee is a rounding error next to what a wrong or delayed decision would cost. A refund returns the price of the document. It does not return the cost of acting on it.
The first Australian case showed the same limit. There the remedy was the final instalment of the fee, after fabricated references and a made-up judge's quote had to be removed. A month later, in November 2025, a C$1.6 million health workforce report for Newfoundland and Labrador was found to have false citations. That makes three governments in one year, all handed the same kind of flaw by the same firm.
I think the reason is simple economics. AI has made a hundred-page report cheap to produce, but it has not made a careful reading cheaper. The writing side got faster and the checking side stayed where it was, so the checking has moved to whoever receives the document.
What a buyer can do on Monday
For example, a finance director at a hospital group commissions an options study on merging two sites. Before paying the final invoice, she asks for the spreadsheet behind the appendix and adds up every total herself. For a report with six scenarios, that is a few minutes of work. She also asks which sections AI touched, and she makes the last 20% of the fee payable only once her own check is done.
The last instalment is the one lever a buyer has, and last year's refund was exactly that. The same rule applies to reports your own team writes with AI. In that case you are the consultant, and the person who signs it is the one who has to add up the totals first.