AI Made Tasks Faster. So Why Am I Still Working Late? · Issue #2

Work Time and Quitting Time Turn Out to Be Different

I use AI pretty actively myself. I turn to it when I’m writing, digging up research, or organizing things. Tasks that used to take an hour sometimes wrap up in 15 minutes now. And yet my quitting time hasn’t budged. If anything, it feels like I’m leaving later than before.

Finishing one task faster and doing less work over the course of a day turn out to be two separate things. I start something else with the time I’ve freed up, and by the time I’ve checked over what AI produced, the day is gone.

There’s a study I read that echoed this experience. Aruna Ranganathan and Sungchi Maggie Yeh of UC Berkeley’s Haas School of Business observed a US tech company with roughly 200 employees over 8 months and conducted more than 40 interviews. The study looked at what work employees took on, and how they spent their time, as they used AI.

Efficiency Gains Can Still Mean Higher Total Use

This brings to mind the Jevons paradox. The economist William Stanley Jevons, in his 1865 book The Coal Question, explained that using coal more efficiently doesn’t necessarily reduce total consumption.

If a product needs less coal to make, production costs fall. And if those lower costs drive a big enough increase in output and in new uses for coal, total coal consumption can end up rising. This isn’t a rule that holds every single time efficiency improves — it depends on how much usage grows in response.

You can ask a similar question about AI. If the time needed to produce one report shrinks, keeping the number of reports the same would save time. But once you start producing and reviewing more reports, working hours might not fall at all.

This is simply borrowing the Jevons paradox as an analogy for work. It doesn’t mean that more AI use necessarily means more human labor.

Three Changes the Researchers Observed

The company the Berkeley researchers studied didn’t mandate AI use. Even so, as employees adopted it voluntarily, workloads still appeared to grow.

First, the scope of work broadened. AI made it possible to try tasks outside one’s own role. But that often meant someone else had to review the new output. One person starting a task faster didn’t eliminate the checking and revising the team as a whole still needed.

Second, it became easier to start working during breaks. Firing off a quick request to AI during lunch or between meetings became more common. It’s a brief request, but it means work creeping into time that was supposed to be rest.

Third, more work happened in parallel. Handing several tasks to AI at once meant a steady stream of checking results and issuing the next instruction. Even as hands-on execution time fell, the amount of work to manage could grow.

At first, people welcomed being able to do more. But once that expanded workload became the new baseline expectation, some reported staying constantly busy and finding it hard to step away from work. The researchers describe this shift as work intensification.

This is an ongoing observational study at a single company. It isn’t representative of every workplace, and it doesn’t prove that heavy AI users are the first to burn out. Some of it overlaps with my own experience, but it’s worth reading with that scope in mind.

Feeling Faster and Actually Being Faster Aren’t the Same

Other surveys have turned up different things.

Upwork surveyed 2,500 executives, employees, and freelancers across the US, UK, Australia, and Canada in 2024. Among workers using AI, 77% reported at least one experience of AI increasing their workload or reducing their productivity. That included time spent reviewing output or learning the tools. Since this is survey data, it should be kept separate from actual measurements of working hours.

METR’s 2025 experiment involved 16 experienced developers completing 246 tasks in open-source repositories they were already familiar with. Under the conditions in place at the time, having AI tools available made completion take 19% longer on average. Yet after the experiment, participants still estimated that AI had made them about 20% faster.

This result shouldn’t be generalized to all development work today. In a follow-up study published on 2026-02-24, METR noted that the tools may have since improved. But it also noted that factors like how participants chose which tasks to do without AI affected the results, making it hard to reliably estimate how much things have actually improved.

What strikes me about this research isn’t which specific tool wins or loses, but the method of measurement. Feeling faster isn’t enough on its own. You need to look at the total time from when a task starts to when review and revision are finished.

Oswarld’s Lens

There’s a pattern I’ve seen repeat itself while building GTM strategies. When a new tool is introduced, the question that comes up first is always, “What more can we do with this?” The question “What can we stop doing because of this?” comes later — if it comes up at all.

So this research wasn’t hugely surprising to me. Being able to do more work with AI, and whether that work is actually necessary, are two separate judgments.

When a leader asks about the impact of adoption, a team member might answer, “Things got faster.” Even so, if you overlook the time spent checking output, the review work pushed onto other teams, and the tasks that carry on after hours, you can misjudge how much load your team is actually carrying.

Learning to write good prompts and choosing the right tools matter. But I think you also need to decide what to stop doing, and where “good enough” ends.

The Berkeley researchers suggest pausing before important decisions to check goals and counterarguments, batching non-urgent AI notifications instead of reacting to each one, and setting aside time to talk with colleagues directly. The point is to build ways of working so that judgment and rest aren’t lost to a constant stream of incoming results.

If you’re applying this at work, I’d suggest tracking three things together, rather than just counting how many tasks got done before and after adopting AI: work that actually disappeared or shrank, new review and management work that appeared, and tasks still unfinished by the time you clock out. That’s how you find out where the saved time actually went.

References & Further Reading

UC Berkeley Haas, interview with researcher Sungchi Maggie Yeh — Covers the findings and recommendations of a study observing one tech company. 2026-02-18.
Upwork, AI-Enhanced Work Models — Survey results and methodology from the 2024 survey.
METR, 2025 Experienced Developer Experiment — Explains the tasks, the tools available at the time, and the study’s limitations.
METR, February 2026 Follow-up Study — Covers the changed development environment and the difficulty of measurement.
William Stanley Jevons, The Coal Question — The original text discussing the relationship between efficiency gains and coal consumption.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.