After the 2025 Australian government report refund, should Deloitte require human review of every AI-generated passage in client deliverables, even if it slows delivery?

Introduction

In October 2025, Deloitte Australia agreed to partially refund the Australian government for a report, worth about A$440,000, after errors including incorrect and non-existent references were found; a revised version disclosed that generative AI had been used in its preparation. The episode drew wide attention to how professional services firms use AI in client work. Deloitte is at the same time investing heavily in AI, committing US$3 billion through 2030 to modernise delivery and promoting tools like Zora AI and internal assistants such as PairD. The question: should Deloitte require human review of every AI-generated passage in client deliverables, even if it slows delivery? This is a product and policy trade-off between speed and productivity on one side and quality, trust and accountability on the other. I'll treat the review process as a product: something used by thousands of practitioners every day.

Step 1

Clarifying Questions (3 minutes)

• Context: What counts as an "AI-generated passage"?

• Why it matters: Grammar suggestions and fully drafted analysis carry very different risks.
• Hypothetical answer: Any text, table, citation or figure substantially produced by a generative AI tool.
• Impact: We may need risk tiers rather than one rule.

• Business context: How much delivery time does AI save today?

• Why it matters: Sizes the cost of slowing down.
• Hypothetical answer: Internal estimates suggest AI drafting saves 15–25% of time on research and documentation tasks.
• Impact: A heavy review process could erase much of that gain.

• User impact: Who would do the review?

• Why it matters: Review capacity is limited.
• Hypothetical answer: Engagement managers and subject experts, many of whom are already stretched; a large share of drafting happens in USI teams.
• Impact: The review must be efficient and well-designed.

• Technical: Can tools detect which passages were AI-generated?

• Why it matters: Enforcement depends on tracking.
• Hypothetical answer: Firm-approved tools can log AI use; external tools cannot be tracked reliably.
• Impact: Policy must push work into approved, logged tools.