sk most teams how they're getting on with AI, and the answer is upbeat. People feel fluent. They use it daily, for drafts, research, first passes at almost anything. New research published this week suggests that confidence is running well ahead of the results it's actually producing.

WalkMe's 2026 AI at Work Pulse Survey, the third in the series, surveyed 2,037 working Americans who use AI on the job. 90% said they feel confident using it. Only 24.6% said it worked on the first try. Just over half, 50.2%, said they had spent more time trying to get AI to do a task than the task would have taken them manually.

A second, independent piece of research adds weight to the same picture. SAP's Value of AI Report found 79% of businesses have experienced rework, delays or backlogs caused by low-quality AI output.

Put those together and a specific, useful problem appears. It isn't that people are avoiding AI, or using it badly out of ignorance. Most people can operate the tools comfortably. The gap is between feeling capable and checking whether the output was actually right, useful, or worth the time it took.

That distinction matters more for a smaller business than a larger one. A large organisation with a dedicated data or AI team is more likely to have someone whose job is to notice when AI output needs rework, and to measure it. A 20 to 100-person business rarely has that role. AI use spreads informally, person by person, and nobody is specifically responsible for checking whether it's landing. Confidence becomes the only signal anyone is tracking, because it's the only one anyone can see.

This is precisely the gap the Value dimension of a structured AI diagnosis exists to close. Not "is the team using AI," which nearly every business can now answer yes to. The harder, more useful question is: how would you know if it's working? Which tasks are genuinely faster and better because of AI, and which ones are quietly costing more time than they save, dressed up as progress because everyone feels comfortable doing them?

The honest answer for most SMEs right now is that nobody is measuring this at all. That isn't a criticism. Measuring AI output quality wasn't a skill anyone needed two years ago, and most training available still focuses on how to use the tools rather than how to judge what they produce. Confidence training is everywhere. Judgement training is rare.

The fix isn't more tool training, and it isn't banning AI until someone works out the risk. It's building a habit, small and specific, of checking outputs against outcomes on a regular basis: did this actually save time, was the quality good enough to use as-is, and if not, who caught it before it went further. That's a governance and value question as much as a technical one, and it's exactly the kind of question a proper AI Accelerator Diagnostic is built to surface across a business, not just in one team or one tool.

None of this is a reason to be cautious about AI. It's a reason to be specific about what "working" actually means, rather than assuming that comfort with a tool is the same thing as getting good results from it. The businesses that do well over the next year are unlikely to be the ones using AI the most. They're more likely to be the ones who took the time to check.

Go wisely.

WalkMe "The AI Confidence Trap: Feeling Fluent Isn't the Same as Being Effective," via GlobeNewswire, 25 August 2026; SAP Value of AI Report