AI saved an hour. Your boss added three meetings.

AI can make a task quicker while the total workload grows. People may take on more tasks, produce more material that needs review or stretch work into previously quiet time. To understand productivity, measure the complete workflow and decide what the saved capacity is for. Otherwise, every saved hour becomes an invitation issued on behalf of the employee.

University of California's account of Berkeley Haas research describes work intensification in an eight-month study at one US technology company. It is an important qualitative warning, not proof that AI always increases work in every setting. Gallup's October 2026 analysis also shows that reported benefits differ by worker and outcome. Local evidence matters more than a universal productivity slogan.

Separate speed, demand and useful capacity

Speed is how quickly a task is completed. Demand is how much work arrives. Useful capacity is the time and attention available to produce acceptable outcomes. Improving the first does not automatically improve the third when the second is allowed to expand without a decision.

A team may draft reports faster and then receive requests for twice as many reports. Each report may generate more comments, meetings and decisions. That can be worthwhile if the extra work creates value. It should be a conscious choice, with the additional coordination included in the cost.

Ask who owns the trade-off. If every stakeholder can request another output but nobody can reduce the workload, the AI rollout has increased supply while leaving prioritisation unresolved. A faster tool is now serving an unlimited queue.

Use arithmetic that includes the recipient

Consider a hypothetical task that took 120 minutes to prepare and 20 minutes to review. AI reduces preparation to 30 minutes, while verification and recipient clarification take 50 minutes. Net time saved is 60 minutes: 140 before, 80 after. The example illustrates a calculation, not a measured result from a client or study.

Now suppose management doubles the number of requested outputs because the first draft is faster. Total effort becomes 160 minutes instead of the original 140. The tool improved the unit task, while the demand decision increased workload. Those are two separate effects and should be discussed separately.

For your own experiment, record preparation, checking, clarification and rework for comparable tasks. Add setup and learning time where relevant. Estimate uncertainty honestly. A five-case sample can help identify a problem without supporting a sweeping company-wide savings claim.

Watch for work expanding into every gap

AI can make a small action feel easy enough to do between meetings, during lunch or after the official end of the day. Individually, each action seems minor. Together they can remove the pauses that allow people to recover, think and finish existing work.

Do not infer wellbeing from activity logs. Talk to the team about actual demands and working patterns. Some employees may appreciate flexibility; others may feel pressure to remain available because the tool makes another task technically possible. Set expectations rather than assuming every extra output is freely chosen.

CIPD's July 2026 wellbeing guidance treats work demands and job design as part of the picture. My operational recommendation is to include availability boundaries in the rollout, alongside training and quality checks. The tool does not get to negotiate working hours through gradual calendar erosion.

Decide what the saved time should buy

Choose a purpose: reduce backlog, improve quality, create learning capacity or take on a defined additional outcome. A team may need a combination. Make the choice explicit enough that people can tell whether the change is working.

For example, reserve part of the capacity for checking and part for a customer problem the team previously could not reach. Review whether that is happening. If the time is immediately consumed by unrelated urgent requests, the capacity plan has changed and deserves a fresh decision.

Avoid converting an early estimate into an immediate permanent output target. Initial savings can change as unusual cases appear, demand grows and the tool is used on harder work. Establish a period for observation and a decision point rather than announcing that every task should now take half as long.

Reduce the output nobody needs

Ask recipients which reports, summaries or updates they actually use. AI can make obsolete work cheaper to produce, which is sometimes the most efficient way to keep an obsolete process alive. Stopping an unnecessary deliverable may help more than automating it.

The workslop research is useful when polished output moves unfinished reasoning to colleagues. Set a minimum quality standard for the handoff and make checking effort visible. A shorter, verified answer may create more capacity than a comprehensive document with a decorative conclusion.

For the practical standard, see how to reduce AI workslop. For the financial version, read hidden automation costs. The same principle applies: follow the work past the point where the demo ends.

Run a workload review with a real stop option

Choose one recurring workflow, agree a baseline and test the new method over a defined period. Examine quality, effort, working patterns and recipient usefulness. Invite people to name the cases where the assisted method was worse. Without that permission, the review becomes a search for positive anecdotes.

Continue, narrow or stop based on the evidence. It is possible for a tool to be useful for routine cases and unsuitable for sensitive exceptions. It is also possible for the tool to work while the new demand policy does not. Keep those decisions separate so you can fix the right problem.

If your team remains busy without clearer outcomes, look at the work system. I help connect AI adoption with priorities, responsibility and real delivery. Bring one task that got faster and one team whose day got longer. Together they tell a more useful story than the number of licences activated.

Questions people ask

Can AI increase workload even when tasks become faster?

Yes. More requests, expanded scope, checking work and activity outside normal boundaries can consume the time saved on an individual task. Measure total effort and useful outcomes rather than only draft speed. The effect varies by workflow and organisation; it should not be assumed to occur everywhere.

How should a team calculate AI time savings?

Compare similar work before and after, including preparation, verification, recipient clarification, corrections and relevant setup costs. Then account for changes in demand. Keep illustrative estimates separate from measured results and record uncertainty, especially when the sample is small.

What should managers do with time saved by AI?

Agree the intended use of capacity before increasing expectations. Options include quality, backlog reduction, learning and a defined new outcome. Review actual workload and availability boundaries after rollout, and keep a route to narrow or stop uses that create more effort than value.