AI ROI calculator: a model your CFO will accept
AI ROI measures the value AI creates (hours saved, output gained, errors avoided) against what it costs (licenses, usage, rollout and review time). Most estimates fail because they assume every employee uses the tool and every saved hour becomes output. This calculator corrects for adoption and realization, so the number survives a finance review.
Below are the calculator, the formula with a worked example, the corrections most estimates skip, and the questions finance will ask. Every number in the calculator is an example value to replace with your own, and every statistic is listed with its source at the end. Last checked October 23, 2026.
AI ROI calculator: what is AI worth to your team?
Enter your seats, the tool's price and the hours it is supposed to save, then correct them for how many people use it, how much of the saved time becomes output and the time AI adds. The calculator returns the annual benefit, annual cost, ROI, payback in months and the break-even hours saved per user per week.
Every number below is an example value, not a benchmark. Replace each one with your own; the results update as you type.
Licenses you pay for, used or not. Example value: 50 seats.
The license price plus any usage fees, averaged per seat. Example value: $30.
What the provider, a pilot or a survey says a regular user saves. Example value: 3 hours.
The share of seats whose holders use the tool in a typical week. Example value: 60%.
The share of saved time that becomes useful output. Example value: 50%.
Pay plus payroll taxes, benefits and overhead, per hour. Work it out with the employee cost calculator. Example value: $50.
One-time, in year one: setup, training sessions and practice. Example value: 4 hours.
Time spent checking and fixing AI output. Example value: 0.5 hours.
52 weeks minus holidays and leave. Example value: 48 weeks.
Results for the example values
- Annual benefit
- $108,000
- 2,160 realized hours a year × $50 an hour
- Annual cost
- $64,000
- Licenses $18,000, rollout and training $10,000, review and QA $36,000
- ROI
- 69%
- Net value of $44,000 in year one
- Payback
- 2.2 months
- $10,000 one-time cost recovered at $4,500 a month
- Break-even hours saved
- 1.8 hours
- Per active user per week for an ROI of 0%; you entered 3
- Active users
- 30
- 50 seats × 60% adoption
ROI 69%, payback 2.2 months, break-even at 1.8 hours saved per active user per week.
The share link carries only the nine numbers above, never a name, email or company.
Start with the break-even figure. If the hours saved you can measure sit well above it, the case holds on modest assumptions; if they sit close to it, the answer depends on adoption and realization, which the rest of this guide shows how to measure.
How do you calculate AI ROI?
AI ROI is the annual benefit minus the annual cost, divided by the annual cost. The benefit is the saved time that becomes useful output, valued at what an hour of that person's time costs the company; the cost is licenses plus the time people spend learning the tool and checking what it produces.
The model behind the calculator, in full:
- Active users = seats × adoption rate.
- Realized hours a year = active users × hours saved per active user per week × realization rate × working weeks a year.
- Annual benefit = realized hours a year × loaded hourly cost.
- Annual cost = licenses (seats × cost per seat per month × 12) + rollout and training (seats × training hours per seat × loaded hourly cost) + review and QA (active users × review hours per week × working weeks × loaded hourly cost).
- ROI = (annual benefit minus annual cost) ÷ annual cost × 100.
- Payback in months = the one-time rollout and training cost ÷ the monthly net value, where the monthly net value is (annual benefit minus licenses minus review and QA) ÷ 12. Licenses and review recur every month, so only the one-time cost has to be paid back.
- Break-even hours saved = the hours saved per active user per week at which ROI is 0%: the annual cost per seat ÷ (adoption rate × realization rate × working weeks × loaded hourly cost).
The loaded hourly cost is what an hour of an employee's time costs the company, not their hourly pay: pay plus payroll taxes, benefits, equipment and overhead, divided by the hours they work. The employee cost calculator works it out from a salary.
A worked example
Fifty people get an AI assistant at $30 per seat per month, and the provider says it saves three hours a week. These are the calculator's example values, not benchmarks:
Worked example with the calculator’s example values
- Active users: 50 seats × 60% adoption = 30 active users
- Realized hours a year: 30 active users × 3 hours saved a week × 50% realization × 48 weeks = 2,160 hours
- Annual benefit: 2,160 hours × $50 loaded hourly cost = $108,000
- Licenses: 50 seats × $30 a month × 12 months = $18,000
- Rollout and training: 50 seats × 4 hours × $50 = $10,000, one-time
- Review and QA: 30 active users × 0.5 hours a week × 48 weeks × $50 = $36,000
- Annual cost: $18,000 + $10,000 + $36,000 = $64,000
- ROI: ($108,000 minus $64,000) ÷ $64,000 = 68.75%, or 69% rounded
- Payback: ($108,000 minus $18,000 minus $36,000) ÷ 12 = $4,500 of net value a month after licenses and review, and $10,000 one-time cost ÷ $4,500 = 2.2 months
- Break-even hours: Per seat, the year costs $1,280 ($360 in licenses, $200 in training and $720 in review), and each claimed hour saved per active user per week is worth $720 a year per seat (60% × 50% × 48 weeks × $50), so 1,280 ÷ 720 = 1.78, or 1.8 hours saved per active user per week rounded
The tool that "saves three hours a week" breaks even at 1.8 hours. That gap, more than the headline ROI, is the number to test.
Why are most AI ROI estimates wrong?
Most AI ROI estimates multiply the hours a tool could save by every seat, count every saved hour as output and leave out the time AI adds. Correcting for adoption, realization and hidden costs takes the worked example from 1,900% to 69%.
| Step | Annual benefit | Annual cost | ROI |
|---|---|---|---|
| Naive estimate: every seat used every week, every saved hour becomes output, licenses are the only cost | $360,000 | $18,000 | 1,900% |
| Correct for adoption (60%) | $216,000 | $18,000 | 1,100% |
| Correct for realization (50%) | $108,000 | $18,000 | 500% |
| Add rollout and training (4 hours per seat) | $108,000 | $28,000 | 286% |
| Add review and QA (0.5 hours per active user per week) | $108,000 | $64,000 | 69% |
Adoption: not everyone with a seat uses it
A seat is a purchase, not a user. In the third quarter of 2026, 34.8% of employed US adults aged 18 to 64 had used generative AI for work in the previous week (Federal Reserve Bank of St. Louis, Real-Time Population Survey). In Gallup's May 2026 survey of 22,573 employed US adults, 52% used AI in their role at all, but only 30% used it a few times a week or more and 15% daily.
How a company rolls out the tool matters as much as the tool. In a late-2024 survey of 25,000 workers in 11 occupations across 7,000 Danish workplaces, about 40% had used AI chatbots at work where the employer neither encouraged use nor provided tools or training, against 93% where the employer encouraged use, provided an enterprise chatbot and trained people (Humlum and Vestergaard, NBER working paper, revised March 2026).
Use the adoption rate you measure: the share of seats whose holders used the tool in a typical recent week, not the share who have ever signed in.
Realization: saved time doesn't automatically become output
Saved time goes somewhere, but not necessarily into more or better output. In the same Danish study, 85% of AI chatbot users said they moved the time AI saved them to other job tasks, yet the researchers found no effect of chatbot use on earnings or recorded hours two years after ChatGPT's launch, ruling out average effects larger than 2%. Firms report the same gap: around 70% of firms in the US, UK, Germany and Australia actively use AI, but more than 80% report no impact on employment or productivity over the past three years (Federal Reserve Bank of Atlanta, survey of almost 6,000 CFOs, CEOs and executives, March 2026).
So set the realization rate from evidence: what the saved time went into, and whether the team's own output measure moved. Until you can show either, use a low rate and test it.
Hidden costs: rollout, training and review
The license is rarely the whole cost. People spend time learning the tool, and someone has to check what it produces. In the Danish study, about 8% of chatbot users had taken on entirely new tasks because of chatbots, rising to about 17% in workplaces with active employer initiatives, and those tasks included reviewing AI output for quality and compliance. In the worked example, licenses are $18,000 of the $64,000 annual cost and review time is $36,000.
How do you measure AI ROI inputs: actual usage, not survey claims?
Measure what people do, not what they report: weekly active users from the AI tool's admin console or sign-in logs, time in AI tools from app and website analytics on work computers, and time saved from timed comparisons of the same tasks with and without AI. Use surveys to learn why people use AI or don't, not how much time it saves.
Self-reported time savings are the weakest input. In a randomized trial of early-2025 AI tools, 16 experienced open-source developers took 19% longer to finish their tasks when AI was allowed, yet afterwards they estimated that AI had sped them up by 20% (METR, July 2025). METR said in February 2026 that those results no longer reflect current AI tools, and that developers' self-reported speedups "can be quite unreliable".
| Input | Weak source | Better source |
|---|---|---|
| Seats and cost per seat | The list price | The invoice, including usage fees and add-ons |
| Adoption rate | Seats assigned, or anyone who ever signed in | Weekly active users from the tool's admin console or sign-in logs, or time in AI tools by team from app and website analytics |
| Hours saved | The provider's claim or a one-off survey | Timed samples of the same task with and without AI, or a pilot team compared with a similar team without the tool |
| Realization rate | An assumption | The team's own output measure, such as tickets resolved per hour or days to close the month, before and after rollout |
| Loaded hourly cost | Hourly pay | Pay plus payroll taxes, benefits and overhead, from the employee cost calculator |
| Rollout and training hours | Left out | Hours spent on setup, training sessions and practice, from calendars or a short log |
| Review and QA hours | Left out | Two weeks of reviewers logging the time they spend checking and fixing AI output |
Our guide to AI adoption metrics covers the adoption side in detail: eight metrics, how to collect each one, and a benchmark table to fill in for each team.
What will finance ask about your AI ROI?
Finance will ask three questions: when the money comes back (payback), how wrong the inputs can be before the case fails (sensitivity), and whether the gain came from AI or from something else (attribution). Answer each before the meeting.
Payback: when does the money come back?
Payback is the number of months the monthly net value takes to cover the one-time cost. Licenses are paid as you go, so payback mostly reflects the rollout and training effort: in the worked example, $10,000 of rollout and training is recovered by $4,500 a month of net value in 2.2 months. If the monthly benefit doesn't cover licenses and review, there is no payback, however long you wait.
Sensitivity: how wrong can the inputs be?
Adoption and realization are the inputs most likely to be wrong, so show the ROI across a range of both. In the worked example, the case survives lower adoption but not lower realization: at 25% realization, ROI is negative at every adoption rate in the table.
| Adoption rate | 25% realization | 50% realization | 75% realization |
|---|---|---|---|
| 40% adoption | -31% | 38% | 108% |
| 60% adoption | -16% | 69% | 153% |
| 80% adoption | -5% | 89% | 184% |
The break-even hours in the calculator do the same job in one number. Rerun the calculator with your pessimistic adoption and realization rates, and check that the hours saved you can measure stay above the new break-even.
Attribution: was it the AI?
Output rises and falls for many reasons, such as seasonality, new hires or a new process. To credit AI, compare with a baseline that didn't get it: the same team before rollout, or a similar team that gets the tool later. Researchers used that design when a company introduced an AI assistant to its customer support agents in stages, and found that it raised issues resolved per hour by 15% on average across 5,172 agents (Brynjolfsson, Li and Raymond, The Quarterly Journal of Economics, 2025). A staggered rollout gives you the same comparison at no extra cost.
How does Intelogos measure real AI usage and time?
Intelogos is a workforce analytics platform that shows how work time is spent across apps and websites, without screenshots or keystroke logging, and lets managers ask questions about it in plain language. It records the AI apps and websites people use on work computers like any other tool and groups them in a built-in AI time category for AI tools and assistants, so two inputs of the ROI model come from measured use instead of survey answers.
- Which AI tools each department uses, and for how long. On the Tools page, filter to the AI time category and a department to see every AI app and website that department used in the date range you pick, with its total time, activity and share of all tool time.
- Time in AI tools next to other work. The Time Categories chart on the Dashboard and Profile pages shows time in AI tools alongside development, communication, documents and the other kinds of work.
- Before and after a rollout. Pick a date range before the rollout and one after it to see whether time in AI tools grew in the departments that got the tool, and where it didn't.
What it doesn't capture matters as much. Intelogos doesn't take screenshots. Intelogos doesn't record what people type; it records only whether keyboard or mouse input happened. So it shows how much time a department spends in AI tools, not what anyone asked them or whether the answers were right; the output measures and review samples above cover that. It covers work computers with the Intelogos agent installed. An AI tool Intelogos doesn't recognize yet appears in the Other time category until someone moves it to AI on the Tools page. The Tools page is available to owners, and to admins and managers with the App Management permission.
The same view shows AI tools nobody approved; our guide to shadow AI covers what to do about them.
Frequently Asked Questions
What is a good ROI for AI?
A good AI ROI is one that stays positive after you correct for adoption, realization, training and review time, and still holds at pessimistic adoption and realization rates. There is no reliable published benchmark, and many companies can't yet say: in ISACA's May 2026 poll of more than 3,400 digital trust professionals, 22% said AI's return had met or exceeded their expectations, 23% said it was too early to tell and 22% didn't know. In the worked example above, the corrected ROI is 69%, and it turns negative if only a quarter of the saved time becomes output.
How long does payback take for AI tools?
Payback depends on two numbers: the one-time rollout and training cost, and the monthly net value after licenses and review time. In the worked example, $10,000 of rollout and training is recovered by $4,500 a month of net value in 2.2 months. If the monthly net value is zero or negative, the tool never pays back, however long you wait.
How do you measure productivity gains from AI?
Compare the same work with and without AI using the team's own output measure, not self-reported time savings. Record a baseline before rollout and compare it with the period after, ideally against a similar team that gets the tool later. Track time in AI tools alongside it, so you can tell a team that didn't use the tool from one that used it without effect. Self-reports alone mislead: in METR's 2025 trial, developers who were 19% slower with AI believed they had been 20% faster.
Sources
Every statistic in this guide comes from the sources below. All were checked on October 23, 2026.
- Federal Reserve Bank of St. Louis, FRED, "Generative Artificial Intelligence, Use Last Week for Work: Employed Adults", updated October 20, 2026: the Real-Time Population Survey's Generative Artificial Intelligence Adoption Tracker by Alexander Bick, Adam Blandin and David Deming, a nationally representative online survey of US adults aged 18 to 64. 34.8% of employed adults used generative AI for work in the previous week in the third quarter of 2026.
- Gallup, "Organizational AI Adoption Jumps Six Points", July 20, 2026: Gallup Panel web survey of 22,573 employed US adults, May 6 to 20, 2026, margin of error ±0.9 percentage points. 52% use AI in their role, 30% a few times a week or more and 15% daily.
- Anders Humlum and Emilie Vestergaard, "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI," NBER Working Paper 33777, May 2025, revised March 2026 (previously circulated as "Large Language Models, Small Labor Market Effects"): surveys of AI chatbot adoption in 11 occupations in Denmark, run with Statistics Denmark, the latest in late 2024 with 25,000 workers across 7,000 workplaces, linked to administrative records through December 2024. Adoption of about 40% without employer initiatives and 93% with encouraged use, an enterprise chatbot and training; 85% of users reallocating time savings to other tasks; about 8% and 17% taking on new tasks; no effect on earnings or recorded hours, ruling out effects larger than 2%.
- Federal Reserve Bank of Atlanta, "Firm Data on AI", Yotzov, Barrero, Bloom and colleagues, working paper 2026-3, March 24, 2026: survey of almost 6,000 CFOs, CEOs and executives in the US, UK, Germany and Australia.
- METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", July 10, 2025: randomized trial with 16 experienced open-source developers and 246 tasks. And METR, "We are Changing our Developer Productivity Experiment Design", February 24, 2026, on why the earlier results are out of date and on self-reported speedups.
- Erik Brynjolfsson, Danielle Li and Lindsey Raymond, "Generative AI at Work," The Quarterly Journal of Economics 140(2), published February 4, 2025: the staggered introduction of an AI assistant for 5,172 customer support agents, which raised issues resolved per hour by 15% on average.
- ISACA, "AI use accelerates while governance and ROI lag", May 5, 2026: 2026 AI Pulse Poll of more than 3,400 digital trust professionals. 22% say AI's return has met or exceeded expectations, 23% that it is too early to tell and 22% that they don't know.