Deloitte's AI Report Refund: A$97,000 Back, 5 Cases, 3 Clauses to Copy

Deloitte refunded about A$97,000 on a report Australia's government paid close to A$440,000 for, after fabricated citations and a made-up court quote were found in it. That is roughly 22% of the fee. The Deloitte AI report refund is the only confirmed cash refund among five public cases of AI generated report errors we could document. In every case, someone outside the client found the problem first.

The short answer. Deloitte repaid over A$97,000, the final instalment of an A$440,000 contract, after Australia's welfare department found fabricated references in its review. Across five documented cases, refunds were small or absent because the client accepted that the findings still stood. Buyers who want real remedies need a citation warranty and a fee holdback written before the work starts.

What the Deloitte AI report refund actually returned

The Department of Employment and Workplace Relations hired Deloitte to review the Targeted Compliance Framework. That is the IT system that automates penalties for welfare recipients. The contract was executed in December 2024 and valued at about A$440,000, according to the AusTender notice cited by CFO Dive.

The report was published in July 2025. University of Sydney academic Dr Christopher Rudge then found citations to reports that do not exist, plus a fabricated quote from a Federal Court judgment. The Australian Financial Review carried his findings in August 2025.

A revised version went up in early October 2025. The Register reported that it stripped out more than a dozen bogus references and footnotes. It also added a disclosure: Deloitte had used "a generative AI large language model (Azure OpenAI GPT-4o) based tool chain licensed by DEWR and hosted on DEWR's Azure tenancy."

The refund came later that month. A DEWR spokesperson told CFO Dive that Deloitte had refunded over A$97,000. On a A$440,000 contract that is 22%. The department said the substance of the review was retained, with no changes to the recommendations.

The number I would not celebrate. Coverage treated the refund as accountability. I read it the other way. The client paid A$343,000 for a report it now knows contained invented sources, and the supplier kept that money because the conclusions survived.

That outcome was not negotiated from a clause. It was a goodwill gesture under political pressure. A private buyer without a Senate committee behind it should expect less, which is the gap examined in the breakdown of what AI contracts let buyers recover.

The case ledger: five consulting errors on the record

We collected every case we could verify where a consultancy or Big Four firm published or delivered AI-assisted work with fabricated content. The test for inclusion was strict. The error had to be documented by a named outlet, and the firm or the client had to respond on the record.

Three cases are paid government work. Two are thought leadership the firms published themselves. We kept those two because they show the same failure without a client in the loop to absorb it.

Dates are when the errors became public, not when the work was delivered. The GPTZero counts are the vendor's own classification. Two government documents with fabricated citations, Newfoundland's Education Accord and South Africa's draft national AI policy, are excluded because they were written in house, not by consultants.

Why three of five are Deloitte. Do not read the table as a Deloitte problem. Deloitte has the most public sector advisory work in these jurisdictions, and public sector work is the only kind that gets audited in public. A private client that finds a fake citation in a strategy deck has every reason to stay quiet.

The ledger is a sample of what became visible. It is not a measure of how often this happens. The real rate is unknown, and anyone quoting one is guessing.

What the two thought leadership cases add. EY Canada's report on cyber threats in loyalty systems was withdrawn in May 2026. GPTZero, which sells AI detection and citation checking, found that 16 of 27 citations were fabricated, misattributed or broken. EY said it was reviewing the circumstances of publication.

KPMG International pulled a report on agentic AI in customer experience in June 2026. GPTZero found only 5 of 45 citations led correctly to their source. UBS called the case study about it "factually incorrect," according to The Register.

These matter for a buyer because they are the firms' own marketing. If a firm's flagship research ships with invented sources, its client deliverables run through the same production line with less outside scrutiny.

Who caught the errors, and it was never the client

This is the finding I think buyers should sit with. In all five cases, the person who found the fabrication had no contract with the firm and no role in accepting the work.

The DEWR report surfaced within about a month, because one specialist reader knew the cited literature personally. The Newfoundland plan sat for roughly six months. The KPMG report sat for about eight. Each of those reports had already informed decisions by the time anyone checked.

Your acceptance process probably looks like these clients' processes. Someone reads the executive summary, checks the recommendations against the brief, and signs. Nobody opens the footnotes. That is the gap a hallucinating model exploits, and it is the same gap tested in the audit of deep research tools that invent sources.

What each client did about AI generated report errors

The responses split into three kinds. Some clients asked for money, some changed procurement rules, and some kept buying. Several did all three at once.

Australia: a refund, then new rules

Beyond the refund, Department of Finance officials told a Senate committee in December 2025 that they were reviewing Deloitte Australia's listing as an approved supplier. Bloomberg Tax reported the hearing. We found no published outcome of that review.

The bigger change was policy. The APS AI Plan, released by Public Service Minister Katy Gallagher in November 2025, requires suppliers on the management advisory and people panels to declare planned AI use when they quote. New clauses state that consultants remain fully responsible for their services "regardless of whether generative AI is used," according to the Canberra Times.

Newfoundland and Labrador: disclosure rules and a pending claim

"AI was not used to write the report; it was selectively used to support a small number of research citations," Deloitte Canada told Fortune. It said it stood behind the recommendations and would correct the citations.

By March 2026 the province had changed procurement so vendors must disclose whether and how they intend to use AI. A health department spokesperson said the government was "reviewing options related to compensation," the Canadian Press reported. CPA Newfoundland and Labrador opened an investigation after a citizen complaint the same month.

Canada's federal government: keep buying. Employment and Social Development Canada had awarded Deloitte a contract worth up to C$1.1 million for advice on deploying AI. In December 2025 the department defended it. It said that if Deloitte did not respect the contract's terms, "appropriate action will be taken." That sentence only bites if the terms say something about AI.

Western Australia: recovery requested, findings accepted

The WA government paid A$166,000 for a cost-benefit study on merging three universities. Guardian Australia found one scenario where A$514m plus A$145m was totalled as A$399m, a shortfall of A$260m. In another, A$396m plus A$116m became A$823m. A spokesperson said the state was "extremely disappointed" and had been advised the errors "do not affect the modelling." The state wrote to Deloitte seeking expedited cost recovery.

The A$311m overstatement is about 1,870 times the fee. That ratio is our own arithmetic, and it is the clearest illustration of why a fee-capped remedy is the wrong frame for advisory work.

Why the refund equalled the last instalment

The DEWR refund was the final payment under the contract. That is not a coincidence. When a supplier has been paid most of its fee, the only money the client controls is whatever it has not yet released.

Look at the Newfoundland contract. The province paid C$1,598,485 in eight instalments over two years, and the plan was published in May 2025. By the time The Independent found the citations in November, the money was gone. The province is now "reviewing options," which is what a buyer says when it has no lever.

The findings defence closes the gap. Every paid case shares one sentence. DEWR said the substance was retained. Deloitte Canada said the corrections "do not impact the report findings." WA was advised the errors did not affect the modelling. Each time, the client accepted that framing.

I think that framing is wrong. A conclusion supported by a nonexistent source is a conclusion supported by nothing at that point. It may still be right, but the client paid for the support, not just the conclusion. Most consulting contracts only define a defect against the deliverable's purpose, so the supplier wins the argument by default.

This is the same structural problem covered in the CFO clauses for AI vendor contracts. The remedy exists only where the contract defines the failure in terms the supplier cannot reinterpret.

Three contract clauses for AI quality assurance in consulting work

These are written for a buyer commissioning research, reviews or strategy work from any adviser. Adapt them with your own counsel. They are drafting starting points, not legal advice, and the defined terms will need to match your master agreement.

Clause 1: AI use disclosure and approval

"Before commencing the Services, the Supplier must disclose in writing each generative AI system it will use, the tasks each system will perform, and any Buyer data each system will process. The Supplier must not use a generative AI system for any undisclosed task without the Buyer's prior written approval. Each Deliverable must include a methodology statement repeating this disclosure."

This mirrors the notice and approval model in the Digital Transformation Agency's AI model clauses, which King and Wood Mallesons summarised in August 2025. The point is not to ban AI. The point is that a disclosure made only after the errors surface, as in the DEWR case, is a confession rather than a control.

Clause 2: source verification warranty

"The Supplier warrants that every source cited in a Deliverable exists, is accessible to the Buyer, and supports the statement for which it is cited. With each Deliverable the Supplier must provide a citation register listing each source, a working link or copy, and the name of the person who verified it. Any citation that breaches this warranty is a Defect, whether or not it affects the Deliverable's findings or recommendations."

The last sentence carries the weight. It removes the findings defence that kept every paid case in this ledger cheap for the supplier. The register also makes a buyer's spot check fast: pick 10 citations, open them, and you know within an hour whether anyone checked.

Clause 3: defect remedy and fee retention

"The Buyer will retain the Retention Amount until the later of 90 days after acceptance and 90 days after publication of the Deliverable. If a Defect is notified within 12 months of acceptance, the Supplier must correct and reissue the Deliverable at its own cost within 10 business days. The Supplier must also pay the Buyer the Defect Refund set out in the Schedule. The Defect Refund is not limited to fees that remain unpaid."

Set the Retention Amount at a level that the supplier would notice. The time limits above are my suggested defaults, not an industry standard. The DEWR, Newfoundland and KPMG errors surfaced between one and eight months after publication, so a 12 month notification window covers the cases on the record.

Where this argument is weakest

Five cases are a small base, and three involve one firm. The ledger tells you how these failures surface and how clients respond. It cannot tell you how often consulting deliverables contain fabricated content, and it would be dishonest to imply otherwise.

The findings may genuinely stand

The clients in the three paid cases were advised that the conclusions held, and nobody has published evidence that they do not. A strategy review can be directionally right with bad footnotes. If your interest is the decision rather than the document, a large refund may matter less to you than a fast correction.

AI attribution is partly contested

Deloitte's own internal review of the DEWR report reportedly blamed human error rather than AI misuse, according to The Register's account of AFR reporting. In Newfoundland, Deloitte said AI supported only a small number of citations. Fabricated references predate language models. The clauses above are written to work either way, which is deliberate.

Suppliers will price the risk

A citation warranty and a long retention period move risk to the supplier, and suppliers charge for risk. Expect higher quotes or pushback on the clauses, especially from smaller firms. My view is that the price is worth paying for any report that will inform a public or board decision, and not worth paying for internal brainstorming material.

Frequently asked questions

Where to start this week

Pull the last three external reports your team paid for and check 10 citations in each. Open every link and confirm the source says what the report claims. If all 30 hold, you have a baseline. If any fail, you have a conversation to have before renewal, and you will have found it before a journalist did.

Then send the source verification warranty to whoever owns your consulting master agreement, and ask whether it can go into the next statement of work. It is the cheapest of the three clauses to adopt, and it is the one that would have changed the most outcomes in this ledger.

Related reading