A stitch in time

A Big Four firm billed a government nearly half a million Australian dollars for a report quoting a judgment that was never written. Nobody inside the firm caught it. A university researcher reading it for other reasons did.

Deloitte Australia delivered a 237-page assurance review to the Department of Employment and Workplace Relations. It examined the IT system that automates penalties in the country's welfare programme.

The report quoted a Federal Court judgment. The judge had never said it.

It cited academic papers by named professors at the University of Sydney and at Lund University in Sweden. The papers did not exist.

The contract was worth A$440,000. Deloitte returned A$97,000. The corrected report added one sentence nobody had seen before, confirming that Azure OpenAI had been used in the drafting.

Now hold on to one detail. The review passed through a Big Four quality process, was signed off, was delivered to a government client, and was published on a department website in July.

Nobody in that chain found the fabricated quote. It was found by Chris Rudge, a welfare law researcher at Sydney University, who was reading the report for his own reasons and noticed the references pointed at nothing.

The check did not come from inside the process. It came from a stranger, and by accident.

The same thing is happening in Washington

In August, Politico reported on the small office in the House of Representatives whose entire job is turning what a member of Congress wants into language that will hold as law.

Bills are arriving there drafted by chatbots. Statutes cited that do not say what the draft claims. Definitions that do not mean what the member intended.

Money routed as a tax credit when it was meant to be a deduction.

One congressional staffer described it in two words. “Absolutely terrifying.”

But the error rate is not the story. This is the story: one adviser said the office now spends more time fixing AI-drafted legislation than it would take to write it from scratch. AI did not reduce the work. It moved the work downstream, to someone who never asked for it.

These are the places with the reputation and the responsibility to check. That is what should worry you.

A Big Four assurance review and a congressional drafting office are not careless environments. They are built for suspicion. And it still got through.

If it can get through there, it can get anywhere.

Now look at your own floor

Those places have something your business does not. An answer to check against. A judgment says a thing or it does not. A paper exists or it does not. Someone can look it up, and eventually someone did.

Most business decisions have no such thing.

Take pricing. Cost plus. Value based. Competitive. Blended. Four defensible formulas and no correct one, which is exactly the problem. Once somebody picks, it stops being a choice and becomes a system. It flows into quotes, into margin targets, into forecasts, into next year's plan. Nobody rechecks it, because there was never a right answer to check it against.

And whether you have looked at this or not, your teams are already working this way. Ideas, decisions and policies drafted by an AI model and passed upward as work. Not because anyone approved it, but because it is faster.

Intelligence without skin in the game

These tools are exceptionally powerful. They compare things faster than any team I have worked with. I have spent three years helping organisations install them and I would do it again tomorrow.

They have one serious limitation. They are bound by the direction they are given, and once they take it, they do not leave it. Feed it one faulty number and it carries that number forward as fact and important decisions can flow from it. Machine cannot be held responsible for any of the consequences .

Intelligence without skin in the game is not intelligence. It is brute force of logic. And logic is a double-edged sword. One of those edges is always facing you.

A stitch in time

The proverb survives because small failures are cheap to fix and ruinous to leave.

Every policy, decision and document produced in your organisation this week has AI somewhere in the chain. Somewhere in that chain is the gap that needs a stitch.

Leave the one stitch today and you may not get to make the nine. It does not grow linearly: one, two, three. It grows exponentially: two, four, eight, sixteen, thirty-two.

Give it twelve more months and your documents will be so thoroughly mixed with model output that nobody will be able to separate the human part from the rest. Not the auditor. Not the regulator. Not you.

Deloitte got caught because somebody outside happened to read 237 pages closely. That is not a control. That is luck.

Which decisions in your organisation are machine-assisted, and which are machine-made?

Sources: Associated Press, 7 October 2025 / Financial Times, October 2025 / CFO Dive, 21 October 2025 / Politico, 17 August 2026, and secondary coverage in Gizmodo and TaxProf Blog / Legal Data Hunter primary-source report, April 2026.