AI can make your workday longer — unless you cap the work it creates

AI is very good at removing the small obstacle between “I should do this” and “this is done.” That is the appeal. It is also the trap.

When drafting, summarising, coding, researching, and replying become easier, the work does not automatically shrink. The boundary of acceptable work can expand instead. More tasks look feasible. More ideas stay alive. More “quick wins” slip into the gaps that used to be pauses.

The result is a productivity paradox: the tool gets faster while the day gets denser.

The time saving is real. So is the expansion

A 2025 NBER field experiment followed 7,137 knowledge workers across 66 firms. Workers were randomly selected to receive access to a generative AI tool integrated into applications they already used for email, meetings, and writing.

Among the 80% of treated workers who actually used the tool, email time fell by two hours per week. They also reduced the time they worked outside regular hours. That is useful evidence. AI can remove friction from real work.

But the researchers found no change in the quantity or composition of workers’ tasks from individual-level AI access alone. The tool saved time without automatically turning that time into different, higher-value work.

That distinction matters. A faster email workflow is not the same thing as a redesigned job. If nobody decides what the recovered time is for, it usually gets absorbed by the next message, request, or vaguely valuable idea.

A different kind of signal comes from an eight-month UC Berkeley Haas ethnography at a 200-person technology company, including more than 40 interviews. The researchers observed three forms of work intensification:

• People took on work that previously belonged to someone else—or would not have been attempted at all.
• AI-assisted work seeped into lunch breaks, the time before meetings, and evenings.
• Workers kept more threads running at once, sometimes with several AI processes operating in the background.

This was an in-progress study of one company, so it is not a universal measurement of what AI does everywhere. It is still a useful warning. When capability rises faster than boundaries change, “more possible” quietly becomes “expected.”

AI is entering an already crowded workday

The surrounding system is not exactly begging for more activity.

Microsoft’s 2025 Work Trend Index, based on aggregated Microsoft 365 productivity signals and a survey of knowledge workers, reported that the average worker received 117 emails and 153 Teams messages per weekday. Its telemetry showed interruptions by a meeting, email, or notification every two minutes during core hours. Nearly half of surveyed employees said their work felt chaotic and fragmented.

Those figures come from Microsoft’s own ecosystem, so they are not a neutral census of every workplace. They are nevertheless a helpful diagnostic. If your day is already a conveyor belt of messages and meetings, adding a tool that makes it easier to process more work may increase throughput without creating calm.

AI does not need to be the cause of the overload to become its accelerator.

The practical answer: set a capacity cap

The useful question is not, “What else can AI help me do?”

Ask: “What existing work will this make easier, and where does the extra capacity stop?”

A capacity cap is a simple rule that prevents efficiency from becoming an open invitation. Use four parts.

1. Define “done” before you open the tool

Write the finish line in one sentence:

By 15:00, five customer replies are reviewed and sent. No campaign, analysis, or follow-up sequence is being created.

Then use AI inside that boundary. It can draft, sort, or suggest. It does not get to redefine the assignment halfway through.

This sounds almost insultingly obvious. That is usually a sign that it belongs in the system.

2. Reinvest saved time—or deliberately close it

If AI saves 30 minutes on email, decide where those 30 minutes go before the week starts. Put them into one named block: a project milestone, a recovery break, a run, or simply the end of the workday.

“Free time” is not a plan. It is unclaimed capacity, and unclaimed capacity attracts requests.

For teams, make the choice visible. A saved hour should either fund a defined outcome or remain protected. It should not vanish into a larger pile of “while we are here” work.

3. Treat new AI-assisted work as a trade

Every new task needs one of three labels:

• Replace: it removes an older task.
• Defer: it is useful, but not this week.
• Own: someone has explicitly agreed to carry it within existing capacity.

There is no fourth category called “the AI made it easy.” Ease is not ownership.

This is especially important with agentic tools. An agent can keep searching, drafting, checking, and proposing long after the original job was complete. Give it a stopping condition, a maximum number of outputs, and a clear handoff point.

4. Track workload, not just minutes saved

For one week, record four numbers:

1. Minutes saved.
2. Finished outcomes.
3. New tasks created because the work felt easier.
4. Minutes worked outside your normal boundary.

If the first number rises but the second does not, you have an efficiency gain without an outcome gain. If the third and fourth numbers rise too, the tool is helping you run a larger treadmill.

That is not a failure of motivation. It is a design problem.

A better job for an AI assistant

The strongest assistant is not the one that keeps you busy at higher speed. It is the one that helps you close loops.

That means turning a vague intention into one finished next action, keeping new work tied to a project, surfacing what is blocked, and stopping when the agreed outcome is complete. It means protecting the user from the assistant’s own ability to generate more options.

For example, “help me with customer follow-up” is an invitation to create an entire miniature department. A bounded version is clearer:

Find the five overdue customer replies, draft responses using the existing context, flag anything uncertain, and stop after the drafts are ready for review.

The second request saves effort without manufacturing a new workload.

AI can absolutely give time back. But time only becomes a benefit when somebody decides what it is for—and what it is not for.

Set the cap before the tool starts. Finish the important work. Then stop. A shorter to-do list is still a productivity result, even when no dashboard celebrates it.