Every week a list of the twenty best free AI tools comes out. The format works, and there’s a reason it almost never works for the person reading it: it assumes the bottleneck on your productivity is a shortage of tools.
In 2004 Victor González and Gloria Mark directly observed professionals at work in offices, minute by minute. The result became one of the most cited figures in the field: people spent on average about three minutes on an activity before switching, and about twelve minutes on a working sphere before moving to another. Later studies by the same group, using automatic on-screen activity logging, found average dwell times on a single window down to roughly 47 seconds.
That’s the context the list of twenty tools arrives into. And the reason adding more of them, free or not, isn’t neutral.
What switching tools actually costs
The cost of context switching has been measured in the lab for decades, and it isn’t small.
| Study | What it measures | Result |
|---|---|---|
| Rubinstein, Meyer & Evans 2001 | Task switching in the lab | Loss of productive time up to ~40% in the worst conditions |
| Mark, Gudith & Klocke 2008 | Interruptions in a real office | ~23 minutes to return to the original task |
| Leroy 2009 | Attention residue | Part of attention stays on the previous task, and performance on the new one drops |
| Ward et al. 2017 | Smartphone presence | Available cognitive capacity reduced even with the phone off and face down |
Sophie Leroy’s study is the one that best describes the everyday sensation. What she calls attention residue isn’t the distraction during the interruption: it’s that, when you return, part of your attention is still on the task you left half-finished. And the residue is larger when you leave the previous task incomplete with no clear expectation of when you’ll resume it.
Adrian Ward and colleagues’ 2017 result is more uncomfortable still: the mere presence of participants’ own smartphones on the desk reduced their available cognitive capacity, even when the phone was switched off and face down. You don’t have to use it. Knowing it’s there is enough.
Every new tool is one more place where something might ask for your attention.
The productivity paradox, second edition
There’s a historical precedent worth keeping in mind before expecting miracles.
In 1993 Erik Brynjolfsson published in Communications of the ACM a paper on the productivity paradox of information technology: investment in computing was enormous and aggregate productivity data showed no trace of it. Robert Solow’s line that made it famous came earlier: the computer age was visible everywhere except in the productivity statistics.
In 2021 Brynjolfsson himself, with Daniel Rock and Chad Syverson, proposed an explanation that has held up well: the productivity J-curve. General-purpose technologies require complementary investments (process reorganisation, new skills, different data) that are booked as costs long before the benefits become measurable. At first measured productivity falls. Then it rises, if and when the reorganisation happens.
Applied to your individual case, the model says something precise: adopting a new tool makes you less productive for a while. Adopting twenty at once means staying on the descending part of the curve forever.
And “free”?
It’s worth being explicit about a distinction the lists rarely make.
Free tools fall roughly into three categories: reduced versions of paid products, products financed by data collection, and products in a user-acquisition phase that will become paid. None of the three is dishonest in itself, but they imply different things.
The questions that matter before letting a tool into your workflow are three, and they’re the same for a free product and a paid one:
- Is the data I put in used to train models, and can I opt out?
- If the service shuts down or changes its price tomorrow, what do I lose and in what format can I take my work with me?
- Does the company I work for permit this tool?
The last one causes the most real problems, and it’s also the one the lists never mention.
A tool you use every day is worth more than ten better tools you open once a month.
How to choose them without wasting time
- Start from the task, not the tool. Write down the three activities that cost you the most time in a normal week. If you haven’t measured them, measure them first: two weeks is enough.
- Adopt one at a time, for at least two weeks. That’s the minimum to get off the descending part of the J-curve. Less than that measures novelty, not usefulness.
- Compare against your real time, not against the promise. Same task, before and after, timed. It’s the only way not to fool yourself, and the research on AI shows people fool themselves a lot: generative AI.
- Keep one tool per category. One conversational assistant, one place for notes, one place for tasks. Two tools for the same job always means zero tools for that job.
- Uninstall anything you haven’t used in a month. It costs nothing in money. It costs in attention, and that’s the cost Ward and colleagues documented.
- Check the data settings on day one, not later. If there’s an option to exclude your conversations from training, turn it on before you paste anything work-related in.
- Fix the request before changing the tool. The difference between a mediocre result and a useful one comes far more often from the request than from the product: writing effective prompts.
My take: a criterion ages better than a list
I don’t publish a list of twenty tools because I think it’s the format that has done the most damage to the productivity of everyone who followed it, and because any such list will be wrong in six months. I prefer a criterion that ages more slowly.
I also suspect that looking for the right tool is one of the hardest forms of procrastination to recognise, because it looks a great deal like work. You read, compare, try things, take notes. At the end of the day you’ve produced nothing, but you don’t feel guilty.
I’ll close with a personal bet I can’t prove: most people would get more done using one conversational assistant and one place to write well than with any optimal combination of twenty different applications.
Frequently asked questions
So I shouldn’t use free AI tools? You should use few of them, well. The advice isn’t against tools, it’s against accumulation, which has a measured and almost never accounted cost.
How many tools are too many? The literature gives no threshold. A practical test: if you can’t recall from memory what each one does and when you last used it, it’s too many.
Are free versions really enough? For most office tasks, yes. Differences become relevant at high volume, with long documents, and for enterprise integrations.
Is it risky to put work documents in? It depends on the tool and your company’s policy. The general rule holds: never paste third parties’ personal data or confidential information into unapproved services.
How do I know whether a tool is actually saving me time? By timing the same task with and without it. It’s tedious and it works, and the research shows subjective impressions can be wrong even in sign.
Sources
- González, V. M., & Mark, G. (2004). Constant, Constant, Multi-tasking Craziness: Managing Multiple Working Spheres. CHI ‘04, 113-120. The three minutes per activity and twelve per working sphere.
- Mark, G., Iqbal, S., Czerwinski, M., Johns, P., & Sano, A. (2016). Neurotics Can’t Focus: An In Situ Study of Online Multitasking. CHI ‘16. Automatic measurement of on-screen dwell time.
- Rubinstein, J. S., Meyer, D. E., & Evans, J. E. (2001). Executive Control of Cognitive Processes in Task Switching. Journal of Experimental Psychology: Human Perception and Performance, 27(4), 763-797.
- Mark, G., Gudith, D., & Klocke, U. (2008). The Cost of Interrupted Work: More Speed and Stress. CHI ‘08, 107-110.
- Leroy, S. (2009). Why Is It So Hard to Do My Work? The Challenge of Attention Residue. Organizational Behavior and Human Decision Processes, 109(2), 168-181.
- Ward, A. F., Duke, K., Gneezy, A., & Bos, M. W. (2017). Brain Drain: The Mere Presence of One’s Own Smartphone Reduces Available Cognitive Capacity. Journal of the Association for Consumer Research, 2(2), 140-154.
- Ophir, E., Nass, C., & Wagner, A. D. (2009). Cognitive Control in Media Multitaskers. PNAS, 106(37), 15583-15587.
- Brynjolfsson, E. (1993). The Productivity Paradox of Information Technology. Communications of the ACM, 36(12), 66-77.
- Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The Productivity J-Curve. American Economic Journal: Macroeconomics, 13(1), 333-372.
- Noy, S., & Zhang, W. (2023). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science, 381(6654), 187-192. The real gains when the tool is one and the task is defined.
This article was written following the process described in how we write. If you find a figure reported wrongly, write to me: I’ll correct it and note the correction at the bottom of the article.



