The Deloitte AI report refund: anatomy of a missing review step

In December 2024, Australia's Department of Employment and Workplace Relations commissioned Deloitte to run an independent assurance review of its Targeted Compliance Framework, the automated system that penalises jobseekers who miss welfare obligations. The 237-page report was published in July 2025, and it looked exactly like what a top-tier firm delivers: structured, referenced, fluent.

Inside the polish sat fabrications. A quote attributed to a federal court judgment that the judgment does not contain. Citations to academic papers that were never written, attributed to real professors at Lund University and the University of Sydney. Deloitte later confirmed that a generative AI toolchain based on Azure OpenAI GPT-4o had been used in producing the report, and agreed to repay the final instalment of the contract. A corrected version, with the AI use disclosed, replaced the original.

The errors were not found by a procurement checklist, a second AI, or the firm's own review layers. They were found by Chris Rudge, a welfare academic at the University of Sydney, who read the report closely and noticed something specific: papers attributed to his own colleagues that he knew did not exist.

That detail is worth sitting with. Every institutional safeguard between a global consultancy and a national government let the fabrications through, because at every layer the report was reviewed for how it read, not for whether its sources held. The one reader who checked was the one who knew the terrain well enough to be surprised. Verification did not fail because it is hard. It failed because nobody upstream considered it their job.

Nothing about this failure is unusual to AI. It is the pattern we have written about across this series: models restate weak or invented material in the same confident register as solid material, so a fabricated footnote reads exactly like a real one. Fluency is what review processes are tuned to catch, and fluency was flawless.

The economics made it worse. Checking a claim means opening the source behind it, and a 237-page report carries hundreds of claims. Without tooling that makes each check cheap, the honest cost of full verification looks prohibitive, so organisations quietly skip it and hope. The Deloitte case is simply what that hope looks like when it lands on the front page, with a refund attached.

The fix is not to ban AI from deliverables. The report's problem was not that AI helped write it; it was that no cheap path existed from each claim back to its evidence, so nobody walked that path before the client did.

A working review step has three parts. First, provenance by construction: every factual claim in the draft carries a link to the exact passage it came from, so checking is a click, not a research project. Second, a verification pass scoped to what matters: the claims you would be most embarrassed to retract, the numbers that drive the recommendation, anything surprising. Third, ownership: the person whose name is on the deliverable runs the pass, because the Deloitte story shows what happens when everyone assumes the layer below did it.

On that last point, the story generalises beyond consulting: whoever signs, checks. It is the same rule teams use for any AI-assisted work product.

This is the posture Tatsulok is built around. Answers come only from your documents and curated collections, every claim links to its exact passage, and gaps are stated honestly instead of filled with plausible inventions. The review step that was missing in the Deloitte report is the interface: you read the claim, you open the passage, you decide. Proposals stay proposals until a human accepts them.

This week's drill, no signup needed. Take one AI-assisted document you or your team shipped recently and pick the three claims that would hurt most if they were wrong. Trace each one to its source. If all three trace cleanly, you have a review habit worth keeping. If one does not, you have found the same gap Deloitte found, at a much better price.