An AI agent can now fill a form for me, revise a file, reconstruct a thread of 30 emails that I once understood and then forgot, or bring me back into a project I left two or three weeks ago (things that required a really high amount of willpower and time in the past, right? I think all of us have delayed some tasks that are not that hard or time-consuming because we perceived them as such, especially when a long time had passed since we worked on them). Thus, these are exactly the tasks I usually find myself delaying. They may take only an hour, sometimes a lot less, but they carry the cost of finding the file, recovering the context, and remembering what I had decided. Now I can ask an agent to do much of that work in minutes.
I think many people, including people who use chatbots every day and are even comfortable using agents to some degree (and I mean here tools like Claude Code and Codex, which work locally on your computer and on your files, not Q&A chatbots), still do not understand how accurate agents have become for bounded work. Maybe I do not fully understand it either (or don’t use them to the max; see the FOMO? You will see how all of the above connects to this feeling in a minute). They need context, and the first attempts may require correction. But essentially, at least in my daily use, after a few rounds they can learn how I organize files, what I want checked, and how I like a task completed. I trust them with mundane work in a way I did not a year ago (even three months ago). Tasks that took hours take minutes today.
And by the way, this is more than my impression. In an experiment with 444 professionals, Shakked Noy and Whitney Zhang found that ChatGPT reduced the average time for a writing task from 27 to 17 minutes, while quality also improved. A later randomized study across 66 firms found that workers assigned an AI tool spent 1.3 fewer hours per week in Outlook on average. Among regular users, the reduction was about 3.6 hours. The time saving is real, at least for some tasks and some parts of work.
With that said, funnily enough, my days have not become shorter. I think you, dear reader, already have a sense of where this is going.
And I think, in the most basic form, the reason could be this: my work does not arrive as a fixed basket of tasks. It is elastic. When an agent makes a project cheaper to resume (in time, commitment, costs, friction.. whatever), I just resume it. When an analysis becomes easier to run, another analysis becomes reasonable. A dormant idea becomes an active commitment. So basically, the saved hour is “absorbed” before I experience it as free time, and the number of things I can plausibly do keeps growing.
A large Danish study offers a useful counterweight to the task-level experiments. Across 25,000 workers in 11 AI-exposed occupations, researchers found no measurable change in recorded hours or earnings two years after ChatGPT. Simply said, they did find work being reorganized through new tasks involving content generation, oversight, and integration. Faster tasks can coexist with an unchanged workweek because the contents of the workweek change.
Simply said, the first two studies show that AI can make specific parts of work faster. The third found that, even with those savings, recorded work hours did not measurably change.
Sources: Noy and Zhang (2023); Dillon and colleagues (2025); Humlum and Vestergaard (2026).
Different studies and time scales; not a causal sequence.
I need to be careful with the story here. Personally speaking, I am in clinical residency. I have more roles, more research interests, and more collaborators around the world than I had before. So my work, as for many people building a career, has naturally increased in volume and responsibility, with or without AI. I cannot separate normal career expansion from AI-enabled expansion in my own life. Still, when I talk with peers in research, academia, and clinical medicine who use agents heavily, I hear a similar description: output feels a lot easier. Things that took me days are sometimes even automated in the background, but I don’t find myself hiking a lot more. Just hanging out more? Reading my favorite book series more? Sitting with family more? I don’t feel that I am fulfilling the “AI dream” (the promise of letting the AI agent do your stuff in the background while you go and just have fun).
I think that, consciously or not, the expected output rises. The expectation comes from bosses and collaborators, but also from ourselves. In a sense, I think that even our own and others’ perceptions of how much effort and time things take have changed that much, so we just say yes and yes and yes.
Sometimes I catch myself thinking that any minute in which an agent is not running is a lost minute. I can be working on one thing while another agent searches files, reviews a draft, or reconstructs a project. Parallel work creates a strange new guilt, in which an idle computer begins to look like unused capacity. And because the visible part of a task became faster, the task itself starts to look cheap, even when the judgment around it still takes time.
So essentially, the work is not really less, nor does it consume less time. The work has moved. Yes, absolutely, I spend less time searching, formatting, and reconstructing context. But at the same time, I spend more time specifying what I want, checking what came back, deciding what matters, and taking responsibility for the result. And even more simply, I am doing more and more, and expecting to do more and more.
Don’t misunderstand me, I am genuinely enjoying this expansion, at least currently. As a past gamer especially, I find working with agents fun, and I think many people outside gaming do too. It is magical in a sense, because we love as humans to see how things are created, how an imagination, an idea, or a mere simple thought could turn into an image, file, video, sentence, or whatever we want it to be. And seeing the advancements with our own eyes (from how GPT-4 hallucinated every two facts and then started creating images and files, to where we are today) is also magical. And maybe, to a degree, this is also a reason why many find themselves “glued” to the chair and just doing more and more. At the end of the day, I am doing projects that would previously have stayed as notes, and I think I am finding more of myself through them. But I also spend more time at the computer, while the promised life outside it has not automatically appeared. And yes, perhaps saved time has to be claimed before productivity claims it, and maybe we are still in the initial transition phase, learning what and how to prioritize, and what is enough. And then maybe, just maybe, we unlock the “AI dream.” But currently, every minute saved arrives already assigned. As one who believes in living intentionally as much as we can (and mind you, this can also be a toxic pursuit, but that is for an entirely different essay), I think even with all the caveats above, even now, especially now, we should start thinking about how to make agents really save us time and help us do more and more of whatever we want, at least for those whose work agents can help with.
Sources and evidence notes
These sources bound the claims in the essay. They support task-level time savings, measured changes in application use, and the absence of a clear reduction in total recorded work hours.
- Noy S, Zhang W. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science. 2023;381(6654):187-192. https://doi.org/10.1126/science.adh2586
- Dillon EW, Jaffe S, Immorlica N, et al. Shifting Work Patterns with Generative AI. NBER Working Paper 33795. 2025. https://www.nber.org/papers/w33795
- Humlum A, Vestergaard E. Large Language Models, Small Labor Market Effects. NBER Working Paper 33777. Revised March 2026. https://www.nber.org/papers/w33777