You bought into the AI thing. Your team uses it every day now. On paper it's working, because everyone says they're saving hours. But the work isn't getting out the door any faster, and you can't figure out where the time went.
There's finally a name for where it went. It's called botsitting.
A report called the Work AI Index came out June 10, 2026. Glean's Work AI Institute ran it, and it's not the usual vendor survey. They polled 6,000 full-time workers across the US, UK, and Australia, and academics from Stanford, Berkeley, and a handful of other universities co-authored it. So the numbers carry weight.
Here's the headline. 87% of workers use AI, and it saves them about 11 hours a week. That's a quarter of the workweek. But only 13% of organizations say they're actually performing significantly better.
That gap is the whole story.
Where the 11 hours goes
Workers spend an average of 6.4 hours a week botsitting. Feeding the bot context it's missing. Checking its answers. Debugging confident-but-wrong output. Re-running the same prompt because the first two tries were garbage.
That's 37% of their AI time. Slightly more than the 36% they spend actually producing work with it. For every productive hour, they spend roughly another hour making the output usable.
Do the math and the real number is about 4.6 hours saved, not 11. The 11 is the slide you show the board. The 4.6 is what your team actually lives.
Now the part that hits a small shop the hardest. One IT leader in the writeup put it plainly: the productivity gain was never real savings, it was a transfer of labor from the person who made the output to the person who inherited it. Someone ships work they didn't verify, and it lands downstream on whoever has to fix it. On a big team that gets absorbed quietly as rework nobody budgeted for. On your team, that downstream person is probably you. There's no spare bench to soak it up.
Why the hidden work shows up
Here's the part most people get wrong. Botsitting isn't proof the AI is bad. It's proof nobody engineered the layer around it.
Four things are usually missing, and none of them is the model.
- No shared context. The AI can't see what your business knows, so a human re-supplies it by hand on every prompt. 53% of workers in the study said the information they need isn't even reachable by their AI. Giving a tool a login is not the same as giving it context.
- Disconnected tools. 77% of people bounce between multiple AI tools every week, and a third use four or more. Work gets scattered, so people copy-paste between apps and re-run the same prompt somewhere else hoping for a better answer.
- No defined ownership. Nobody decided what the AI owns and what the human owns. So every worker improvises that boundary on their own, every time.
- No verification built in. Checking gets left to the individual at the end. So it either eats hours, which is botsitting, or it gets skipped and bad work ships.
And here's why debugging AI specifically wears people down. When the output is wrong, you have to reason backward from a bad answer without knowing which missing piece caused it. That's detective work, not your actual job. It burns focus faster than the work itself would have.
Picture the difference:
Prompt with no context:
"Write the onboarding email for the new client."
-> the bot guesses the tone, invents details, misses your deadline policy
-> you spend 20 minutes fixing what it guessed
Same prompt, with the context layer doing its job:
"Write the onboarding email for the new client."
-> the system already knows your voice, the client's plan,
the deadline policy, and the last three emails you sent
-> you spend 2 minutes approving it
Same prompt. The only thing that changed is what the system already knew before anyone asked.
One more finding, because it flips the usual advice. The people getting the most out of AI don't botsit less. They botsit more. 40% of their AI time versus 33% for everyone else. The difference is they're picky about what they hand off, they invest in feeding rich context up front, and they build verification in instead of bolting it on after. They're 4.4 times more likely to be proud of what comes out. The goal was never zero oversight. It's oversight that's designed in, not firefighting a system that was never set up.
This isn't one vendor's number either. A separate report this year clocked the same supervision burden at nearly 8 hours a week. Different study, same shape.
What "built right" actually means
This is the whole reason LTFI exists. It's our system for doing the work of a much bigger team with two people, and it only holds together because the layer around the AI is engineered, not improvised. Shared context the tools can actually reach. Connected systems instead of fifteen tabs. Clear lines on what the AI handles and what a human signs off on. Verification that runs before anything ships, not after a client catches it.
That's not a new platform you go buy. The study is clear on this too: the fixes don't need a new vendor or a new headcount plan. They need someone to engineer the context and integration layer once, so your team stops paying the tax every day.
We run our own shop this way. Our entire daily content operation runs end to end on it, with a quality gate that throws out bad output before a human ever sees it. We build the same thing for clients.
If your team is saving 11 hours and somehow handing 6 back, that's not an AI problem. That's a layer that was never built. Subscribe free at kief.studio and grab the companion resource that breaks down the four gaps and what to check for in your own setup. And if you want a set of eyes on it, the first conversation is free, no commitment.