Why Businesses Need Human Review Before AI Content Goes Live

It’s becoming a familiar kind of story: a report goes out under a trusted name, and within weeks it turns out large parts of it were never true. That’s exactly what happened to KPMG in October 2025, when it pulled its own report, Redefining Excellence in the Age of Agentic AI, after several of the organizations it described pushed back on what had been written about them.

UBS, the UK’s National Health Service, Swiss Federal Railways, and Transport for London all told the Financial Times that the report misrepresented how they actually use AI. When detection firm GPTZero checked the sourcing afterwards, it found that only five of the report’s 45 citations pointed to real, intact sources.

Deloitte Australia had almost the identical experience the same month. The firm had delivered a report to Australia’s Department of Employment and Workplace Relations when a University of Sydney researcher checked the footnotes and found fabricated academic references and a misquoted line from a federal court judgment. 

Deloitte later confirmed it had used Azure OpenAI to help draft the document, and it agreed to partially refund the AU$440,000 contract once the errors went public.

Neither firm is a small operation experimenting carelessly. Both are among the largest professional services companies in the world, and both were caught by someone outside the building, not by anyone inside it whose job it was to check first.

Why Grammar Checking Isn’t Real Review

A mistaken assumption. A misplaced trust in fluent-sounding prose. An error that made it all the way to publication before anyone thought to check.

That’s the pattern behind almost every AI content failure that’s gone public, and it has nothing to do with grammar. A spell-checker finds typos; it has nothing to say about a citation that reads perfectly well and simply isn’t real, because AI models state fabricated information in exactly the same confident tone as accurate information. 

Readers of the Chicago Sun-Times and Philadelphia Inquirer found that out in 2025, when a syndicated summer reading list ran with 10 of its 15 recommended titles invented outright.

The sentences were clean. The books didn’t exist.

A similar pattern showed up in an Australian courtroom a few months later. In August 2025, King’s Counsel Rishi Nathwani apologised to the Supreme Court of Victoria after AI-generated submissions in a murder trial cited nonexistent case judgments and misquoted a speech to the state legislature. The errors weren’t caught until the judge’s own associates went looking for the cases and couldn’t find them, pushing the resolution of the case back by 24 hours.

Nothing about the grammar tipped anyone off in any of these cases. The content simply wasn’t true, and it reached readers, a government department, and a court before anyone checked.

The Four Things Human Review Must Actually Catch

A review process worth having goes well past language polish. There are four distinct failure points worth checking every time AI-assisted content is about to go out the door.

Factual accuracy: Every statistic, quote, and claim needs to be traced back to its original source, not just checked for internal consistency. AI models can produce citations and attributions that are grammatically indistinguishable from real ones, which is exactly what tripped up both KPMG and Deloitte.

Brand voice alignment: The more content a business produces this way, the more its AI drafts tend to drift toward generic, overconfident phrasing. Left unchecked, that “semantic drift” slowly wears down a brand’s distinct voice until it reads like everyone else’s AI-assisted content.

Customer and legal risk: Some lines read as guarantees, or create a promise nobody intended to make, or need a disclaimer the AI had no way of knowing to add. Someone needs to be reading specifically for language that could mislead a customer or expose the business legally.

Missing context: AI tends to flatten out the exceptions and caveats that someone with real, hands-on experience of a topic would naturally include. That missing nuance is often the exact thing that separates content that’s genuinely useful from content that just sounds authoritative.

A Simple Review Process Teams Can Actually Use

None of this calls for a heavy compliance process. Run five checks, consistently, and most of the failures above get caught before they become headlines.

Source-check pass: Trace every factual claim back to its primary source before anything else happens.

Voice pass: Read it aloud and ask whether it sounds like something the brand would actually say, not just something that’s technically correct.

Risk pass: Flag anything that reads as a guarantee, a legal claim, or a sensitive topic that needs extra care.

Context pass: Ask what someone with direct experience of the topic would add that’s currently missing.

Named sign-off: One accountable person approves it. “AI-assisted” isn’t a substitute for someone actually putting their name to the work.

Deloitte’s report is a good example of why that last step matters. The firm stood by the report’s substance even after the fabricated references came out, but the damage to its credibility had already happened in public, in the press, rather than getting caught in-house first.

The Regulatory Backdrop Making This Non-Optional

Human review is moving from best practice to expectation. Article 14 of the EU AI Act requires high-risk AI systems to be built so human operators can genuinely monitor, understand, and step in on their outputs, not just rubber-stamp them.

Enforcement of those obligations begins in December 2027, but the underlying principle, that a person needs to be able to catch and correct an AI system’s mistakes, is already shaping how regulators and courts think about who’s accountable when something goes wrong.

Courts, for their part, are already treating “the AI made the error” as no defence at all. In the Victorian murder trial, Justice James Elliott was direct about this: counsel bears full responsibility for anything filed with the court, and using AI output isn’t acceptable unless it has been independently and thoroughly verified beforehand.

The accountability landed on the people who filed the work, not on the tool that produced it, and that’s very likely where the wider conversation about AI accountability is headed too.

Review Is What Makes AI Adoption Sustainable

The businesses avoiding these headlines aren’t the ones avoiding AI. They’re the ones that never let it publish unsupervised. 

KPMG, Deloitte, and two of America’s most established newspapers all had access to the same generative tools everyone else does, yet what separated their outcomes from the businesses that don’t end up in the news was whether a person actually checked the output before it went live.

Human review isn’t friction slowing AI down. It’s the difference between AI working as a genuinely fast first draft and AI becoming a liability with your name, and your client’s trust, attached to it.