Western Australia seeks refund from Deloitte over AI errors in A$166,000 report

The state government of Western Australia announced on October 25, 2026, that it is seeking an expedited refund from Deloitte over an independent review of public university structures that cost A$166,000. An investigation revealed that the final deliverable contained AI-generated citations to non-existent studies, as well as cost discrepancies amounting to hundreds of millions of dollars.

Deloitte acknowledged using generative AI tools during the drafting and research phases of the assignment, describing the findings as minor referencing errors that it intends to correct in a revised submission.

What happened

In late 2025, the Western Australian state government commissioned Deloitte to deliver an independent assessment on the viability of merging several public universities. When the report was submitted, analysts and journalists reviewing the appendices found that multiple academic papers cited in the bibliography did not exist. At least one reference included an unedited system tag from a consumer chatbot interface.

Beyond fabricated references, mathematical inconsistencies appeared in the fiscal modeling sections. Cost estimates for institutional integration reported in summary tables diverged from underlying appendices by more than A$200 million. Following these disclosures, state officials stated they were extremely disappointed with the quality of the deliverable and initiated formal procedures to recover the taxpayer funds paid for the contract.

Why it matters for business

This incident marks a turning point in how corporate clients treat AI-assisted professional services. For years, advisory firms have incorporated automated tools into research workflows without explicit client disclosure or mandatory source-level verification.

When an advisory firm delivers work with unverified AI hallucinations, the damage extends beyond embarrassment:

1. Direct financial recovery: Clients are no longer accepting corrections or apologies. They are invoking breach-of-contract clauses to demand full refunds.
2. Reputational and evidentiary exposure: Once AI hallucinations appear in public filings or regulatory submissions, opposing parties and audit committees can challenge every finding in the document.
3. The liability of unreviewed output: The court or client holds the signing partner responsible for the entire text, regardless of whether a junior associate or an automated model drafted the initial paragraphs.

How to apply to day-to-day work

Organizations do not need to ban AI tools to prevent citation and calculation failures. They need strict verification boundaries that no deadline is allowed to override:

• Enforce the human citation rule: Never allow a document to leave your organization if citations, case references, or external quotes were not manually retrieved and verified against primary sources by a named team member.
• Isolate calculations from generative text models: Generative language models predict text sequences; they do not calculate reliably. Keep financial projections, payroll models, and balance sheets inside verified spreadsheets or deterministic tools rather than asking a language model to total rows in prose.
• Document AI assistance in client agreements: Clarify upfront where automated systems assist your team and where human review begins. Transparency builds client trust and protects your firm from breach claims.