The two hours
Among people who use AI at work, self-reported time savings averaged 5.4% of work hours, roughly 2.2 hours in a 40-hour week (Bick, Blandin & Deming, “The Rapid Adoption of Generative AI”, NBER WP 32966 / Federal Reserve Bank of St. Louis; nationally representative US survey). A second large study puts the figure lower, at 2.8% of work hours or about 1.1 hours (Humlum & Vestergaard, NBER WP 33777; 25,000+ workers across 7,000 Danish workplaces, linked to administrative records). The defensible range is therefore about 1 to 2 hours per week per user; we display the upper study and round it down to two. Neither study was run by a vendor.
The staff figures
Two hours per person, multiplied by headcount, rounded down from the 2.2 hours the study reports. The multiplication assumes every coach on staff becomes a regular user, which is exactly what usually doesn't happen, and the reason the adoption figure sits next to it. Both savings figures also assume a 40-hour week; in-season weeks routinely run well past that, which these numbers do not account for.
Which work the savings come from
Drafting and writing show the largest measured gains: task completion time fell 40% and rated quality rose 18% in a randomized trial of 453 professionals (Noy & Zhang, Science, 2023). Routine requests and triage showed 14% more resolved per hour, rising to 34% for the least experienced staff, across 5,179 support agents (Brynjolfsson, Li & Raymond, NBER WP 31161). Email savings ran to more than half an hour a week for regular users. All measured, not self-reported. The same email trial found no significant reduction in meeting time, and no clean figure exists for data entry or reporting, which is why this page doesn't claim savings on scheduling or logistics. That trial (Dillon, Jaffe, Peng & Cambon, “Early Impacts of M365 Copilot”, 2025; 6,000+ workers at 56 firms) was run by Microsoft, which makes the product it tested.
The 6 in 10
In that same trial, only about 40% of workers handed a license went on to use the tool regularly; the remaining ~60% is the figure quoted above. Adoption is what the research supports as the lever. No study isolates how much time training itself saves, so we don't claim one.
The caveats
Both headline figures are self-reported and may exceed what could be measured objectively; the Danish study found no detectable change in recorded working hours despite users reporting savings. Nobody has yet measured a worker's total weekly hours saved with objective instrumentation.