58% of Leaders Made a Big Decision on Wrong AI Insight: GWI

What did GWI's AI trust survey find?

GWI surveyed 1,883 senior marketing, insights and agency professionals in the US, UK and Singapore, and 58% said they had made a significant business decision based on an AI-generated insight that later turned out to be wrong or misleading. The report, Artificial Confidence: The AI Trust Gap, came out on November 10, 2026.

These are heavy AI users. According to GWI's report page, 88% use some kind of AI tool every day, and 80% say they frequently run into AI findings that are inaccurate, misleading or untrue. The press release adds that 82% reported at least one negative consequence of bad AI output inside their business.

"AI sounds just as certain when it's wrong as when it's right," said Tom Smith, GWI's founder and CEO, in the release. A wrong answer from a chatbot comes back in the same confident tone as a right one, so nobody stops to check it.

What did the wrong AI answers cost these businesses?

Mostly time, and sometimes real money. The most common results were internal rework, delayed decisions and customer complaints. More than a quarter of the leaders reported a financial loss, and about one in five said a bad AI insight weakened a campaign's performance.

GWI warns that these costs get absorbed into daily work instead of being traced back to the bad insight. A business counts the hours AI saved and never adds up the hours spent cleaning up after it. 85% of respondents rated the risk of decisions based on inaccurate AI as significant, so they know. Most of them still don't check.

How often do leaders check what AI tells them?

Not often. Only 36% said they verify AI-generated insights most of the time (43% in the US, 29% in the UK), and 22% check only occasionally. Another 36% cross-check by asking a different AI tool, which GWI compares to asking one machine to mark another's homework.

Old information is a second leak. 63% said they had unknowingly acted on data that was out of date, and 72% agreed that AI interfaces make old information look more relevant than it is. If your AI assistant learned your prices in March, it will quote March prices in November with a straight face.

One caveat. Superpower Daily's write-up notes the survey ran in July 2026, covered people at companies with at least 500 employees or 20 million pounds in turnover, and is self-reported. A roofing company with six trucks isn't in this sample. I'd still bet the pattern is worse for small businesses, because there's usually nobody else in the building to catch a bad answer before it goes out.

What does this mean if you run ads, answer leads or close sales?

Treat AI output like a new hire's first report. It's fast and often useful, and you check it before you spend money on it. For a small business the risk shows up in three places: ad numbers, lead follow-up and what your AI agents tell customers.

1. Check ad claims against your bank account. If a dashboard or AI assistant says a campaign is profitable, pull the real spend and the real revenue yourself. Our free ROAS calculator takes about a minute and gives you a number you can defend.
2. Don't let one AI grade another. Check against a source that can't make things up: your CRM, your payment deposits, a call recording.
3. Hunt for stale facts. Prices, hours, service areas and current offers are what AI chat and voice agents repeat to customers. Update the documents they learn from every time something changes.
4. Read what your agents said. Pull a handful of AI call and text transcripts every week and compare them to what actually happened on the calendar.

We run AI agents that call and text leads for our own business, and we check their work. In our own calendar data, over 45 days, calls our AI agents booked went 10 showed and 14 no-shows. No-shows were mostly booked 2 to 5 days out, and one happened because a booking had the wrong time zone. The agent's summary looked fine, and the calendar showed the problem. In KenjiAI every call, text and booking lands in the CRM with the conversation attached, so you can read what a Voice AI agent said before you trust its numbers.

Is AI still worth using for marketing decisions?

Yes, and GWI's respondents aren't backing away from it. They want AI checked against real people: 80% said understanding customer motivations still requires human research, and 75% said AI will never fully replace talking to customers directly.

I've been running ads since 2013, and every new tool comes with the same trap: the report looks better than the bank account. Pick one decision you made this month based on an AI answer, like an ad budget change, a new offer or a lead list, and check it this week against one real number.

FAQ

What percentage of business leaders made a decision based on wrong AI insight?

In GWI's November 2026 report Artificial Confidence: The AI Trust Gap, 58% of 1,883 senior marketing, insights and agency professionals said they had made a significant business decision on an AI-generated insight that later proved wrong or misleading. The survey covered the US, UK and Singapore.

How often do businesses verify AI-generated insights?

Only 36% of the leaders GWI surveyed said they verify AI insights most of the time, and 22% check only occasionally. Another 36% check one AI tool's answer with a different AI tool instead of a real data source.

What are the costs of inaccurate AI output for a business?

GWI found 82% of respondents had at least one negative consequence. The most common were internal rework (38%), delayed decisions (36%) and customer complaints (34%), followed by financial loss (27%), reputational damage (26%) and weaker campaign performance (21%).

How can a small business check AI marketing numbers?

Compare what the AI says against sources it cannot invent, such as actual ad spend, payment deposits, CRM records and call recordings. Recalculate return on ad spend yourself, and read a sample of your AI agents' call and text transcripts every week.

Originally published at kenjiai.com. Published by the Media Traffics | KenjiAI team.