AI time savings at work range from no gain or even a slowdown in some expert tasks to roughly two hours per week in a large knowledge-worker experiment. OECD evidence suggests whole-job savings of about 2.8% to 5.4% of work hours among users. The result depends on the task, worker, adoption rate, review burden, and whether saved capacity is captured.
AI
AI time savings at work range from no gain or even a slowdown in some expert tasks to roughly two hours per week in a large knowledge-worker experiment. OECD evidence suggests whole-job savings of about 2.8% to 5.4% of work hours among users. The result depends on the task, worker, adoption rate, review burden, and whether saved capacity is captured.
Key findings
• A randomized experiment across 66 firms and 7,137 knowledge workers found active users spent two fewer hours per week on email in the second half of a six-month trial.
• The same experiment found reduced after-hours work but no detectable change in the quantity or composition of tasks beyond individual time use.
• Danish workers reported about 25 minutes saved per day when using AI chatbots, yet administrative data showed no measurable earnings or recorded-hours effect larger than 2% after two years.
• OECD summarizes whole-job time savings of 2.8% to 5.4% of work hours among users, much lower than some task-specific experiment results.
• A P&G product-innovation experiment found 16% faster completion for individuals and 13% for teams using AI.
• Experienced open-source developers in a 2025 randomized trial took 19% longer when AI was allowed, showing that assistance can add overhead in complex, familiar work.
Methodology and definitions
This review separates three measures that are often mixed:
Task speed is the change in time needed for a defined assignment. It can be large when the task is well matched to AI.
Whole-job time savings is the share of total work hours recovered after accounting for all tasks and days when AI is not used.
Captured capacity is the saved time the organization converts into more output, faster service, lower overtime, avoided hiring, or another measurable outcome.
The denominator matters. A 30% speed improvement on a task that consumes 10% of the week produces at most a 3% gross weekly saving before review and adoption. If only half of eligible cases use AI, the potential falls again.
We prioritize randomized or quasi-experimental workplace studies, administrative labor evidence, and OECD or Stanford research syntheses. Results are reported for the population and tools tested, not as universal benchmarks.
Two fewer email hours per week
The NBER Shifting Work Patterns with Generative AI study is one of the largest workplace experiments available. Researchers randomly selected workers across 66 firms to receive a generative AI tool integrated into applications for email, meetings, and writing. The study covered 7,137 knowledge workers over six months.
In the second half of the experiment, 80% of treated workers had used the tool. Those users spent two fewer hours on email per week and reduced time working outside regular hours.
The result is meaningful because it reflects real workplace software and a multi-month period. It does not mean every treated worker saved two hours, because the estimate highlighted users. It also does not show that the firms gained two hours of output per user.
The researchers did not detect shifts in the quantity or composition of workers’ tasks from individual-level AI provision. The recovered email time may have improved worker well-being, created unobserved slack, or been used in ways the study could not distinguish.
Some authors were employed by Microsoft, the tool provider, at the time of the study. The disclosure states that Microsoft reviewed the paper for privacy concerns while authors retained discretion over results and estimates.
Twenty-five minutes per day without earnings effects
The NBER working paper Still Waters, Rapid Currents, revised in March 2026, linked large-scale AI-adoption surveys with administrative labor records in Denmark.
An NBER research summary reported that 83% of workers used AI chatbots when employers encouraged them, with average reported savings around 25 minutes per day. The paper found widespread new tasks in content generation, AI oversight, and AI integration.
Despite reported productivity benefits and task reorganization, the researchers found no measurable effects on earnings or recorded hours at worker or workplace level. Their estimates ruled out effects larger than 2% two years after ChatGPT’s launch.
That does not prove the saved minutes were imaginary. Workers may complete the same work with less effort, use the time for unmeasured activities, or face organizational constraints that prevent productivity from reaching pay and hours. The study shows that time savings do not automatically appear in labor-market outcomes.
Why OECD estimates are smaller than task studies
The OECD’s SME workforce report reviews experimental studies with impressive performance gains on selected tasks: about 14% among customer-service agents, nearly 40% among business consultants in a bounded exercise, and more than 50% among programmers in a coding task.
The report cautions that these tasks were selected because they were suitable for generative AI. Across the full workweek, AI applies to only part of most jobs.
OECD cites a Danish estimate of 2.8% of work hours saved among users and a US estimate of 5.4%. Users often applied generative AI on some rather than all workdays, and only for a fraction of the day. This dilution from task to job is expected.
A company should therefore avoid inserting a 40% laboratory speedup into a payroll model. It should estimate task share, eligibility, adoption, review, and exception handling.
Time and quality can move together
The P&G field experiment summarized by Harvard Business School involved 791 professionals working on product-development challenges. AI reduced completion time by 16% for individuals and 13% for teams.
Individuals with AI produced work comparable in quality with two-person teams without AI. Teams using AI produced the highest-quality solutions; ideas in the top 10% were three times more likely to come from AI-assisted teams than from individuals without AI.
This result shows that speed need not require lower quality in a suitable creative workflow. It remains one company’s innovation setting using an internal GPT-4-powered tool. It does not establish the same effect in accounting, support, software maintenance, or regulated decisions.
When AI makes work slower
METR’s early-2025 developer study randomized whether 16 experienced open-source developers could use AI on 246 real tasks in repositories they knew well. With AI, tasks took 19% longer.
Before the study, developers expected a 24% speedup. After experiencing the study, they still believed AI had made them 20% faster. The mismatch shows why self-reported time savings need validation.
This was a specific population: experienced maintainers working in large, familiar codebases with tools available from February through June 2025. METR explicitly cautions against generalizing the result to most software work.
Its February 2026 follow-up showed raw late-2025 estimates consistent with small speedups, but confidence intervals included zero and participation and measurement biases made the signal unreliable. Tool performance was changing faster than the research design could cleanly track.
How SMBs should estimate time savings
For each workflow, calculate:
gross hours saved = eligible cases x baseline minutes x measured speed improvement x adoption rate
Then subtract:
• Human review time.
• Correction and rework.
• Exception handling.
• Additional data preparation.
• Monitoring and workflow maintenance.
• Time spent prompting or waiting for outputs.
The result is net capacity, not financial value. Use the AI automation payback period framework to connect net capacity to labor cost, increased throughput, faster revenue, avoided overtime, or another captured outcome.
Run a two-week baseline before enabling AI. Sample real cases, not only easy examples. Compare median and tail handling time because an average can hide expensive failures. Track quality alongside speed.
Finally, decide how the saved capacity will be used. “Reps respond to 20 more qualified leads per week” is measurable. “The team has more time for strategy” may be valuable, but it needs a defined output.
Limitations
The studies use different tools, periods, tasks, and populations. Some rely on self-reported time. Workplace telemetry can observe application activity without observing the full value of what workers do.
Fast model change shortens the shelf life of point estimates. Selection can matter: enthusiastic users may adopt more, while experts may avoid studies that require working without AI.
What would change our view
We would raise expected time savings if replicated multi-firm studies showed durable gains across full workflows, including review and exceptions, with stable quality. We would lower them if longer studies found that early speed gains fade, correction costs rise, or workers become slower without developing independent skill.
FAQs
How many hours per week does AI save?
One large experiment found two fewer email hours per week among active users. That is a result for a specific integrated tool and worker population, not a universal average.
What percentage of work time does AI save?
OECD cites estimates of about 2.8% to 5.4% of total work hours among users. Task-specific studies can show larger gains because they measure work selected for AI suitability.
Why can AI make a task slower?
Users may spend time prompting, reading, verifying, correcting, or integrating output. Experts working in complex environments may receive suggestions that do not fit context they already understand.
Are self-reported AI time savings reliable?
They are useful but can be biased. METR found experienced developers believed AI sped them up even when measured completion time was 19% longer.
How should a business convert time savings into ROI?
Measure net time after review and exceptions, multiply by actual adoption, then specify how capacity changes output, overtime, hiring, service, or revenue. Subtract implementation and operating costs.
Get a 20-Minute AI Workflow Audit
AI Operator can measure the manual baseline for one workflow and design a real-case test that separates gross time savings from captured value.