A landmark study of 6,000 full-time knowledge workers, co-authored with researchers at Stanford, UC Berkeley, and Harvard, found that 87% of workers use AI and 75% feel more productive. But only 13% say AI has significantly improved their company's performance. If you lead a team, closing that gap is now the core challenge of managing AI at work.

The stat almost everyone reads wrong

It's easy to read that 13% as a failure. It's actually the needle starting to move.

Rewind a year, and most companies couldn't point to any measurable return. Widely cited MIT research found 95% of enterprise AI pilots showed no measurable financial impact, and a majority of CEOs reported zero revenue or cost benefit. Against that backdrop, 13% reporting significant company-level improvement is a massive jump from zero.

The same thing is happening to individuals. Workers now report saving around 11 hours a week with AI. A year or so ago, Federal Reserve research put the average closer to 2 hours a week, with only the heaviest daily users hitting four-plus.

Both numbers are climbing, and not because the models got smarter. What changed is how people use them: how much context they supply, and whether the work was set up properly in the first place. The upside is increasingly a function of operating skill, not raw model capability.

Everyone is a manager now

A year ago, most of your team did the work themselves. They wrote the doc, built the deck, pulled the numbers, answered the email. Today they do something different. They hand the work to something else, wait, and decide whether what comes back is good enough to ship.

Technically that's "using a tool." But it's also "managing." And almost no one on your team was trained to do it. The tools turned everyone into a power user and, at the same time, into a first-time manager of a fast, tireless, occasionally confidently wrong direct report. Some people are managing it well. Most are winging it.

The self-inflicted tax: botsitting

Those same workers spend about 6.4 hours a week "botsitting": feeding the AI context it didn't have, checking its output, correcting mistakes, re-running prompts, and stitching together disconnected tools.

The instinct is to blame the tool. The data says otherwise. 53% of workers say the information they actually need isn't even accessible to their AI.

That isn't the model being dumb. It's people asking for good work without ever giving the AI what it needs to produce it. Set your systems up right and you don't botsit nearly as much. The botsitting is the bill for skipping the setup. Ten years ago we called it technical debt: the extra work you take on when you cut a corner early.

Same model. Different manager.

Split results by whether the AI had good access to company context, and the difference is stark:

Same models. The only variable is whether someone bothered to give the AI the context it needed. With the right context, the quality failures roughly halve and burnout drops by almost two-thirds. The fixing-afterward problem mostly disappears when you move the effort to the front. Anyone who has managed a human team already knows this in their bones.

This is really a management story

When you delegate to a person, you don't get the time back for free. A good hire saves you X hours by doing the work, but you spend Y hours briefing them, answering questions, and refining the result before it ships. Your real gain is X minus Y.

Every experienced manager knows the secret to shrinking Y: front-load it. Give a clear brief, show an example of what good looks like, set the bar before the work starts. Under-brief someone and you'll spend the week rewriting their output.

AI runs on the exact same equation. The 11 hours saved is your X. The 6.4 hours of botsitting is your Y. And just like with people, the size of Y is set almost entirely by the quality of the brief and the context you provide.

A huge share of the workforce got promoted to manager overnight, with direct reports they never asked for and no training in how to delegate. This isn't a fringe idea. Forbes has run on "when everyone becomes a manager." Ethan Mollick calls management "the AI superpower." Deloitte's enterprise survey now ranks insufficient worker skills as the single biggest barrier to AI, ahead of cost, governance, and the technology itself.

Tellingly, research found that people who fully handed tasks to AI finished fastest but learned the least, while those who directed and questioned it did far better on later work they had to do alone.

What this actually takes

The opportunity in front of every leadership team isn't "get more AI." It's "teach your people to manage." Four moves do most of the work:

1. Supply context once, not every session. Build the brain of context files your AI works inside. Add durable briefs, examples, and "what good looks like" references, so people stop retyping the same background every day. That's where the 6.4 hours goes to die.

2. Teach the brief. Setting clear intent and a quality bar before the work starts is a learnable skill, and it's the biggest single lever on net savings.

3. Make reviewing part of the job. Shipping work nobody checked is a missing checkpoint, not a bad employee. Build verification into the workflow as a normal, expected step.

4. Reward the best managers, not the heaviest users. Usage is a vanity metric. The people who get clean, trustworthy output with the least rework are the ones to study and promote.

This is how a 13% gain becomes a 30% gain. The winners in the data don't out-tool everyone else. They out-manage them. That gap, teaching people to manage AI rather than just prompt it, is exactly what we exist to close.