AI-Generated Report Errors: Deloitte Case ⚡ (2026)
Learn about AI-generated report errors in Deloitte’s 2026 case: fake references, wrong sums, AI hallucinations—and AI report governance lessons for responsible oversight.
When an AI-Aided Consulting Report Gets the Basics Wrong
A report commissioned to assess a possible merger of three Western Australian universities has left its authors with a problem - and it offers a sharp warning for any organisation using generative AI in professional work. Deloitte's cost-benefit analysis included references to sources that do not exist and, according to figures reported by Guardian Australia, an incorrect cost total in its appendix. Western Australia's government is now seeking a refund for the A$166,000 study.
This isn't just a story about whether an AI tool made a mistake. Deloitte said generative AI helped gather research, summarise information and proofread the report. The more important question is what happened after those outputs were produced: who checked the references, who reconciled the figures, and what reassurance the client received before the report went out.
The takeaway here is practical. It's about oversight - how consulting and other knowledge work should be set up so that errors don't slip through simply because they look polished. It also shows why a disclosure isn't the same thing as quality control, and why appendix-level errors can still matter even if the headline modelling is said to be unchanged. Finally, it's a reminder that clients should be able to ask what will be checked, not just what tools were used.
What happened in the Deloitte report
Deloitte prepared a cost-benefit analysis for a proposal to merge the University of Western Australia, Curtin University and Murdoch University. The Western Australian government paid A$166,000 for the study.
The report disclosed that Deloitte used generative AI for research gathering, summarising information and proofreading. That matters because it tells readers something about how the report was made. But disclosure doesn't answer the obvious follow-up questions: was the work checked properly, and was the checking done at the right level?
The first issues raised publicly were references to sources critics said could not be found. Times Higher Education reported that the problem was flagged in October, beginning with a Reddit post. Deloitte described the references as minor errors, and the government asked the firm to correct and reissue the report.
Then more scrutiny came. Guardian Australia reported incorrect sums in the report's appendix. One example it gave was that costs associated with a Curtin and UWA merger were listed as A$399 million, while it said they should total A$659 million. This article has not independently verified the appendix figures; it relies on Guardian Australia's reporting for that discrepancy.
A government spokesperson said the state was "very disappointed" and had written to Deloitte to speed up a refund. The spokesperson also said the government was told the errors didn't change the report's modelling or findings, and that Deloitte would correct and reissue it. Deloitte did not respond to Guardian Australia's questions, according to the publication.
Taken together, those details point to two problems. First, the deliverable: the report included fabricated references and arithmetic errors. Second, the relationship with the client: the government is looking for its money back, while the firm is expected to correct what was delivered. The source material does not confirm the final refund amount or the final status of the revised report, so neither should be assumed.
A disclosed tool is not a verified result
It's tempting to focus on the headline fact that AI was used. That's relevant - but it doesn't explain the full failure. Generative AI can write convincingly, while still getting details wrong, including citations that look legitimate but don't exist. It can also summarise text without guaranteeing that the summary stays faithful to what's actually in the source.
In a professional report, the publisher is still responsible for the claims it puts forward, even if an individual, a model, or a combination produced the first draft. The tool describes part of the workflow, not the quality of the final output. It doesn't tell a client whether every citation was checked, whether figures were recalculated, or whether an accountable subject-matter expert reviewed the work.
That distinction matters for responsible AI use. This isn't only "Was AI involved?" It's "What checks were applied to the parts of the work where errors would be expensive?"
Why factual errors matter even when findings remain unchanged
The Western Australian government was reportedly told that the errors didn't change the report's modelling or findings. That might be correct. But it doesn't make the errors harmless.
A cost-benefit analysis isn't just a conclusion. It's also an evidentiary record: assumptions, inputs, sources and calculations that let a client understand how the conclusion was reached. If a reference can't be found, the reader can't easily verify the evidence behind the claim. If a total doesn't match its component costs, the reader can't be confident that the arithmetic was handled correctly without redoing the work.
That creates a trust problem that's separate from whether the headline recommendation changed. A reasonable reader would ask: if one number is wrong, have other numbers been checked too? If a citation was fabricated, are the other sources reliable? The existence of one error doesn't prove the whole report is broken. But it does mean confidence needs to be rebuilt through verification, not waved through because the "main result" is unchanged.
Errors create work for the client
A commissioned report is supposed to reduce the client's burden. When the client - and/or journalists, researchers or other reviewers - has to go hunting for sources and retesting calculations, some of that burden comes right back.
This isn't limited to fixing a single line in a document. It can mean time spent locating and checking references, rerunning calculations, reviewing other sections, briefing decision-makers, and deciding whether the revised version can be trusted. When a report feeds into public policy, errors can also pull attention away from the actual policy question the report was meant to address.
So "the findings are unchanged" isn't a complete answer. It addresses only one question: whether the reported errors affect the result. It doesn't address how the errors happened, whether the whole report was rechecked, or what level of assurance the client should reasonably expect for the corrected version.
Accuracy is part of the product
Consulting work is often judged by conclusions, but a credible conclusion depends on the work behind it. A clean-looking document with weak source control isn't the same thing as a verified analysis. References and calculations aren't decoration. They're part of what the client paid for.
This becomes especially relevant when AI is used to speed up drafting. If an organisation saves time writing, then spends that time later chasing invented citations and reconciling basic totals, the workflow hasn't necessarily become more efficient. It may have simply shifted labour from production to repair - and made that repair harder to scope, because the errors can be difficult to spot at a glance.
So the business case for AI should include the cost of review and correction. Speed at the drafting stage is only one slice. The more useful question is whether the full end-to-end process produces reliable work with less total effort and an acceptable risk level.
The real issue is the workflow, not just the AI
Generative AI isn't a guarantee of accuracy, and it shouldn't be treated like one. But this case shouldn't be reduced to "any use of AI makes a report untrustworthy." The real issue is whether the work was organised so that the tool's weak spots were caught before the report reached the client.
An assistant using AI to draft a rough summary creates a different risk than an author relying on AI-generated references without checking them. A model can help organise information, but it can't make a non-existent source real. It can calculate or restate a number, but a professional workflow still has to verify inputs and recalculate arithmetic where it matters.
Match controls to the task
Not every AI-assisted task needs the same level of scrutiny. Editing an internal draft for tone isn't the same as producing a public cost-benefit analysis. A workable control system starts with the intended use of the output, and the real-world consequences if it's wrong.
For a report that may inform institutional or government decisions, basic controls should include:
- Source verification: Each cited source should be opened and checked against the claim it supports.
- Calculation checks: Important totals should be reproducible from visible inputs, with independent recalculation for material figures.
- Traceable assumptions: The report should spell out which assumptions drive the analysis and where they came from.
- Human accountability: A named reviewer or responsible team should own the final work, not just rubber-stamp that an AI tool was used.
- Version control: Corrections should be documented so the client can see what changed and why.
These aren't "AI-only" requirements. They're normal quality controls - made more important when plausible-sounding errors can be introduced and carried forward quickly.
Disclose use, but make disclosure specific
Saying that generative AI was used is a good start. It's still not enough for a client trying to judge risk. A more useful explanation distinguishes between tasks: was AI used to search for sources, summarise material, draft passages, proofread, or analyse data? Which outputs were checked against original material? Were any figures generated or transformed with AI assistance?
This doesn't mean every report needs pages of technical documentation. The detail level should fit the importance of the deliverable. Still, clients should be able to understand where AI contributed and what verification was performed. Otherwise, disclosure becomes closer to a disclaimer than an account of how the work was done.
The same applies inside organisations. A vague statement like "AI may be used responsibly" leaves key decisions to individual workers. Teams need clear expectations for higher-risk tasks: what tools are allowed, what data can be entered, what evidence must be kept, and which outputs need independent checking.
A practical framework for clients and consulting firms
The Deloitte case points to a simple operating framework: define acceptable use before work begins, build verification into the delivery process, and spell out what happens if errors are found. Arguing about whether AI should be allowed in consulting tends to miss the point.
Before the work starts
Clients can include AI-related questions in the project brief or contract discussion. The goal isn't automatically to ban a tool. It's to remove ambiguity about responsibility and quality.
Ask the provider to explain:
- Which parts of the work may involve generative AI.
- Whether AI-generated research, summaries, calculations or citations will be independently verified.
- How the team will handle confidential or sensitive client information.
- Who is accountable for checking the final report.
- How corrections will be handled if factual errors are identified after delivery.
These questions are useful even if the provider says AI won't be used. They help establish an expected standard of care and make it easier to compare proposals beyond just cost and turnaround time.
During production and review
The project team should keep the evidence trail close to the output. For each material claim, reviewers should be able to find the supporting source. For each important total, they should be able to see the inputs and repeat the calculation. If AI has drafted a summary, the reviewer should compare it against the underlying source rather than treating fluent prose as proof.
Review should also be proportionate. Checking every sentence with the same intensity can be wasteful and inefficient. Focus first on claims that affect the conclusion, figures that drive the economics, and citations supporting contested or consequential statements. Then sample lower-risk material to see whether the controls are working consistently.
This isn't an excuse to ignore the rest of the document. It's a way to use limited review time where errors would do the most damage.
When a problem is found
A correction process should make clear what went wrong, how wide the issue may be, and what the client needs to do next. Fixing a faulty citation or correcting a total may be part of it - but the response should also say whether related references and calculations were checked again.
A practical correction note can cover:
- What was wrong and where it appeared.
- Whether the error affected any other figures, assumptions or conclusions.
- What checks were repeated elsewhere in the report.
- Which version is now authoritative.
- What remediation, including a refund where applicable, is being offered.
That approach helps both sides. The client gets a clearer basis for deciding whether to rely on the revised work. The provider can also show that the fix addresses the underlying quality issue, not just the visible example.
The commercial context matters, too. Deloitte previously agreed to partly refund Australia's federal government for a separate A$440,000 report in 2025 that also contained AI-generated errors, according to the source material. Rival firms KPMG and PwC have also faced problems this year involving AI-written reports. These examples don't prove that every consulting firm uses the same controls, or that all AI-assisted work is defective. They do reinforce the point that verification and remediation should be treated as procurement issues, not as an afterthought.
Conclusion
This Western Australian report is a reminder that AI can speed up parts of professional work without taking responsibility for the result. Deloitte disclosed generative AI use, but the report still contained non-existent references and, according to Guardian Australia, an incorrect cost total. The state is seeking a refund, and the firm is expected to correct and reissue the report.
For organisations adopting AI, the practical lesson is to treat verification as part of production - not optional cleanup after delivery. Set expectations up front, check consequential evidence and calculations, and require a clear account of any correction. That won't eliminate every error, but it makes errors easier to catch, contain and fix.
If you commission AI-assisted research or analysis, review your project requirements now: define what must be verified, who owns final sign-off, and what happens when the work is wrong. Asking whether a tool was used is the easy question. Building a process that catches mistakes is the important one.