Workers say AI gives them about 11 hours a week back. The same workers report spending 6.4 hours a week looking after the AI.
Both numbers come from the 2026 Work AI Index, a survey of 6,000 full-time digital workers in the US, UK and Australia, fielded between December 2025 and January 2026 with researchers from Stanford, Berkeley, UCL and four other universities. It gave the second number a name that stuck: botsitting.
Resist the urge to subtract. The 11 hours is what people estimate they saved against an imagined week without AI. The 6.4 is what they report actually doing. Different bases — and only one of them is close to a diary entry.
That tension is the story. The split underneath makes it concrete: workers now spend 37% of their AI time botsitting, 36% using AI, and 27% learning and building agents. Supervision has quietly overtaken production as the main thing people do with AI.
Botsitting is the time you spend managing AI instead of doing your work: feeding it context, supervising what it produces, debugging what it gets wrong, and cleaning up afterwards. The 2026 Work AI Index puts it at 6.4 hours per week — 2.3 hours of context-feeding, 2.2 hours of supervision, 1.7 hours of debugging, and 0.2 hours of cleanup and tool-switching.
What Is Botsitting, and Where Do the Hours Go?
Botsitting is the overhead of delegation, and it breaks into four uneven buckets. Three of the four shrink with practice. One should not.
| Activity | Hours per week | Can practice reduce it? |
|---|---|---|
| Feeding context — explaining background the model does not have | 2.3 | Yes, substantially. Reusable context beats re-explaining. |
| Supervising output — reading, judging, deciding whether to trust | 2.2 | Partly. Better verification habits make it faster, not optional. |
| Debugging errors — repairing what came back wrong | 1.7 | Yes. Much of it traces back to thin context upstream. |
| Cleanup and tool-switching | 0.2 | Marginally. Mostly a tooling and integration question. |
The pattern holds across independent research. BCG's fourth annual AI at Work study, covering 11,749 workers in 14 markets, found that 47% of the workers surveyed report spending more time managing and directing AI than doing the work itself. Not a fringe experience. Nearly half.
And it has a cost beyond the clock. The Work AI Index found that for every 10% more time workers spend feeding AI context, they are 25% more likely to report feeling worn out.
Why the Saved Hours Don't Show Up in the Numbers
Because most of them are never redirected, and self-reported savings are the least reliable number in this field. This is where individual experience and company results stop agreeing.
Start with redirection. BCG found that 42% of regular frontline users report saving at least a full workday per week through AI — and that 66% of regular frontline users get limited or no guidance on what to do with that time. More than half don't redirect it into strategic work. An hour saved with no plan attached is an hour absorbed.
Then the harder finding. A February 2026 NBER working paper surveying nearly 6,000 senior executives across the US, UK, Germany and Australia found that 69% of firms actively use AI — and nine in ten reported no impact on employment or productivity over the previous three years. Their own forward expectation was modest: about 1.4% productivity growth over the next three years.
The sharpest evidence that self-reports are unreliable comes from the people who went looking. In METR's 2025 randomised trial, 16 experienced open-source developers took 19% longer on 246 real tasks when allowed to use AI — while still believing afterwards that AI had sped them up by 20%. Perception and measurement pointed in opposite directions.
What happened next matters more than the headline, and it is why we cite METR rather than a vendor survey.
METR did not stop there. In February 2026 the team announced it was redesigning the experiment, reporting later-2025 results of −18% speedup for the original developers (confidence interval −38% to +9%) and −4% for newly recruited ones (−15% to +9%). Both intervals cross zero. METR's own reading is that it is "likely that developers are more sped up from AI tools now" — while flagging two selection effects that widen the error bars. It is missing the most AI-optimistic developers, who won't join a study that asks them to work without AI. And it is missing their best tasks: 30% to 50% of participants said they chose not to submit some work because they did not want to do it without AI, which "implies we are systematically missing tasks which have high expected uplift from AI".
So the honest summary is not "AI makes you slower". It is that the effect is moving, the error bars are wide, and the one team that measured it carefully trusted self-reports less over time, not more. Extend that scepticism to your own sense of how much time AI saves you.
Where You Point AI Matters More Than How Much You Use It
Breadth of application predicts benefit almost linearly — and most people are using AI for the tasks it helps with least. Gallup's July 2026 workplace study of 22,573 employed US adults puts numbers on both halves of that sentence.
Reported positive productivity impact, by how many distinct purposes someone applies AI to:
| Number of uses | Report positive productivity impact |
|---|---|
| 1–2 purposes | 45% |
| 3–4 purposes | 66% |
| 5–6 purposes | 78% |
| 7 or more purposes | 90% |
Now the mismatch. Gallup found the highest-yield uses are coding assistance and automation (77% positive each), presentations (76%) and analytics (75%). The two most common uses are writing (51% of users) and research (49%) — which score 68% and 65%.
So the average worker has picked two of the lower-yield applications and stopped there. That is not a technology problem. It is a habit nobody has been given a reason to break.
The Skill That Shrinks the Tax
Botsitting falls when context stops being retyped and starts being reused. The largest bucket — 2.3 hours a week — goes on telling the model things it could have been given once, properly, in a form you can reuse tomorrow.
That is the practical core of context engineering: stop optimising the wording of a single request and start managing what the model can see. Standing background about your role, your customer, your constraints and your standards belongs in a reusable brief, not in the third paragraph of every prompt.
Anthropic's Economic Index suggests this is learnable rather than innate. Users with six or more months of experience are measurably more likely to have a successful exchange even after controlling for task type — and the behavioural difference is that experienced users work iteratively with the model rather than handing over the whole task and hoping.
Experience changes how people delegate, not just how often. Long-tenure users take on harder work and collaborate across several turns instead of one-shot delegation. That habit is what turns 6.4 hours of supervision into a shorter, sharper loop — the kind of thing daily practice builds and one-off training does not.
Supervision is the bucket you should not try to eliminate. The Work AI Index reports that 41% of workers say they sometimes deliver AI-generated work they couldn't explain if asked, and that 12% knowingly ship AI-generated output they believe is wrong (section 05 of the report, which breaks the behaviour into three forms). Those numbers rise, not fall, when people optimise for speed. The goal is faster verification, not less of it.
What to Practise This Week
Five habits, each aimed at a specific bucket:
- Write your context once. A standing brief on your role, audience, constraints and quality bar — reused, not retyped. Targets the 2.3-hour bucket directly.
- Verify claims, not prose. Read the output for the two or three factual assertions it rests on, and check those. Reading for tone is how unexplainable work ships.
- Ask for the reasoning before the answer on anything consequential. It is faster to audit a short chain of logic than to reverse-engineer a polished paragraph.
- Add a third and fourth use. If you only write and research with AI, you are sitting in Gallup's 45% band. Analytics, automation and preparation work move you up it.
- Keep some work AI-free. Microsoft's 2026 Work Trend Index found that 43% of the group it calls Frontier Professionals deliberately work without AI sometimes to keep their skills sharp, against 30% of other AI users.
None of these is a tooling change. They are practice — the working definition of AI fluency — and practice is what closes the AI productivity paradox at the level of an individual working day.
The 11 hours are an estimate. The 6.4 are a diary. Which one moves is a question of what you practise between now and the next model release.
Frequently Asked Questions
- What is botsitting?
- Botsitting is the time workers spend managing AI rather than doing their own work: feeding it context, supervising its output, debugging its errors and cleaning up afterwards. The 2026 Work AI Index puts it at 6.4 hours per week, from a survey of 6,000 full-time digital workers in the US, UK and Australia.
- Does AI actually save time at work?
- Workers say it does — around 11 hours a week in the 2026 Work AI Index. That figure is a self-reported estimate of time saved, not a measurement, and it cannot simply be netted against the 6.4 hours the same survey records people spending on managing AI. The two numbers have different bases.
- Why don't AI time savings show up in company results?
- Saved time is rarely redirected. BCG found 66% of regular frontline users get limited or no guidance on what to do with it. And a February 2026 NBER study of nearly 6,000 senior executives found nine in ten firms reported no impact on productivity or employment over the previous three years.
- How do you reduce time spent supervising AI?
- Attack the largest component first: context-feeding accounts for 2.3 of the 6.4 hours, and reusable context beats re-explaining. Then verify claims rather than reading for tone, and widen where you apply AI — Gallup found reported positive impact rises from 45% at one or two uses to 90% at seven or more.