Almost every AI productivity guide is a list of apps. This one starts with the arithmetic, because the arithmetic is what decides whether any of it helps you.
The largest real-world study of AI chatbots at work, covering 25,000 workers across 7,000 Danish workplaces, found users saved an average of 2.8% of their work hours. Not 30%. Not half a day. Under three per cent, and 80% of those users poured the saved time straight back into other job tasks. That number is the honest starting point for using AI for time management. AI does five things genuinely well: triaging an inbox, estimating how long work will really take, drafting the message that says no, summarising so you can skip things, and defending your focus blocks. What it cannot do is reduce the number of promises you have made. If your calendar is over-committed by a third, a tool that returns two hours a week is not a fix.
Most time management advice fails because it treats three separate problems as one problem. I’ve watched this play out in workshop rooms for years: everyone in the room says “I don’t have enough time,” and by the end of the exercise it turns out they’re describing completely different situations.
Here are the three, and they need different responses.
Fragmentation. Your hours exist but they arrive in useless pieces. Microsoft’s analysis of aggregated Microsoft 365 signals found the average worker is interrupted every two minutes during core hours, which works out at 275 pings a day across meetings, emails and chats, on top of 117 emails and 153 Teams messages. Half of all meetings land between 9 and 11 in the morning and 1 and 3 in the afternoon, which is exactly when most people think best.[1]
Coordination. The work about the work. Asana’s Anatomy of Work Global Index, based on 9,615 knowledge workers across six countries in November 2022, found 58% of the average day goes on coordinating work rather than doing it, across 8.8 apps.[2]
Volume. You have said yes to more work than fits in the hours you have. No arrangement of the hours makes the total smaller.
My honest take, after running this diagnostic with hundreds of learners: roughly two thirds of the people who come to an AI productivity session have a volume problem and are hoping it’s a tooling problem. That’s not a criticism. A tooling problem is much nicer to have, because you can solve it by Thursday.
A woman on a corporate training day last year, an operations manager, opened her inbox in front of the group to show me how bad it was. Four thousand unread. She said she’d stopped reading it entirely and now relied on people chasing her by phone. That inbox was not a time management failure, it was a triage failure, and it took twenty minutes to fix with a prompt.
Paste in a list of subject lines and senders and ask for three buckets: needs a decision from me, needs an answer but not a decision, and needs nothing. Then ask a second question that people forget: which of these gets worse if I leave it a week? That’s the one that reorders your day. We’ve written a fuller version of this in our guide to building an AI inbox triage workflow, and the sorting logic transfers to any queue you own.
Ask for a range rather than a number: “Give me a best case, likely case and bad case in hours for each of these six tasks, and say what makes the bad case happen.” Models aren’t psychic, but they’re unsentimental, and they don’t share your optimism about Thursday afternoon. The value is having something written down to argue with.
This is the most underrated use and the one people are most embarrassed to admit they need. Give it the request, your actual constraint, and the relationship: “Decline this without damaging it, offer one alternative, three sentences, no apology stacking.” Most people don’t overcommit because they can’t say no. They overcommit because writing the no takes fifteen minutes of dread and saying yes takes four seconds.
The point of a summary isn’t to read faster, it’s to decide whether to read at all. Ask for the summary plus one line: “What would I miss if I only read the summary?” If the answer is nothing, you’ve bought back the meeting, the thread, or the fifty-page deck.
Calendar assistants can reschedule flexible work around fixed commitments automatically, and that genuinely helps at the margins. Reclaim’s own product pages describe scheduling and automatically defending habits and focus time, moving them when a meeting lands on top.[14] It works. It also cannot protect a block from you, the harder problem.
Notice what’s common to all five: AI is operating on information and drafts, never on the decision itself. The moment the tool starts deciding what matters, you have handed over the only part of time management that was ever really yours.
The measured time savings are much smaller than the marketing, and I’d rather you heard that from me than found it out in month three.
The best real-world evidence comes from Denmark. Anders Humlum and Emilie Vestergaard surveyed around 25,000 workers across 7,000 workplaces in eleven occupations exposed to AI chatbots, in two rounds fielded in late 2023 and 2024. Users reported saving an average of 25 minutes on a day they used the tools, which across their full working time works out at 2.8% of work hours.[5] Marketing professionals whose employers actively encouraged use did best, at 6.8%. Teachers at unsupportive schools reported 0.6%.[5]
The US picture is similar. Economists at the Federal Reserve Bank of St. Louis, using a nationally representative survey in November 2024, found users saved 5.4% of their work hours, about 2.2 hours in a 40-hour week.[6] Pooling their 2025 surveys and counting everyone including non-users, that drops to 1.6% of all work hours.[7]
Self-reported savings as a share of work hours, plus one randomised trial that measured actual completion times instead of asking. Sources: Humlum and Vestergaard (2025), Federal Reserve Bank of St. Louis (2025), METR (2025).[5,6,7,8]
That last bar deserves its own paragraph. The research nonprofit METR ran a randomised controlled trial with 16 experienced open-source developers across 246 real tasks in their own repositories. Developers expected AI to speed them up by 24%. It slowed them down by 19%. And afterwards, having just lived through it, they still estimated AI had made them 20% faster.[8]
That gap between what happened and what people believed happened is, for my money, the single most useful finding in this entire field.
I want to be fair to METR too, because they’ve been unusually honest since. When they tried to repeat the study in late 2025 with 57 developers and over 800 tasks, they found their own design had broken: developers were refusing to participate, or quietly withholding the tasks they most wanted AI for, which biased the result downwards. They published that rather than the headline.[9] Their current read is that late-2025 tools probably did speed developers up, but their data is too compromised to say by how much. Both things are true: the perception gap was real, and the slowdown finding is a snapshot of one setting in early 2025, not a law of nature.
I used to think the failure mode was people trusting AI output too much. I now think it’s subtler and more annoying than that: the failure mode is work that looks finished and isn’t, which then costs someone else their afternoon.
Researchers at BetterUp Labs and Stanford’s Social Media Lab named this “workslop”: AI-generated content that appears polished but lacks the substance to move the task forward. In their survey of US desk workers, 41% said they had received it, and each instance cost nearly two hours of rework.[10] Read that as a time management finding rather than a quality one. Your two minutes saved became someone else’s two hours.
Then there’s the part almost nobody mentions when they sell you on time savings. In the Danish study, AI chatbots created new workloads for 17% of users: some doing more of the same tasks, some taking on genuinely new ones that only exist because the tool does.[5] And of the time users did save, 80% went straight back into other job tasks. Fewer than one in ten took a break with it.[5]
So the machine gives you twenty-five minutes and the job takes them back. That’s simply what happens to slack in a system that has none, but it does mean “AI will save you time” is only half a sentence.
And then there’s the version of failure I see most often in training rooms, which has no research behind it at all, only pattern recognition from a few hundred learners. Somebody’s actual problem is that they’ve promised four people the same Tuesday. Instead of having one uncomfortable conversation, they spend a fortnight comparing calendar tools and migrating tasks between apps. It feels productive. It is procrastination with a project plan attached.
Try this before you read on: add up the hours of everything you have currently promised somebody, then subtract your actual uncommitted hours this week. Most people who do this honestly are over by somewhere between 20% and 50%.
Now put the evidence next to it. The best case from the research is a couple of hours a week, and that’s for people who use AI heavily.[6] If you’re over by a third of a 40-hour week, that’s thirteen hours. Two does not close thirteen. No prompt closes thirteen. The only thing that closes thirteen is dropping something, and that’s a conversation, not a tool.
| What it feels like | What it usually is | Can AI help? |
|---|---|---|
| Busy all day, nothing finished | Fragmentation: interruptions every couple of minutes, meetings sitting on your best hours | Partly. It defends blocks and summarises what you missed, but it can’t stop your colleagues |
| Everything overruns your estimate | Optimism: you plan for the best case and hit the likely one | Yes. Ranged estimates are the single highest-yield use |
| You can never find where things stand | Coordination: 58% of the day spent on work about work | Yes, mostly. Triage, summarising and status-chasing are its strongest suits |
| Behind no matter what you do | Volume: you have promised more than the hours hold | No. It can draft the message that gets you out. It can’t make the decision |
A diagnostic based on the interruption and coordination data cited in this article, and on which problems the measured time savings could plausibly touch.[1,2,5,6]
There’s a piece of older research I keep returning to here, because it complicates the standard story. Gloria Mark’s team at UC Irvine ran a lab experiment with 48 participants doing an email task, some of them interrupted every two minutes. The interrupted group finished faster, not slower. They also reported significantly higher stress, frustration, time pressure and effort, and they wrote shorter messages.[4] People compensate for interruption by working faster and giving less, and they pay for it in how the day feels.
Which reframes the goal: you’re optimising for a day that doesn’t cost you everything to get through, rather than for hours reclaimed. Only one of those shows up on a dashboard.
One more note on a number you’ve probably seen. The claim that it takes 23 minutes and 15 seconds to refocus after an interruption is usually attributed to Mark’s team, and I went looking for it in the primary papers. It isn’t in the 2005 field study of 24 information workers, which reports that people spend about 11 minutes in a working sphere before switching and that 57% of those spheres get interrupted.[3] It isn’t in the 2008 experiment either. So I’m not citing it, and neither should the next person who quotes it at you in a meeting.
Ninety minutes on a Friday, then fifteen minutes each following Friday. No new software required, and deliberately smaller than anything a vendor will propose.
The workflow described in this section. Steps 2, 3 and 6 use AI. Step 4, the one that decides the outcome, does not.
Step three is where people flinch, so let me be specific about it. Don’t count your working week as 40 hours. Count the hours that aren’t already meetings, then take 25% off for the interruption load, because the interruptions are not optional and pretending otherwise is how the plan dies by Wednesday.[1]
For step six, the prompt that works better than any tracking app I’ve tested: paste last week’s list and this week’s, and ask “Which items moved and which have been on this list for more than two weeks? For each stale item, ask me one question about why.” The questions it asks are frequently uncomfortable and usually correct.
If you want the wider version of this, with the capture and review layers around it, our guide to building a personal AI workflow system puts the same logic into a full week.
One rule while you run it: no client names, salary figures, performance notes or anything commercially sensitive into a general consumer chatbot. Use initials or role labels. This is boring advice and it is the advice that stops a time management project becoming an incident report.
I’m including this section reluctantly, because tool lists are how most AI productivity advice avoids saying anything. So here’s what’s actually true about the features people ask me about, checked against the vendors’ own documentation in August 2026.
ChatGPT scheduled tasks do work for recurring reminders, daily briefings and monitoring for a change. The limits are real and worth knowing: tasks can’t run more than once an hour, and active task limits run from 3 on Go and 5 on Plus up to 15 on Pro, Business and Enterprise. Delete the chat a task lives in and the task pauses.[12]
Gemini in Gmail has a “Help me schedule” function that reads your availability and your guests’, proposes four slots by default, and handles up to 20 guests on a single schedule. It needs an eligible Workspace or Google AI plan.[13] For back-and-forth scheduling, this is one of the few AI features where the time saved is obvious and immediate.
Reclaim starts free for one user, with Starter at $10 per seat per month on annual billing (or $12 month to month) and Business at $15 annually, $18 monthly.[14] Worth checking the current page, because these plans have been restructured more than once.
Microsoft 365 Copilot is the one where people most often quote me a price that turns out not to be the price they would pay, usually because there are two current list prices and they have heard only one. Microsoft’s enterprise pricing page lists Microsoft 365 Copilot at $30 per user per month paid yearly.[17] Its business page lists a separate Copilot Business add-on from $21 per user per month paid yearly, promotionally discounted to $18.[15] Which one applies to you depends on how you buy, and either way it requires a qualifying Microsoft 365 licence underneath.
Todoist Assist breaks a vague task into sub-tasks, rewrites tasks to be more actionable, and builds filters from plain-language descriptions. Task Assist is a browser extension you have to enable, on Pro and Business plans rather than free ones.[16]
Every one of these is real and some are genuinely good. None of them appears anywhere in the evidence as the thing that moved the number, and I’d rather you spent your ninety minutes on the six steps above than on evaluating any of them.
Four failure patterns, in the order I see them.
Automating a commitment you should have cancelled. The most efficient possible version of a report nobody reads is still a waste. Before you build the workflow, ask who reads the output and what happens if it stops. Sometimes the honest answer ends the project.
Believing your own time savings. The METR developers were wrong by 39 percentage points about their own productivity, in a direction that flattered the tool.[8] You will be wrong too. Which is why step six of the reset is measuring what actually took what, not what you remember taking what.
Adding a tool to a fragmented day. Another app is another notification source. Asana found knowledge workers already juggling 8.8 apps.[2] If a new AI tool doesn’t retire something, you have added surface area, not capacity.
Outsourcing the judgment along with the typing. Microsoft’s 2026 Work Trend Index, which surveyed 20,000 AI-using knowledge workers across 10 markets between February and April 2026, found that the most advanced users behave differently in one specific way: 53% of them say they deliberately pause before starting work to decide what should be done by AI and what by a human, against 33% of everyone else. And 43% say they intentionally do some work without AI to keep their own skills sharp, against 30%.[11] The best users are the ones who choose when not to use it.
The one thing I’d have you take from all of this: use AI to make your commitments visible and your estimates honest, and then do the hard part yourself. The people I’ve watched genuinely get their time back weren’t the ones with the best stack. They were the ones who finally wrote down everything they’d promised, looked at the total, and told somebody no.
Less than the marketing suggests. A study of around 25,000 workers across 7,000 Danish workplaces found chatbot users saved an average of 2.8% of their work hours, roughly 25 minutes on a day they used the tools. A nationally representative US survey by the St. Louis Fed put it at 5.4% of hours for people who used AI in the previous week, about 2.2 hours in a 40-hour week. Heavy, encouraged users do better than that. Light users do considerably worse.
It can rearrange an overbooked calendar and defend focus blocks, but it cannot reduce what you have promised. If your commitments exceed your available hours by a third, that is about thirteen hours in a standard week, and the measured savings from AI top out at a couple of hours. Add your commitments up honestly, compare the total to your uncommitted time, and if the gap is large the fix is a conversation with whoever is expecting the work.
Ask for ranged estimates rather than single numbers. Paste your task list and ask for a best case, likely case and bad case in hours for each item, plus what would cause the bad case. Then add the likely cases up and compare against your real free hours. It is unglamorous, it takes four minutes, and it surfaces overcommitment faster than any planning system I have taught.
It can. A randomised trial by the nonprofit METR gave 16 experienced developers 246 real tasks in their own codebases and found they took 19% longer when allowed to use AI tools, while believing they had been sped up by 20%. METR has since said its follow-up study was compromised by selection effects and that later tools probably did help. The durable lesson is the perception gap: people are unreliable judges of their own time savings.
Not to start. Run a manual weekly reset for a month first: dump every commitment, estimate it in ranges, total it against real hours, cut what does not fit, protect two blocks, review on Friday. If you are still hitting the same wall after four weeks, then buy something. Reclaim starts free and runs $10 per seat per month on annual billing, and Microsoft 365 Copilot Business is currently listed from $21 per user per month paid yearly, discounted to $18.
Every figure here was checked against its original source in August 2026, and two widely repeated ones were deliberately left out. The claim that it takes 23 minutes and 15 seconds to refocus after an interruption is attributed everywhere to Gloria Mark's UC Irvine research, and it appears in neither the 2005 field study nor the 2008 experiment, so it isn't cited here. The figure that task switching costs 40% of productive time traces to an American Psychological Association page that would not load for verification, so it went too. Prices came from the vendors' own pricing pages, which is how the stale $30 Copilot figure got corrected.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.