57% of businesses now name AI errors as their top threat. Not because AI gets things wrong. Because it gets things wrong while looking completely right.
I ask business owners what they're afraid of when using AI and the answer is almost always the same:
"Having something go out from our business that looks right but is completely wrong."
Not slow outputs. Not tone that needs adjusting. That.
It's the hallucination fear. And for professional services businesses, it's the right fear to have.
The Gallagher 2026 AI Adoption and Risk Benchmarking report came out in February. I came across it recently while pulling research, and one number stopped me. They had surveyed more than 1,200 businesses globally and asked them to name their top AI threat. AI errors and hallucinations came in first, at 57%. Ahead of data privacy. Ahead of job displacement. Ahead of every other concern on the list.
The fear is correct. And most businesses aren't handling it.
It doesn't tell you when it's guessing. You get an answer that's formatted correctly, sounds authoritative, and could be completely made up. A stat that doesn't exist. A citation to a case that was never decided. A case study the model built from pattern-matching because that's what a case study is supposed to look like.
The confident liar problem
AI makes fewer mistakes than most people assume. On repetitive, high-volume tasks it often outperforms humans. That's not the issue. The issue is what happens when it is wrong. A person who doesn't know the answer hesitates, leaves a gap, says they'll check. AI fills the gap with something that looks exactly like an answer.
That's the hallucination trust gap. Not that AI makes mistakes. Every tool does. The gap is that AI makes mistakes in a way that looks like certainty. And when your clients are paying for your judgment, that matters.
A wrong answer from an AI is a problem you can fix.
A wrong answer that you passed along as your own professional assessment is a different category of problem entirely.
The same Gallagher report found that less than half of businesses have any AI risk management framework in place. No incident response plan. No protocol for when AI output is wrong. That gap, between "we use AI" and "we know how to catch what it gets wrong," is where professional reputations take damage quietly, before anyone notices.
Better prompting won't fix this. Neither will switching models. What fixes it is reading the output before it goes out. Checking the sources. Verifying anything that carries your name.
You probably already do a version of this manually. Read the AI output twice, feel vaguely uneasy, send it anyway when you're busy.
When I work with expertise-led businesses on AI implementation, the verification layer is part of what we build. It is not an afterthought. Not because AI is untrustworthy across the board, but because the professional is accountable, and their clients are counting on their judgment, not the model's.
If you're asking whether your safeguards are sufficient, they probably aren't. Reach out. I'm happy to discuss.
The technology will keep improving. Hallucination rates will come down. But "average accuracy" doesn't protect the individual document that goes to your most important client on your most important day. The accountability still lands on the professional who signed off.
That part isn't changing.
Source
Gallagher 2026 AI Adoption and Risk Benchmarking Report (February 2026, 1,200+ global businesses surveyed).