The capture gap: why your team saves hours with AI and your company doesn't

A team I work with started saving about four hours a week each once AI was properly part of their day.

Their company's numbers did not move at all.

Both of those things are true. The gap between them is the most useful thing to understand about AI at work right now.

Most companies can't see the savings

Researchers at the US National Bureau of Economic Research asked roughly 6,000 CEOs, CFOs and senior executives across the US, UK, Germany and Australia what AI had done to their productivity over three years (the paper is here). Close to 90% said nothing measurable had changed.

That number gets passed around as proof the whole thing is a bubble. I don't read it that way.

I think it points to something much more specific, and much more fixable.

The time savings are real. They leak.

You save twenty minutes on the Monday report. Those twenty minutes don't turn into twenty minutes of better work.

They turn into inbox, or a meeting that quietly expands to fill them. Nothing structural changed, so nothing structural shows up in the numbers.

Spread that across a team of forty and you get a lot of slightly less tired people, and a P&L that looks exactly the same. That's the capture gap.

Same tools, different decision

The companies that do see a return aren't using better models. They have the same subscriptions as everyone else.

What they did differently was sit down and name which specific tasks should stop needing a person. Not "we're adopting AI", and not a licence rollout with a training session attached.

A list of tasks, with names against them, that a person no longer touches. That's a workflow decision, and no tool will make it for you.

Signs you have a capture gap

You probably have one if most of these sound familiar:

• People say AI saves them time, but nobody can name a task that stopped happening.
• You measure AI by licences bought or training sessions run, not by work removed.
• Every team has a favourite tool, and no shared list of what each one is for.
• Headcount plans and delivery timelines look exactly as they did before the rollout.

None of these mean the training failed. They mean nobody turned saved minutes into a decision about the work.

How to close the gap

1. List the recurring work, team by team. Every task people do weekly, with a rough time against it. An hour per team is usually enough.
2. Sort each task into three buckets. What AI can do alone, what people do with AI, and what stays entirely human.
3. Decide where the freed time goes before you automate. If nobody decides, it goes to email.
4. Measure the work that stopped, not the hours saved. "The weekly report no longer needs anyone" is something a board can see. "People feel faster" isn't.

The third step is the one almost everyone skips. It's also the one that turns individual time savings into something the company can bank.

The bucket that matters most

The bucket that surprises people is the human-only one. There's always work that costs more to check than to just do yourself.

Handing that work to a model is how teams end up in the 90%. They automate it, spend the saving on checking it, and conclude AI doesn't work.

Knowing where that line sits is the actual skill. Everything else is software.

One thing to do this week

Pick one team and run the list with them: every recurring task, a time against each, sorted into the three buckets.

Then ask one question before anyone touches a tool. If this task disappears, what does the team do with the hours instead?