Two point two hours a week.
That is what the St. Louis Fed measured for workers who already use generative AI. Multiply by 52 and you get about 114 hours per employee per year. Roughly three work weeks. Not a vendor slide. Source: St. Louis Fed Real-Time Population Survey (November 2024 data, published February 2025).
Be precise about who that covers. That 114 is the average for users -- people who opened the tool in the prior week. Not every seat on the payroll. Economy-wide, counting non-users, the Fed put the hit at about 1.4% of total hours. Users only: about 5.4% of their work week. Three weeks back is real money when you treat it like capacity, not a vibe.
The number is not the ceiling
114 hours is a slice of a much larger mess.
Knowledge workers already burn a huge share of the week on "work about work": chasing status, hopping between apps, hunting files, sitting in meetings that should have been a note. Asana's Anatomy of Work put unnecessary meetings around 103 hours a year, duplicative work around 209, and talking about work around 352. You feel the pattern either way.
GenAI does not invent a new pool of time. It chips at that pool. Writing and communications are the most common workplace use. Searching for information is next. Meeting prep, notes, and follow-ups take a bite. Admin (invoices, scheduling, ticket triage, CRM updates) eats another. Classic workflow automation -- form to sheet to email -- was already reporting about 3.6 hours a week saved in large workforce surveys before chatbots owned the headlines. Annualize that and you are near a full work month.
If someone sells you "AI will give everyone Friday off," check the math. Five percent of a week is useful. It is not twenty. You do not fund a four-day week on Fed averages alone. You fund better throughput and fewer after-hours catch-ups.
Intensity beats "we bought licenses"
Having a login is not adoption.
In the same Fed data, prior-week users split hard. About a third saved an hour or less that week. About a quarter saved roughly two. About one in five saved four or more. Daily users pull far more than once-a-week dabblers. That gap is the whole ROI story. Frequency is the difference between "I tried it once" and "I got a workday back every two weeks."
Small businesses that already use AI tend to self-report stronger effects than the Fed average -- higher savings, plus revenue and efficiency claims in the high 80s and 90s in SMB surveys. Selection bias and real pressure both show up. Small teams feel the hour. They also feel the rework when the tool is wrong.
Where the hours actually come from
Strip the sci-fi. This is what businesses are actually automating in 2025 and 2026:
Inbox and first replies. Draft the first answer. Route the lead. Human on the edge cases.
Meeting notes to action items. Prep packs, summaries, who-owns-what. Some internal deployments report 25 to 35 minutes saved across prep, attendance, and follow-up per meeting.
Proposals and status report drafts. First pass only. Humans still own numbers and promises.
L1 tickets and FAQ deflection. The questions you answer the same way every week.
Invoice and expense data entry. Extract, validate, hand to accounting.
Lead routing and appointment booking. Capture to CRM to calendar without a copy-paste marathon.
Cross-app data moves. Form to CRM to chat to sheet. Boring. Expensive when a person does it.
That list is not glamorous. It is where the 114 hides.
What real teams measured
An education organization piloting workspace AI for educators reported roughly 9 hours a week back on admin, research, planning, and feedback. That is not "AI replaces teachers." That is "stop burning skilled people on paperwork."
A cautious corporate measurement from another large workspace-AI rollout put savings closer to 5.6 hours a month per employee via usage analytics -- about 67 hours a year. Lower than the Fed user average, and a useful sanity check.
Agent-style deployments (AI that completes a defined task, not just drafts text) reclaim serious hour totals at enterprise support volume. For a five-person shop: one workflow, one owner, one metric. Vendor writeups of small firms claim things like 20 hours a week of admin recovered. Treat those as illustrations until you measure your own baseline.
We see the same shape in our own shop. Kief Studio is two founders. We automated the roles we did not hire for. Content research, drafting, quality scoring, publishing, multi-platform distribution, reporting packs -- the pipeline runs on timers. LTFI is how we describe that stack. The point is not our tools. The point is capacity without a headcount chart that matches a 10-person agency.
Time saved is not money banked
Average enterprises convert only about 41% of AI-generated time savings into measurable business outcomes. Top performers convert closer to 71%. Everyone else "feels faster" and watches the calendar fill back up.
Glean's Work AI Index 2026 named the drag: botsitting. Workers spent about 6.4 hours a week managing, checking, and fixing AI output, often more time than they spent producing with it. Separate leadership surveys put roughly 40% of claimed AI time savings lost to rework.
So the Fed's 114 only sticks if you kill the rework loop. Templates. Clear "good enough" bars for L1 work. A human gate where mistakes are expensive. A named owner who turns free hours into booked work, not longer lunches and more meetings. Saved hours that become on-the-job leisure are still a personal win. They are not a P&L win. The Fed flagged that risk. So should you.
What the hours are worth (and what to do)
Back-of-envelope only: 114 hours times $50 fully loaded is about $5,700 per person per year. At $75 (agency, MSP, or senior knowledge work loaded rates), call it about $8,550. A practical stack -- workspace AI, one automation layer, review time -- often costs less than that per seat if people use it weekly. If they do not, you bought expensive wallpaper. Add process automation on top of chat and the ceiling rises (the 3.6 hours/week workflow number annualizes near 187 hours).
Pick two workflows that already hurt. Track hours for one week the old way. Wire automation into those two only. Daily use. Same person owning the metric. Compare next week. Good candidates: first-response email, meeting notes to tasks, invoice capture, lead routing. Bad candidates: "we'll add AI everywhere and see."
Daily use is the line between one hour a week and four. Workflow design is the line between four hours that vanish into busywork and four hours that show up as delivered work, lower overtime, or one more client without another hire.
114 hours per employee per year is available to people who use the tools and redesign the work around them. It is not magic. It is arithmetic.
We build this kind of automation for creators, small businesses, and agencies who want the hours banked, not just discussed. First conversation is free. No commitment. kief.studio/contact
Practical breakdowns and companion guides for members: free membership at kief.studio.