Deloitte has handed in another report that doesn't add up

514 plus 145 equals 399

Page 83, in an appendix, a cost table. First row: what merging two universities would cost before day one, A$514 million (Australian dollars throughout). Second row: the following five years, A$145 million. Bottom row, the total: A$399 million.

You don't need a calculator to see the oddity. The table compares three cost scenarios, low, median and high. With that addition, the high scenario becomes the cheapest of the three.

The table sits in a Deloitte report for the Western Australian government, which cost A$166,000 according to the Guardian. The government told the paper it was extremely disappointed, and it has written to the firm to speed up a refund.

It's not the first time. In 2025, Deloitte already had to hand back part of its fee on a report for the Australian federal government, over dubious references.

A report on merging three universities

Western Australia is weighing a merger of three of its public universities: the University of Western Australia (UWA), Curtin and Murdoch. Deloitte's report, a 219-page rapid cost-benefit analysis dated 26 November 2025, was released by the government on 11 October 2026 alongside two other studies, to support the next steps of the project.

Three days later, on 14 October, the Australian Financial Review flagged dubious references. According to Times Higher Education, which reports the investigation: three quotations that appear to be credited to the wrong authors, two references that don't say what the report claims they say, and a link to an unrelated article. Deloitte answers that these are minor referencing errors with no effect on the conclusions.

On 26 October the Guardian published the two sums. The one on page 83, off by A$260 million. And another in the low scenario of a three-way merger: 396 plus 116, which is 512, written as 823. A gap of A$311 million. Higher education minister Tony Buti first called them typos, then said the next day he was angry, the paper reports.

We redid the sums

The report is online on the government's website. We went through Appendix D, which details the costs of the three merger cases, each under three scenarios: nine totals in all.

The Guardian's two errors are there. There's also a third, smaller one that the article doesn't mention. In the median scenario of the Curtin and UWA merger, 421 plus 122 makes 543, and the table says 533. A A$10 million gap. The other six totals are right, and so are the subtotals. Only the bottom line goes astray.

The errors don't stay on their row. Just below, the net cost after expected savings starts from the wrong totals, and becomes wrong in turn.

What these errors do to the conclusions, nobody can say. The report's headline figures (A$175 million, A$450 million and A$522 million depending on the merger case) are discounted values, not the appendix totals, and we couldn't trace the link between the two. The government tells the Guardian the errors affect neither the modelling nor the conclusions, and that Deloitte will reissue a corrected report.

Did the AI miscount?

Nothing in the record answers that. The Guardian sets the dubious references, the wrong sums and Deloitte's use of AI side by side, without establishing a link between them. A wrong sum in a spreadsheet is easily made by hand: all it takes is a number typed in where a formula should be.

The report itself says what the AI did. Section 4.8 explains that generative AI was used to gather documentation, synthesise information and proofread the text. It states that no AI output, neither modelling nor drafting, was used to prepare the report. And it adds that everything AI-assisted went through a full quality review by Deloitte professionals.

In the references, a link to a World Health Organization recommendation carries the suffix utm_source=chatgpt.com, which means it was copied from ChatGPT. The Guardian spotted it, and we found it in the file. That fits the declared use, and it says nothing about the sums.

The same report writes "Curtin University of Technology" 55 times, a name the university hasn't used since 2010 according to the Guardian, and six headings carry the typo "UNIVERISTY". Nothing serious, but a proofread catches this kind of thing. This one didn't.

Almost a year in the client's hands

Delivered in late November 2025 and published in October 2026, the report spent nearly a year with the government that commissioned it. In a document meant to inform the merger of three universities, the bottom line of a cost table stayed wrong: neither the firm before delivery nor the client afterwards fixed it.

A consultancy sells expert time, and part of that time is checking. A report that says it was fully reviewed is like a receipt stamped "checked": the stamp is only worth something if someone recounted. Here, the sentence about the quality review and the appendix that contradicts it are in the same file.

The second time in a year

Deloitte has been through this before. In 2025, the firm delivered to the federal Department of Employment and Workplace Relations a review of the system that automates penalties for welfare recipients, on a contract worth about A$440,000. Christopher Rudge, a researcher at the University of Sydney, had found non-existent references in it, and the Australian Financial Review then broke the story in August.

The corrected version, posted in October 2025, mentioned in an appendix a toolchain based on GPT-4o, without Deloitte attributing the errors to AI. The firm agreed to repay the contract's final instalment. At a Senate hearing on 8 October 2025, the Department of Finance put the refund at A$97,587, according to the Canberra press, and acknowledged it had learned of the errors from the newspapers.

The two cases share one feature that has nothing to do with technology. Neither the firm nor the client finds the errors. They arrive from outside: an academic in 2025, the press in 2026.

What the client is paying for

Christian Downie, a governance researcher at the Australian National University, told the Guardian that as AI use grows, governments need to revisit their contracts with consultancies, to build in the right incentives and the right penalties. He also wants to know exactly at which stages AI is involved.

That may be the real shift. We already know AI makes mistakes. What changes is that it gives firm and client alike the feeling that part of the work is happening by itself, including the part that consists of checking. For A$166,000, by the Guardian's account, Western Australia received a report that says it was reviewed in full. Its wrong sums stayed in it until the press published them.