The ninety-seven-thousand-dollar skip
In October 2025, Deloitte's Australian arm delivered a A$440,000 report to the Department of Employment and Workplace Relations — an assurance review of welfare-compliance systems. It shipped with fabricated academic references, experts who don't exist, and an invented quote attributed to a federal-court judgment. The fabrications weren't caught by the engagement's review chain. They were caught after publication by a university researcher, Chris Rudge, reading the report. Deloitte disclosed in a revised version that generative AI (Azure OpenAI GPT-4o) had been used, and repaid the final instalment: A$97,587.11. (Some headlines quote larger US-dollar figures — those convert the whole contract, not the refund.)
Every retelling of this story frames it as an AI failure. Read it as a PM instead, because on a delivery timeline it looks like this:
- A deliverable was produced (partly by AI — fine, that's the new normal).
- It went through whatever review it went through.
- It shipped.
- An outsider ran the check the delivery team never did: open the citations.
Step 4 is the indictment. "Every cited source exists and says what the report claims" is not an exotic quality bar — it's a fifteen-minutes-per-source acceptance check, mechanical enough that a junior could run it with a checklist. Nobody ran it, because it wasn't defined as a check. The review that did happen was the review that's always happened: senior people reading for argument, structure, tone. Polish review. And AI-fabricated citations sail through polish review, because fabricated content is fluent by construction. That's what generation is.
Why the check has to exist before the build
Here's the subtle part. If you define acceptance checks after seeing the deliverable, you unconsciously define checks the deliverable passes. The document looks complete, so "completeness" feels covered. The citations look real — they have authors, years, plausible journals — so "sources check out" feels covered. Post-hoc review inherits the output's blind spots. Checks written at spec time, before any output exists to anchor you, are the only checks that test what you actually needed rather than what you visibly got.
One more thing the case teaches: there was no regulator to catch this. Consulting-grade delivery work carries no license, no statutory review board — accountability ran entirely through the contract, which is exactly how the money came back. When the enforcement mechanism is contractual, the acceptance check is the compliance function. Nobody upstream of the client is coming to save the deliverable.
This lesson builds the gate: checks defined with the spec, evidence recorded per check, a verdict computed from the evidence — and "looks polished" nowhere in the function.