AI administrative work time savings
Abstract
The best current field evidence does not support one universal percentage for AI administrative work time savings. A 2025 working paper covering 7,137 knowledge workers at 66 firms found that regular users with access to an integrated AI tool spent two fewer hours on email per week during the second half of a six month experiment. [1] A 2023 controlled experiment found 40 percent less completion time and 18 percent higher quality on professional writing tasks. [2] These are task level results, not a promise that every administrative role will save the same number of hours.
Methodology
This review prioritizes experiments with measured behavior over opinion surveys. It includes a six month randomized workplace experiment reported in May 2025 and revised in November 2025, a controlled experiment published in Science on July 14, 2023, and a customer support deployment published in May 2025. [1][2][3] A Microsoft technical report is labeled first party because Microsoft makes the product studied. [4] The ILO occupational exposure index, published May 20, 2025, is used to describe technical exposure, not time saved. [5]
The strongest field result: two fewer email hours per week
The broadest direct time result comes from a randomized field experiment across 66 firms and 7,137 knowledge workers. Workers selected for treatment received an AI tool inside applications they already used for email, meetings, and writing. During the second half of the six month experiment, the 80 percent of treated workers who used the tool spent two fewer hours on email each week. They also reduced work outside regular hours. [1]
This result comes from workplace behavior, but it applies to users rather than everyone offered access. The paper did not detect broader changes in the quantity or composition of work from individual access. It does not isolate executive assistants or show that two hours of lower email activity became two hours of additional output. [1]
An April 2025 Microsoft Research report described an earlier version of the experiment with more than 6,000 workers at 56 firms. Workers who used the tool spent about 30 fewer minutes reading email per week and completed documents 12 percent faster. Nearly 40 percent of workers given access used it regularly during the six month study. [4] Access, adoption, and regular use are different denominators.
Why controlled tasks show larger savings
In the Science experiment, 453 college educated professionals completed assignments such as reports, press releases, and emails. Participants with ChatGPT took 40 percent less time on the second task, while independent evaluators rated output quality 18 percent higher. The study was conducted in early 2023 and published on July 14, 2023. [2]
This is causal evidence for bounded writing tasks, not a weekly administrative workload study. Real work includes missing context, approvals, confidential information, relationship judgment, and corrections across systems. The treatment also shifted effort away from drafting and toward editing. [2]
Applying 40 percent to a full workweek would assume that every hour resembles the experiment and that checking, coordination, and tool setup take no time. The field experiment shows why that conversion is unsafe. [1][2]
Customer support measures throughput, not hours returned
The Quarterly Journal of Economics study followed a staggered AI rollout to 5,172 customer support agents. Access increased issues resolved per hour by 15 percent on average. Gains were concentrated among novice and lower skilled workers, while the most experienced workers saw little benefit. The deployment began in late 2020, access expanded sharply by January 2021, and the peer reviewed article was published online on February 4, 2025, before appearing in the May issue. [3]
Customer support includes repeated writing, retrieval, standard processes, and escalation. Still, issues resolved per hour is a throughput measure. It cannot show how many calendar hours an executive support team will recover. The tool also operated in a narrower setting than an assistant handling unrelated inbox, scheduling, travel, and document work. [3]
The experience pattern matters. AI may shorten the time needed to learn a standard response while adding less for someone who already knows it. A pilot should report results by task and experience level.
Exposure is not measured time saved
The ILO refined its occupational exposure index in May 2025 using task classifications, worker input, expert review, and model based scoring. Clerical occupations remained the most exposed group. [5] Exposure means a tool may be able to assist with part of a task. It does not show that a worker used the tool, finished faster, or maintained quality.
An exposure estimate describes technical potential. A randomized experiment estimates change under specific conditions. A product survey records user perceptions. The measures are not interchangeable, and their percentages cannot be added.
What administrative teams can reasonably expect
The evidence supports a bounded conclusion: AI can reduce time on some email, document, writing, and support tasks, but the saving depends on the task and actual use. [1][2][3][4] Repetitive drafting and retrieval with clear source material are better candidates than work that depends on unrecorded preferences, authority, or sensitive context.
Start with one stable workflow and a baseline. The guide to executive assistants using AI tools covers practical uses and review boundaries. The task management guide for executive assistants can help divide work into repeatable stages.
A faster first draft may be followed by citation checks, tone corrections, or permission work. That can improve quality, but it is not a net saving unless the full workflow ends sooner.
How to measure local time savings
Record completion time for the full workflow, including prompt preparation, source review, correction, approval, and system updates. Compare similar tasks with and without the tool. Keep quality checks constant and record actual use rather than treating everyone with access as a user.
Report median time per completed task, rework, and the number completed to the agreed standard. Separate email triage, drafting, meeting follow up, and document search. For a weekly estimate, multiply a measured saving by the local number of eligible tasks. Do not apply a laboratory percentage to untested hours.
Review privacy, access, and human approval before expanding. Teams that need help defining executive support boundaries can contact ExecutiveSupportStaff after documenting the measurement plan.
Limits of the evidence
None of the studies provides a representative estimate for executive assistants. The field experiments cover knowledge workers or support agents, while the controlled study covers short writing tasks. Tool capabilities and workplace policies also changed after the earliest data periods. [1][2][3]
Less email time can reduce work outside normal hours without increasing output. Faster drafting may free time for review rather than create spare capacity. AI has produced real gains in several designs, but their size and destination remain local questions.
Key findings
- Regular users in a 66 firm randomized experiment spent two fewer hours on email per week during the second half of the study. [1]
- In a controlled experiment with 453 professionals, ChatGPT reduced writing time by 40 percent and raised evaluated quality by 18 percent. [2]
- An AI assistant increased customer support issues resolved per hour by 15 percent across 5,172 agents. [3]
- An earlier first party report found documents were completed 12 percent faster among workers who used M365 Copilot. [4]
- ILO exposure estimates do not measure adoption, time saved, or output quality. [5]
Sources
- Shifting Work Patterns with Generative AI, National Bureau of Economic Research
- Experimental evidence on the productivity effects of generative artificial intelligence, Science
- Generative AI at Work, The Quarterly Journal of Economics
- Early Impacts of M365 Copilot, Microsoft Research
- Generative AI and Jobs: A Refined Global Index of Occupational Exposure, International Labour Organization