Underwrite the AI enablement payback period at nine to thirteen months. If usage holds at three or more days a week and the saved hours land on work that was genuinely backed up, payback arrives nearer three to six. The spread is not hedging; it is the shape of the number, and anyone handing you a single figure has quietly chosen your inputs for you.
Here is what the calculators skip. Payback turns almost entirely on one input, hours saved per person per week, and the two best studies of it disagree by a factor of five. Every benchmark month you have been shown is a bet on which one is right.
So the useful answer is a band, the four inputs that move you inside it, and a date to stop waiting.
The short answer, in bands
Start with the row that fits what you are buying.
| What you are buying | Band to plan on | What has to be true |
|---|---|---|
| One cohort on the team’s own work, plus a follow-up pass | 3 to 6 months | Half the room still uses an assistant three days a week at week eight |
| The same, underwritten for a board paper | 9 to 13 months | Nothing. This is the number that survives a bad case |
| One cohort, no follow-up, no manager asking | No payback | Usage decays before any hour is reallocated |
| Several hundred staff in sequenced cohorts | 6 to 12 months | Later groups cost less per head and copy a working internal example |
| Enablement charged with the full license bill | 18 to 30 months | Rarely the right accounting. See below |
| Training bought for a policy or an insurer | Not a payback question | The deliverable is a dated record, so count completion |
Two rows deserve the argument.
The 9-to-13-month row is what you commit to. It assumes half the room becomes a regular user, takes the lower of the two time-savings estimates, and discounts saved hours because most of them vanish into the day rather than into output.
The 3-to-6-month row is what you steer toward. Not optimism, just what the arithmetic produces once usage is daily rather than occasional. The gap between the two rows is a management gap, not a vendor one.
The only input that matters disagrees with itself by a factor of five
Two serious studies measured hours saved, and the gap between them decides your answer.
The national survey panel run by Alexander Bick of the Federal Reserve Bank of St. Louis with Adam Blandin and David Deming asked US workers how many extra hours they would have needed to finish last week’s work without an assistant. Among people who used one for work, the mean saving was 5.4% of work hours, about 2.2 hours a week on a 40-hour schedule, and 1.4% across all workers once non-users are counted.
The LSE Inclusion Initiative, with Protiviti, surveyed roughly 3,000 workers and 240 executives and found something far larger. Employees who had received AI training reported 11 hours a week saved against 5 hours for those who had not, with 7.5 hours across AI users as a whole.
Both are self-reported and both are credible. One implies a trained employee hands back a day a week; the other, a long coffee break.
Build on the LSE figure and you show breakeven in weeks. Build on the St. Louis figure and you show a year.
Where they agree matters more. Savings track usage intensity: in the St. Louis panel, 33.3% of daily users saved four hours or more in the prior week against 11.5% of once-a-week users, and the trained LSE cohort both used AI more (93% against 57%) and saved more.
So the lever is not which hours estimate you pick. It is how often people use the thing, the one variable a well-run program can change.
Price the hour the way your controller would
An hour saved is not worth an hour of salary, and is worth nothing until someone reallocates it.
Use employer cost, not pay. The Bureau of Labor Statistics puts total compensation for private industry workers at $46.89 per hour worked in June 2026: $32.82 in wages, $14.07 in benefits. Thirty percent of an hour sits outside the paycheck, so a model built on base salary gets marked down on first review.
Then apply a realization factor, the share of saved hours that becomes output you can point at. A salaried team that finishes at the same hour has released no cash. The hour counts when it lands on backed-up work, a hire you skip, a deadline you now meet, or a deliverable that used to get bumped.
For a first engagement, honest factors run 20% to 50%. Below 20% is a team that got faster at nothing. Above 50% means a bottleneck sat next to the trained group and someone moved work onto it.
That factor, not the hours estimate, separates serious models from sales math. Our guide to putting an enablement case in front of finance sets and defends it line by line.
Two clocks are running, and only one of them is payback
Programs get called failures at month three because the wrong clock was read.
The adoption clock moves in weeks. People open an assistant, use it on a real document, and either keep going or drift. You see this by week two, judge it by week eight, and it is the only thing visible early.
The reallocation clock moves in quarters. Capacity shows up, a manager points it at something, and an output changes. None of that happens inside 30 days, whatever the pilot dashboard says.
Payback lives on the second clock. At day 45 you are not measuring return but whether return is still possible. At 70% weekly usage in week six with no output change yet, you are on track; at 20%, you are finished, and patience will not fix it.
The numbers to trust sit in our measurement playbook for proving capability moved.
The arithmetic, worked on a 40-person cohort
Take a 120-person company putting 40 people through hands-on sessions on their own documents. Market rates for scoped engagements run roughly $2,000 to $15,000, so call the invoice $12,000, or $300 a head. Then count the cost the invoice does not show.
| Cost line | Amount |
|---|---|
| Engagement fee | $12,000 |
| 40 people out of production 6 hours, at $46.89 | $11,254 |
| Total first cost | $23,254 |
The room is about half the real cost, which is why per-head price comparisons mislead. Our cost-per-head against cost-per-hour-saved comparison takes that apart.
Now vary only adoption, hours saved per user (2.2, 3.5 and 5.0 a week) and realization.
| Scenario | Regular users | Gross hours/week | Realization | Value/month | Payback |
|---|---|---|---|---|---|
| Underwrite (low) | 20 of 40 | 44 | 20% | $1,787 | 13 months |
| Expect (mid) | 26 of 40 | 91 | 35% | $6,470 | 3.6 months |
| Good (high) | 32 of 40 | 160 | 50% | $16,250 | 1.4 months |
A nine-fold spread from one engagement, driven by how many people kept going and whether anyone redirected the time. Any page quoting a median payback of 4.2 months has picked one point inside this band and called it a benchmark.
One more line changes the picture. Charge the assistant seats to this case, at roughly $20 to $30 a user a month on published business pricing from Microsoft and OpenAI, and the low scenario stretches past two years. Which raises the accounting question nobody in a sales meeting opens.
Why the number in the vendor deck is not your payback period
The $3.70 returned per dollar that appears in nearly every AI training deck comes from IDC’s 2024 AI opportunity study of 2,109 firms. Three things about it rarely reach the slide: Microsoft commissioned it, the returns are self-reported by buyers, and it measures AI spend in general, not training.
The number that goes unquoted is Deloitte’s. In its 2025 review of AI returns, only 6% of firms reported payback inside a year, though 38% expected returns within a year of investing. Most put satisfactory return on a typical use case at two to four years, and even among the projects respondents called their best, 13% saw returns inside twelve months.
The two findings do not conflict once you see what each measures. Deloitte’s respondents are mostly describing platform builds, custom models and data work, where the numerator is seven figures and the dependencies are long. Per-seat enablement is a different shape of spend: small, fast, attached to tools the team already owns.
That is the real case for the budget line, and it is stronger than the inflated one. You are not forecasting a return on a platform; you are making an expense you already committed to produce something, which is the argument in why buying licenses is not the same as building capability.
It also settles the license question. Seat cost belongs in the software line where it already sits, and the enablement case should be judged on whether it raised the yield of that spend. Counting them twice makes every training proposal look bad and says nothing about the training.
Four inputs that move your payback date
Four things move the month you break even. Two of them are yours.
How often people use it after week two. Weekly to daily use roughly triples the share of people saving four hours or more. This is the biggest number in the model, set by what managers do in the month after the session.
Whether the training sat on real work. A session run on the team’s own documents builds habits; one run on sample prompts produces a pleasant afternoon. That is a vendor input, and worth being fussy about in procurement.
Who was in the room. The St. Louis panel found computer, mathematical and management occupations saving 2.1% to 2.5% of total hours against 0.4% in personal service roles. Train the people whose week is made of writing, reading and deciding.
Whether a bottleneck was named first. Realization is not passive. Teams that named the backed-up work in advance reallocate hours within weeks; teams that trained first and went looking later mostly do not.
Everything else, format, length, platform choice, is noise against those four.
Set a kill date before you sign
Write three dates into the plan at purchase.
Day 14, usage check. What share of the trained group used an assistant for real work in the last five days? Under 40% means the session did not land, so ask for the follow-up pass while the skill is warm.
Day 60, cadence check. What share uses one three or more days a week? Over 50% and your mid scenario is live. Under 30% and you are heading for the low case, and the fix is a short clinic plus a manager conversation, not more training.
Day 180, decision date. Has one named output moved? Cycle time on a recurring deliverable, a backlog cleared, a vacancy left open on purpose. If usage is healthy and nothing moved, reallocation is the problem. If usage is dead too, stop and write it up.
That last clause is the discipline the calculator pages omit; a payback model with no kill date is a hope with a spreadsheet attached.
Firms of 100 to 249 employees are now at 32% AI use in any business function by the Census Bureau’s count, so your comparison set is no longer firms doing nothing. It is firms doing this with varying rigor.
Frequently asked questions
How long before AI training pays for itself?
Three to six months when usage holds and saved time gets reallocated, nine to thirteen on conservative assumptions, never when neither happens. Commit to the longer band and manage toward the shorter one. On a $12,000 engagement for 40 people, counting their time in the room, the low case breaks even near month thirteen, the mid case near month four.
What payback period should I put in front of my CFO?
Give a range with the inputs visible. Show the hours assumption, the hourly cost basis, the adoption rate and the realization factor as four numbers your CFO can argue with. A model that can be picked apart survives review; a single confident month count invites someone to find the assumption you buried.
Should license costs count against the training payback?
No, in almost every case. The seats are already committed and sit in the software budget, and enablement is the spend that makes them earn. If finance insists, run both versions and show that seat cost is unchanged by the training decision, so it belongs outside the comparison.
How do I measure AI adoption after a workshop?
Count people using an assistant for real work three or more days a week, from self-report plus whatever admin console your provider offers, at day 14 and day 60. Seats provisioned and message volume both overstate adoption, because one person running bulk tasks can carry a whole dashboard.
Is $3.70 returned per dollar a realistic number for AI training?
Treat it as a ceiling claim, not a planning figure. It comes from a Microsoft-commissioned IDC survey of self-reported returns on AI spend in general, and the same report puts top performers at $10.30, which shows the spread. Build your own arithmetic from hours, employer hourly cost and a realization factor.
What if nobody on my team tracks hours?
You do not need timesheets. Pick one recurring deliverable with a clear start and end, time it now, and time it again 90 days after the session. One measured workflow beats a company-wide hours survey and gives a board a number it will accept.
Does payback get faster on the second cohort?
Usually yes. Per-head cost falls as you sequence cohorts, and later groups copy working internal examples rather than starting cold, which lifts adoption. Gains concentrate in the second and third group, which is why one pilot is a poor basis for a go or no-go call.
What is a realistic realization factor for a first engagement?
Between 20% and 50% of gross saved hours. Use 20% for a team with no named bottleneck and no manager redirecting work, and 35% when one backed-up workflow has been named in advance. Reach for 50% only when a known constraint already costs money and the trained group sits next to it.
When should I call an AI enablement program a failure?
At day 60 if fewer than 30% of the trained group use an assistant three days a week, and at day 180 if usage is healthy but no named output has moved. The first is a learning problem and a follow-up pass is the fix. The second is a management problem, and more training will not touch it.
Key takeaways
- Plan on nine to thirteen months, steer toward three to six, and show the inputs rather than the conclusion.
- The hours-saved input carries a five-fold error bar between the two best studies, so any benchmark month is someone else’s assumption wearing your logo.
- Price hours at employer cost, $46.89 in private industry, then discount by a realization factor of 20% to 50%.
- Half the real first cost is the time of the people in the room, not the invoice.
- How often people use an assistant after week two moves payback more than anything a vendor controls.
- Set the day 14, day 60 and day 180 checks before you sign, and treat day 180 as a decision, not a review.
To run these bands against your own headcount, roles and backed-up work before a number goes to your board, book a free AI readiness call. We will build the arithmetic with you, including the version where the answer is to wait.
Arthur Wandzel
