AI can reduce the time a task takes while leaving someone with more work to do. The difference depends on checking, new assignments and what happens to the time recovered. A faster first draft is useful evidence about one stage of a job. It is not enough evidence to raise the team's delivery target.
For managers, the practical question is whether accepted work takes less total effort across the people involved. Check the writer, the reviewer and the recipient before deciding that capacity has increased. Then make an explicit choice about whether any gain goes toward fewer extra hours, better quality, learning or more output.
What the research can and cannot tell you
A randomized field experiment by Eleanor Wiske Dillon and colleagues covered 7,137 knowledge workers at 66 firms. In the second half of the six-month experiment, employees offered Microsoft 365 Copilot spent about 1.4 fewer hours a week in email sessions than the comparison group. Digital activity outside working hours also fell; average meeting time did not change significantly. These are findings from the current paper's work-pattern analysis, not a promise about every job.
The study measured application activity, not every minute of work or the quality of each output. Participating firms were early adopters, and three authors were Microsoft employees. Its evidence supports the possibility of real time savings while leaving important questions about overall workload open.
A different kind of study helps explain why work can still feel heavier. In an eight-month study at one US technology company, Xingqi Maggie Ye and Aruna Ranganathan observed work and conducted more than 40 interviews. People expanded the tasks they attempted, brought work into breaks and evenings, and ran more activities in parallel. Initial enthusiasm could grow into pressure to keep up with the new pace.
That account describes mechanisms at one roughly 200-person company. It does not establish how common they are or prove that AI always increases workload. The studies ask different questions in different settings. Together, they give managers a reason to examine both task savings and the commitments that follow.
Follow the work beyond the person using AI
The benefit can be quite ordinary. In a September 2026 discussion, a Reddit user described speaking their thoughts, using AI to organize them, then editing and finalizing the result. They found that easier than starting from a blank page. Their account still included human editing, rather than treating the generated draft as finished work.
In the same thread, another Reddit user described receiving long AI-assisted action emails and poorly considered requests. The recipients had to work out what the requests meant and whether they were feasible. These are individual, unverified accounts, not measurements of an organization's performance. Their contrast makes a useful review question concrete: did effort disappear, or did somebody else inherit it?
Suppose an insurance customer-service team uses an approved AI tool to draft replies about changes to customers' policies. The adviser supplies the relevant policy details, edits the draft and sends it to a supervisor when approval is required. The meaningful output is an accurate, understandable reply, not a page of generated text.
If drafting becomes faster but the supervisor must correct more coverage details, the team has not yet demonstrated extra capacity. If the replies are accurate and checking also becomes easier, there may be a gain. Follow the reply through approval and any customer follow-up before changing the number each adviser is expected to complete.
Compare a complete piece of work
Choose a recurring task with a clear finish, such as an approved policy-change reply. Compare similar cases before and after the change. A routine address update and a complicated coverage question are different workloads, even if both end in an email.
Use a small, agreed sample and discuss the results with the people doing the work. Record enough detail to make a capacity decision without turning the review into continuous individual surveillance. The following questions are a starting point for that conversation.
| What to compare | What to record |
|---|---|
Accepted output | Replies approved or cases resolved, with comparable complexity and the same quality standard. |
Preparation and drafting | Time finding inputs, prompting and producing a usable draft. |
Checking and repair | Adviser edits, supervisor review and work returned for correction. |
Recipient effort | Clarification requests, avoidable follow-up and effort required to interpret the result. |
Added commitments | New duties, more parallel cases or shorter promised response times. |
Working pattern | Extra hours, interrupted breaks and whether people can finish within their agreed schedule. |
Do not combine these into one reassuring score. Faster drafting and more evening work can both be true. If the team produces more acceptable replies in the same hours, that may be a productivity improvement, but it does not establish that employees have less work or a less demanding day.
Keep the comparison honest about other changes. Staff absence, unusually difficult cases, a new approval rule or a seasonal surge can affect the result. Repeat the comparison when those conditions change instead of attributing everything to the tool.
Decide where recovered capacity goes
A manager should make the next commitment explicit. For the insurance team, that might mean using a demonstrated saving to clear an existing backlog, improve replies on difficult cases or reduce regular overtime. Those are different choices, with different success measures.
If using recovered capacity could change someone's role, explain which decisions have been made and which remain open. The guide to discussing job uncertainty helps managers answer that employment question before asking for more participation.
Before adding volume, agree three things with the people responsible for delivery.
- What stays fixed. Retain the quality standard, required approvals and agreed working hours during the comparison.
- What changes. Name the additional output or the work that will stop. Include the reviewer and any other team receiving more work.
- When to reconsider. Set a review date and name the signs that require an earlier adjustment, such as growing correction queues or work repeatedly spilling beyond the agreed day.
If the gain will fund learning, put that allocation into the work plan. The guide to making time for AI learning explains how to agree cover and move existing commitments. Do not count the same recovered hour once as learning time and again as extra delivery capacity.
Employees should also be able to report an unsuccessful use without having to defend the whole adoption program. “This helps me organize a reply, but checking it takes longer for complex cases” is useful information for deciding where the tool belongs. The broader people and change guide places that feedback within the work of adoption.
For the next workload conversation, bring one completed task through every handoff. Leave with a decision about targets, quality or time, an owner and a date to check the effect. Tool usage alone should not decide which commitment comes next.
Questions and answers
Does better AI productivity mean employees have less work?
No. Productivity describes output relative to inputs; workload also depends on how much work is assigned and what it demands. A team may complete more accurate replies in the same hours while carrying more cases or switching between more tasks. That is different from a shorter or easier working day.
Before claiming workload relief, compare accepted output, checking, added duties and working hours. The complete-work comparison shows what to include beyond drafting time.
How should managers measure whether AI saves time?
Measure a comparable task from preparation through acceptance, including corrections and downstream clarification. Keep the quality standard visible and ask reviewers and recipients about work they have inherited. Application usage or a self-reported drafting saving alone cannot establish the team's total gain.
For example, compare approved policy-change replies of similar complexity, rather than counting every generated reply as completed work. Record other changes such as staffing or demand, then repeat the comparison before making a lasting target change.
What can employees do when AI leads to higher expectations?
Bring specific examples of the changed work to the person who owns the target. Show what became faster, which checking remains and what additional duties or deadlines appeared. Ask for a priority decision, including what can stop or move if the new target cannot fit within agreed hours.
A useful agreement names the target, quality requirement, support and review date. The capacity decision gives managers a structure for that discussion. It cannot guarantee that every workplace will respond well, but it makes the conflict concrete enough to address.
Should every time saving become a higher delivery target?
No. A demonstrated gain can support additional output, reduced overtime, quality improvement or learning. The choice depends on the team's obligations and the people affected. First check that the saving persists through review and does not create a new bottleneck elsewhere.
If a manager chooses more output, record what changes and when to reassess it. Avoid promising the same capacity to several purposes at once, and revisit the decision if corrections or extra hours rise.
