Deloitte handed back A$97,587 on its A$440,000 government review after a law researcher found citations to work that does not exist. For a regulated firm, the exposure sits in every vendor document already in its evidence file.
A regulated firm runs on documents other people wrote. The penetration test report, the SOC 2 report, the model validation, the consultant's assessment of the BSA program, the policy review outside counsel marked up. Each one lands in an evidence file, and an examiner reads it as your evidence, not the vendor's. When the vendor drafted part of it with a language model and skipped the check, the checking did not disappear. It moved to you.
So treat AI use inside a deliverable as one more link in the vendor's supply chain, which banking guidance already expects you to watch, and hold it to three tests: disclosed in the statement of work, traceable to its sources, and sample-verified before the document enters the file. The Deloitte case in Australia is the clearest public record of what happens when nobody does any of the three.
A$440,000, up to twenty errors, and one reader nobody paid
Australia's Department of Employment and Workplace Relations engaged Deloitte under an order for services dated 23 December 2024 to review the Targeted Compliance Framework and the IT system that runs it, which automates penalties for job seekers on income support who miss their obligations. The final report is dated 4 July 2025. In August, the Australian Financial Review reported that Christopher Rudge, a law researcher at the University of Sydney, had found references in it to academic work that does not exist. Rudge said he found up to 20 errors in the first version, according to the Associated Press. Among them was a quotation attributed to a Federal Court judgment in Amato v Commonwealth, the 2019 Robodebt case, that the court never wrote.
The revised report, dated 26 September 2025, was posted on 3 October. Its update note says it corrects "those citations and reference list entries which contained errors in the previously issued version" and amends "the summary of the Amato proceeding which contained errors." Its methodology section now discloses that part of the technical review "included the use of a generative AI large language model (Azure OpenAI GPT-4o) based tool chain licensed by DEWR and hosted on DEWR's Azure tenancy."
Note who owned the tool chain. It belonged to the client, licensed by the department and running in the department's own Azure tenancy. Supplying the tool did not tell the department what the tool had produced.
The department said Deloitte would repay the final instalment under the contract. At Senate estimates in October 2025, the refund was put at A$97,587, and Finance officials said they had learned of the errors through media reporting, according to the Australian Greens' release and PS News coverage of the hearings. On 14 October, Deloitte Australia's chief executive, Joanne Gorton, wrote to the department's secretary in a letter the Senate committee later published: "We take responsibility for the fact that appropriate review and oversight processes were not followed on this occasion."
That sentence is the whole case. The model did what models do. The review step failed, and the review that mattered came from an academic who spotted a book credited to a University of Sydney colleague that does not exist.
A finished-looking document moves the checking to whoever receives it
BetterUp Labs and the Stanford Social Media Lab named the pattern in Harvard Business Review in September 2025. They called it workslop: "AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task." Their point was the direction of the cost. The sender saves the time. The receiver spends it, reconstructing what the sender never checked.
A consulting review is the extreme case, because its entire value is that someone else did the reading. When the department posted the corrected report, its own page described the change as "a small number of corrections to references and footnotes." Deloitte's update note says the corrections "in no way impact or affect the substantive content, findings and recommendations." That may be true. Nobody outside the drafting team can confirm it without redoing the work, which is precisely the work the department paid A$440,000 to avoid.
The department's secretary, Natalie James, told estimates, as PS News reported: "My people should not be double-checking a third-party provider's footnotes." As a statement about what a client deserves from a vendor, she is right. As a statement about where accountability sits in a regulated firm, it does not hold. In a bank or a credit union, your people do double-check, or you explain to an examiner why they did not. A fintech that serves them hears the same question in its partners' diligence.
The research on how these errors behave makes sampling more urgent, not less. In a field experiment with 758 Boston Consulting Group consultants, published in Organization Science in March 2026, those using GPT-4 on tasks inside the model's range worked roughly 25 percent faster and produced work rated about 32 percent higher in quality, as the Harvard Business School AI Institute summarizes it. On one task built to sit just outside that range, 84.5 percent of consultants without AI reached the correct answer. With GPT-4, 70.6 percent did. With GPT-4 plus a prompt engineering overview, 60 percent did. The best-equipped group did worst on exactly the task the tool could not handle, and nothing in a finished document announces which kind of task produced it.
The deliverable is evidence, and evidence has a supply chain
US banking regulators answered whose problem this is without ever mentioning AI. The Interagency Guidance on Third-Party Relationships: Risk Management, issued by the Federal Reserve, the FDIC, and the OCC and published in the Federal Register on June 9, 2023, states that "a banking organization's use of third parties does not diminish its responsibility to meet these requirements to the same extent as if its activities were performed by the banking organization in-house." For purposes of that sentence, a penetration test report you file is your penetration test. A consultant's review of your BSA program is your review. The guidance binds banks directly, and everyone else in the chain meets it through a bank's diligence.
Courts reached the same place for lawyers. In July 2025, Judge Nina Y. Wang of the US District Court for the District of Colorado fined two of Mike Lindell's lawyers $3,000 each over a brief with nearly 30 defective citations, including cases that do not exist, as the ABA Journal reported. Their explanation, that a draft had been filed by mistake, did not persuade her. The name on the filing carried the responsibility, whatever produced the text.
The guidance also defines the supply chain broadly. A footnote says third parties "may enlist the help of suppliers, service providers, or other organizations, which this guidance collectively refers to as subcontractors." It then lists "the third party's reliance on, exposure to, and use of subcontractors" among the factors to watch in ongoing monitoring. The guidance never names generative AI. My reading, and it is a reading rather than a regulator's stated position, is that a model provider or an AI tool chain inside a vendor's drafting process fits that definition closely enough that I would not want to argue the opposite in an exam.
The Deloitte case adds a wrinkle worth planning for. The tool chain was the client's. A disclosure clause that only asks about the vendor's own subcontractors would have missed it, because the department licensed the tool and Deloitte simply used it. Your clause has to cover any AI tool used to produce the deliverable, including the ones you supplied.
The same logic runs in the other direction. SavvyMoney sits behind more than 1,500 financial institutions, so the same argument applies to us as the third party in someone else's diligence. The clauses below are ones a vendor should expect to sign, not only to send.
Put AI use in the statement of work
The guidance's contract section already holds the template. On subcontracting, it says a banking organization "may want to address when and how the third party should notify the banking organization of its use or intent to use a subcontractor." On performance measures, it warns against terms "encouraging processing volume or speed without regard for accuracy." Those two sentences translate into four clauses for any deliverable that will enter an evidence file.
- Disclosure at signing and in the deliverable. The vendor states up front which AI tools it intends to use, who hosts them, and whether your data will enter them, then states in the deliverable which sections a tool drafted or analyzed. A disclosure that first appears in a corrected version, as it did here, arrives too late to shape your intake review.
- Every citation resolvable in minutes. Case name and number with a paragraph pinpoint, statute section, regulation citation, document title and page. Every direct quotation verbatim and pinpointed. An unresolvable reference counts as a defect, the same as a wrong control description.
- A named human verifier who is not the drafter. That person attests to checking every citation and quoted figure against its source. A name gives the verification step an owner instead of a process diagram.
- Payment tied to verification. Hold back the final payment until the deliverable passes your intake check, and state that a fabricated citation triggers correction at the vendor's cost. Deloitte's refund was the final instalment under the contract, according to the department. That was a holdback by accident. Write one in on purpose.
Most contracts also need a data clause: your data goes only into tools on an approved list. Whether a vendor trains on that data is the wrong first question. Where the data travels and where it rests is the right one, and that argument applies to consultants as much as to software vendors.
Sample-verify before it enters the file
Clauses change incentives. They do not verify anything. Verification is an intake step on your side, and it costs less than it sounds, because a fabricated citation fails in a specific, checkable way: the source does not exist, or it exists and does not say what the document claims.
- Sample before you accept. Pull ten citations or ten percent, whichever is larger, weighted toward legal and regulatory references and any quotation that supports a finding. Resolve each one to its source and read the cited passage.
- Re-perform a slice of the substance. For a penetration test, reproduce two or three findings from the report's own steps. For a control attestation, trace a few sampled control statements to the evidence the vendor says it examined. For a model validation, rerun one stated test on the data the report names.
- Fail the document, not the line. One fabricated source sends the whole deliverable back for a full citation review by the vendor. Fixing the single error in place tells you nothing about the ones you did not sample.
- File the check with the document. Record who sampled, which items, what each resolved to, and the outcome, and store it next to the deliverable.
The last step is the one teams skip, and it is what an examiner will ask for. A deliverable with an intake record attached shows a firm that read what it filed. A deliverable without one shows a firm relying entirely on its vendor's review process, and the Deloitte letter is a vendor stating in writing that its process was not followed. Producing that record as the work happens, rather than reconstructing it later, is the same discipline behind a pipeline built to survive an audit for your own systems.
Start with the documents already in the file
The clauses fix the next contract. The exposure sits in the documents already filed. On Monday, pull every third-party deliverable that entered an evidence file in the last twelve months: test reports, attestations, validations, consulting reviews. Send each vendor one written question: "Was any AI tool used to research, draft, or analyze this deliverable, and if so, which tool, hosted where, and who verified its output?" Sample-verify the deliverables whose vendors answer yes and the ones whose vendors do not answer. Put the four clauses into the next statement of work that comes up for renewal.
Then file the vendor answers alongside the deliverables. If an examiner asks how you know your evidence is real, the answer should be better than the one the department had: an academic in Sydney read it.