First the ceiling: what is there to save at all
According to aggregated office telemetry from 2025, an average worker receives 117 emails a day — most of them looked at for less than a minute — and 153 chat messages per working day. During core hours they are interrupted every two minutes on average. Half of all meetings pile into two bands (9–11 and 13–15), and 57% of them are ad hoc calls with no calendar entry.
There are two numbers for how much of that goes on email, and they do not measure the same thing:
| Measurement | What it measured | Result |
|---|---|---|
| 40 people, 12 working days, activity logging (2016) | all email activity | ~1.5 hours a day, 77 opens a day |
| 6,000 people, 56 companies, randomised, telemetry (2025) | reading email | 2.8 hours a week in the control group |
The hour and a half a day is the whole of email: reading, writing, searching, filing. The 2.8 hours a week is only the reading part of it. The two numbers do not contradict each other — but anyone trying to derive one from the other is doing the arithmetic wrong.
The seven task types
This is the backbone of the article. Behind every line there is a controlled measurement, not an impression.
1. Drafting and writing — the largest measured gain
444 graduate professionals (marketers, grant writers, consultants, analysts, HR staff) were given occupation-specific writing tasks: a press release, a report, an analysis plan, an email.
Average time on task fell from 27 minutes to 17, and the graders' scores went up as well — so it did not just get faster, it got better. The spread in performance roughly halved: the weaker performers were pulled up.
But there is a small-print part that is almost never quoted alongside it. 68% of the treated participants submitted the first output with no editing at all, and after pasting it they spent an average of three more minutes on the task. A large share of the saving therefore came from the fact that they did not check it. That does not invalidate the measurement — but it tells you where its limit is.
2. Email — measurable, but smaller than you think
In that same randomised 6,000-person experiment: whoever got the licence spent 12 minutes a week less reading email (−7%). Whoever actually used it, −18% — more than half an hour a week. Response time got 18 minutes faster.
The number of conversations, though, did not change. The machine makes you read your inbox faster; it does not make it smaller.
3. Closing documents — a measurable gain, with a twist
Same experiment: licence holders closed their text documents nearly half a day faster (−6%), and those who actually used the tool nearly a full day faster (−12%).
The twist: they started more documents as well (+4.6% and +11.2% respectively), and read more of their colleagues' (+6% and +14%). The document got cheaper, so more of them got written.
4. Meetings — exactly zero
This is the most important negative result of the experiment. For total time spent in meetings and for the number of meetings, the authors measured a "small and statistically insignificant" increase — which is to say, nothing.
And what says even more: the spread between companies is enormous, in opposite directions. There are organisations where it fell significantly and organisations where it rose significantly. A meeting is not a tooling question, it is an organisational habit.
5. Customer support replies — a big gain, unevenly distributed
5,172 support agents, staged rollout: issues resolved per hour rose by 15%, and the average conversation got 3.7 minutes shorter.
But the average misleads here. Among novice and lower-performing agents the improvement was above 30%. Among the most experienced, a small speed-up — and a small drop in quality. An assisted agent with two months of experience performed like an unassisted one with six. The detailed recipe and the limits belong to the support replies post, which sets out four documented counter-examples.
6. Complex judgement — here the machine does not slow you down, it makes you worse
758 consultants, a pre-registered field experiment. On eighteen realistic knowledge-work tasks, the people using AI completed 12.2% more tasks, 25.1% faster, at better quality.
The same experiment included one management task deliberately designed to sit outside the model's capability. There, the AI users were 19% less likely to produce the correct answer. They did not get slower — they got worse, and they did not notice.
The largest public-sector rollout points the same way (20,000 civil servants, 12 ministries): the machine is strong at pulling topics and claims out of a document, but it stumbles on nuanced, context-heavy material where human judgement is required. On sensitive HR cases, on complaint handling, on weighing up contradictory considerations, it brought no meaningful benefit.
7. Short, easy tasks — here there is a measured slowdown
This is the most recent and the most awkward result. Three pre-registered studies, 2,691 participants in total, 24 tasks in four categories, all of them under five minutes.
Participants expected an average saving of 55.7 seconds. The actual figure was 7.5 seconds. And on the easy tasks AI genuinely slowed people down: from 60.2 seconds to 70.2.
Anyone who opens a chat window for a three-line email is not saving anything. They are spending on prompt writing exactly the time they would have spent writing.
Perception versus measurement
The seven items above are all measurement. Self-reporting sits somewhere else entirely.
In that same 2,691-person study, participants underestimated their own usage as well: they predicted 33%, and it turned out to be 47%. So they use it more than they think, and gain less from it than they think — at the same time.
The sharpest number on this comes from a developer experiment: participants felt a 20% speed-up while the measurement showed a 19% slowdown. Thirty-nine percentage points between the two. What that means for measuring your own rollout is something the measuring the effect post goes through in detail, with marketing examples; the developer side is the subject of the what AI changed in programming post.
The largest public-sector rollout measured a saving of 26 minutes a day. That figure turns up in the news a lot — but the report itself spells out that it was not a randomised experiment, and the time saving is entirely self-reported, from a pick-a-bracket estimate. 83% of participants reported a saving; 17% reported none.
Where the saved time goes
The strongest data on this comes from a Danish study: 25,000 workers, 7,000 workplaces, linked to administrative labour-market records.
- The average saving is 2.8% of total working time — between 0.6% and 6.8% depending on occupation.
- 80–85% of it moves into another work task. Less than a tenth becomes a break or free time.
- Some users pick up new tasks: where the employer actively supports usage, roughly one in six. And 35% of those new tasks are quality-checking, ethics and compliance for the AI itself.
- On earnings: no detectable effect, with the uncertainty band ruling out even an effect larger than 2%.
A 3,200-person corporate survey shows the same thing from the other direction: 85% of respondents report a saving of one to seven hours a week, but roughly 40% of the saving goes back into rework — fixing, rewriting, checking — and only 14% report a lasting, unambiguously positive balance.
The sentence worth remembering from all this: the machine does not give you time, it rearranges the work. The effort does not disappear, it moves — from producing information to checking information.
Where it is zero or negative: in one table
| Area | Result | Sample |
|---|---|---|
| Total time in meetings | no effect, with opposite spread between companies | 6,000 people, 56 companies, randomised |
| Earnings and hours worked | zero, any effect above 2% ruled out | 25,000 people, administrative records |
| Complex management task | −19% correct answers | 758 people, pre-registered |
| Cognitively easy task | +10 seconds, i.e. a slowdown | 2,691 people, pre-registered |
| Most experienced support agent | small speed-up, small quality drop | 5,172 people |
| Sensitive HR case, policy judgement | no meaningful benefit | 20,000 people, observational |
Common misconceptions
- "I save half an hour a day with it." Maybe — but the measured data says roughly 40% of that goes back into checking and fixing, and the rest into another work task, not into free time.
- "If it worked in one experiment, it will work for me." Of the seven task types, four showed a large gain, two showed zero or negative, and one depends on whether the task sits inside or outside the model's capability. What matters is which one your day is made of.
- "A feeling is data too." Not on this topic. With an interruption every two minutes, nobody can estimate where their hour went — in the stopwatch-timed developer experiment the feeling was off by 39 percentage points.
- "It is bound to cut down on meetings." Not one randomised measurement has shown that. A meeting is an organisational habit, not a writing task.
- "The more I use it, the better." The slowdown measured on easy tasks comes from exactly that. And the measurement shows that earlier use increases use on the next task — while also strengthening the illusion of speed.
What to take away
The question is not "does AI help", but which task types your day is made of. On drafting, on email, on support replies and on closing documents you gain a lot. On meetings, nothing. On short, routine steps you lose. And with complex judgement the risk is not speed but accuracy — that is where the machine is confidently wrong.
And before you roll it out anywhere: measure it first. On this topic the feeling is a demonstrably bad instrument, and even the direction of its error is predictable — it errs upwards.
The concrete workflow for the inbox is laid out step by step in the inbox triage post; and for meetings, where you gain no time but the quality of the notes improves a lot, the meeting notes post gives you a recipe.