When a chief executive says AI is saving her twelve hours a week and the people who report to her say it’s saving them nothing, they can’t both be describing the same business. Yet they are, and the distance between those two accounts tells you more about how the business is run than about the software.

The numbers behind that sentence come from a Section survey of 5,000 white-collar workers, reported by the Wall Street Journal in January. Forty per cent of staff outside management said AI saved them no time at all in a typical week, and only two per cent said it saved them more than twelve hours. Among executives the picture inverted: two per cent said it saved nothing, while nineteen per cent claimed more than twelve hours. Same technology, same twelve months — and a gap that no training budget explains.

Bar chart comparing Section survey responses: 40% of non-management staff report zero hours saved by AI each week against 2% of executives, while 2% of staff and 19% of executives say it saves more than 12 hours a week

The easy reading is that one side is deluded, and I don’t think either is. Executives are telling the truth about their own week, and so are their staff. The difference is what each of them actually does with the software, and that makes the gap a management problem before it’s a technology one.

Two different products

An executive meets AI at the summary end. The dashboard that used to take an analyst a day now refreshes itself, and the board pack more or less drafts itself from the six reports underneath it, so for that job the tool is close to miraculous and the hours saved are real. Someone on the front line meets it at the other end, where the output has to be right before anyone can use it, which means the draft still gets checked line by line and the new “AI-enabled” process often turns out to have added a step. A UX designer told the Journal he’d lost count of the times a language model gave him a confident and completely wrong answer to an accessibility problem. His boss never sees that afternoon; he sees the summary of it.

Why nobody says so

Different exposure would be easy to fix if the two sides compared notes, and most of us assume they do. The Checkr report that Fast Company covered in March suggests otherwise. It found that 45 per cent of managers believe their people are using AI regularly, against 18 per cent of employees who say they are, and that 58 per cent of managers see AI use becoming an unspoken performance requirement, which only 29 per cent of staff have noticed; a third of employees couldn’t say who in the business was even responsible for the AI push. When a tool is an unspoken performance requirement, nobody tells the boss it doesn’t work. They nod, they tick the adoption box, and they do the job the old way after hours. The feedback loop that would correct the executive’s view has been switched off by the executive’s own enthusiasm, and none of us switches it off on purpose. We back the tool loudly because we believe in it, and the louder we are, the harder it becomes for anyone to bring us the bad news.

Amy Gallo made a related point in Harvard Business Review the same month: when we hand our interactions to AI we’re “outsourcing the very moments that create connection”. At the summary end that’s exactly what happens. The manager reads the generated digest of the team’s week instead of asking the team about it, and the one conversation that would have surfaced the wasted afternoon is the conversation the tool has just replaced.

That awkward conversation is the whole fix.

Measuring whether it’s helping

The routine I’d use is short enough to run every fortnight. Before the tool goes in, note how long the work takes now and how much of it comes back for correction, because without a baseline every later claim is a guess. Then ask the people doing the work two questions each week, not one: what took less time because of the tool, and what took more. The second question is the one that matters, since nobody volunteers that answer to a boss who’s excited about it, so it needs to be asked plainly and answered without consequence. Watch the checking and rework that happens downstream of the AI output, because that’s where the front line’s hours go and where the dashboard never looks — a licence in use isn’t a job done. Then decide, in the open, whether the tool stays, and let it go if the answer is no; a team that has seen one ineffective tool retired will be honest about the next one. Gallup found that employees whose managers openly support AI are nearly nine times as likely to believe it helps them, so how a manager behaves around the tool shapes what staff will say about it, and that cuts both ways.

The organisations where this gap stays small share those habits: someone owns the rollout by name, the person leading it sits close enough to the work to see the rework, and an honest “this isn’t working yet” costs nobody their reputation. The tools they bought matter far less than that, which is the same reason bolting AI onto an old structure so rarely works, and why I argued in January that AI exposes bad management long before it improves anything. Bad news has always travelled slowly in the average business, and AI simply makes the delay measurable.

If the numbers in your business look like Section’s, start with how truth travels upward before you touch the tool, and the technology will begin earning the hours the executives already believe it’s saving.