Free calculator · vendor-neutral · formulas shown
AI ROI Calculator
Payback, NPV and multi-year ROI for any AI initiative
AI ROI is the net benefit an initiative delivers over a fixed period divided by what it costs over the same period. Enter who will use the system, the time it saves them and what it costs to build and run, and the calculator returns ROI, payback in months, net present value and a year-by-year cash flow. At the illustrative defaults, a knowledge assistant for 500 people returns 79% over 3 years and pays back within 16 months.
Estimate ROI, payback and NPV
Start from an illustrative preset or type your own figures, for one project or several candidates in turn. Results update as you type, the arithmetic runs in your browser and nothing you enter is stored or sent. The currency selector changes the symbol only; no exchange rate is applied.
Results
ROI over 3 years
- Net benefit
- Payback period
- NPV at 8%
- Annual benefit, year 1
- Annual benefit, steady state
- Total benefit over 3 years
- Total cost over 3 years
- $325,000 one-off + $300,000 a year
- Active users at steady state
- 27,600 hours saved a year
Illustrative estimate. Replace each default with figures from your own pilot and finance team.
Year by year
| Year | Active users | Benefit | Cost | Net cash flow | Cumulative | Present value |
|---|---|---|---|---|---|---|
| 0 | – | $0 | $325,000 | −$325,000 | −$325,000 | −$325,000 |
| 1 | 150 | $439,000 | $300,000 | $139,000 | −$186,000 | $128,704 |
| 2 | 300 | $878,000 | $300,000 | $578,000 | $392,000 | $495,542 |
| 3 | 300 | $878,000 | $300,000 | $578,000 | $970,000 | $458,835 |
| Total | – | $2,195,000 | $1,225,000 | $970,000 | – | $758,081 |
Year 0 carries the one-off costs. Payback spreads each year’s net cash flow evenly over its months; NPV discounts it from the year end, and the present-value total is the NPV.
How the AI ROI calculator works
These are the exact formulas the calculator runs, in order.
- Active users in year t
= target users × steady-state adoption × ramp, where the ramp applies in year 1 only and is 100% from year 2. - Time value in year t
= active users × hours saved per week × working weeks × cost per hour × share of saved time that becomes value - Annual benefit in year t
= time value + (revenue uplift + error and rework reduction + risk reduction) × ramp - Total cost
= one-off costs + recurring annual costs × years in the horizon - Net benefit
= total benefit − total cost - ROI
= net benefit ÷ total cost × 100 - NPV
= −one-off costs + Σ (annual benefit in year t − recurring costs) ÷ (1 + discount rate)t - Payback is the month in which cumulative net cash flow first reaches zero, with one-off costs at month 0 and each year’s net cash flow spread over its twelve months.
ROI and payback ignore when the money arrives; NPV does not. It discounts each year’s net cash flow, so late benefits count for less, which is why a finance team will usually ask for all three.
How to choose each input
Every default is an illustrative assumption. Replace each one as follows.
Target users and adoption
Count the people who will get access, then estimate how many will use the system in a typical week once rollout settles; licensed seats are a ceiling, not a forecast. For scale, the St. Louis Fed’s adoption tracker found that 39.2% of employed US adults had used generative AI for work in the past week by the second quarter of 2026. The year-1 ramp is first-year adoption as a share of the steady-state level: 50% means year 1 earns half the steady-state benefit.
Hours saved per active user per week
Measure this in a pilot. The public figures are self-reported: users in the UK government’s Microsoft 365 Copilot experiment (20,000 licences, 30 September to 31 December 2024) reported saving 26 minutes a day, though the report could not tell how the time was spent. The St. Louis Fed tracker put savings at 2.2% of all US work hours in the second quarter of 2026; spread over the 39.2% who used the tools, that is about 5.6% of a user’s hours, a little over two hours of a 40-hour week.
Fully loaded cost per hour
Use the employer’s cost of an hour worked: pay, benefits and employer taxes, plus overhead if finance loads it. The US Bureau of Labor Statistics put employer compensation costs at $54.00 per hour worked for full-time private-industry workers in June 2026, and $78.88 for management, professional and related occupations. Eurostat estimated hourly labour costs of €34.9 in the EU and €38.2 in the euro area in 2025. Use the rate for the roles that will use the system.
Share of saved time that becomes value
An hour saved is worth its full cost only when it moves to valued work or leaves the cost base as less overtime, fewer contractors or a hire not made. The 50% default is a cautious assumption, not a benchmark; raise it only when the case names the backlog, overtime budget or hiring plan that absorbs the hours.
Other annual benefits
Enter steady-state amounts for gains that are not time: the margin on extra or retained sales rather than the revenue, errors and rework avoided, and risk reduction as expected loss avoided, probability times impact. They ramp with adoption in year 1. Count each gain once.
One-off costs
Implementation and integration covers the build, connectors, identity, logging and security review; data preparation covers cleaning, permissions and evaluation sets; change management and training covers communications, champions and the hours people spend learning. The breakdown of AI implementation costs goes line by line, and the comparison of pilot and production budgets shows what changes at scale.
Recurring annual costs
Platform or licences is the subscription or support contract. Model usage and inference grows with volume, so forecast it from expected use rather than the pilot’s bill; if you host models yourself, size the hardware with the GPU memory calculator. The last line is the people who run, evaluate and review the system; leave it out and ROI is overstated. Half the CIOs in Gartner’s 2026 agenda survey (51% of 2,002 respondents) expected AI to raise the total cost of ownership of technology solutions, and Gartner called AI costs “probabilistic and determined at run time”.
Horizon and discount rate
Three years is a common horizon because platforms, models and prices change faster than that; use five only if the system will run largely unchanged. For the discount rate, use the rate finance applies to comparable investments, typically the cost of capital or a hurdle rate above it.
Worked example: a knowledge assistant for 500 people
The defaults describe an illustrative internal assistant that answers staff questions from company documents. This is the arithmetic the calculator runs on them.
- Active users. 500 × 60% = 300 at steady state. Year 1 runs at a 50% ramp, so 150.
- Hours saved in a steady-state year. 300 × 2 hours × 46 weeks = 27,600 hours.
- Value of that time. 27,600 × $60 × 50% = $828,000. Other benefits add $50,000 ($30,000 of rework avoided and $20,000 of risk reduction), for a steady-state annual benefit of $878,000 and $439,000 in year 1.
- Costs. $325,000 one-off plus $300,000 a year, so $1,225,000 over 3 years.
- Net benefit and ROI. $2,195,000 − $1,225,000 = $970,000, and $970,000 ÷ $1,225,000 = 79%.
- Payback. After year 1 the cumulative position is −$186,000. Year 2 adds $578,000, about $48,167 a month, so the gap closes 15.9 months in, during month 16.
- NPV at 8%. −$325,000 + $139,000 ÷ 1.08 + $578,000 ÷ 1.08² + $578,000 ÷ 1.08³ = $758,081.
Which inputs swing the result
Each row changes one input from the default scenario and reruns the same model the calculator uses.
| Change from the default | ROI | Payback | NPV |
|---|---|---|---|
| Default scenario | 79% | 16 months | $758,081 |
| Hours saved: 1 a week instead of 2 | −5% | Not within 3 years | −$117,171 |
| Hours saved: 3 a week | 164% | 12 months | $1,633,332 |
| Adoption: 40% instead of 60% | 23% | 25 months | $174,580 |
| Adoption: 80% | 136% | 13 months | $1,341,582 |
| Share of saved time that becomes value: 25% instead of 50% | −5% | Not within 3 years | −$117,171 |
| Share of saved time that becomes value: 100% | 248% | 8 months | $2,508,584 |
| Year-1 ramp: 25% instead of 50% | 61% | 21 months | $554,840 |
| Recurring costs 25% higher | 51% | 19 months | $564,798 |
| Recurring costs 25% lower | 120% | 15 months | $951,363 |
| One-off costs 25% higher | 68% | 18 months | $676,831 |
| Discount rate: 12% instead of 8% | 79% | 16 months | $671,294 |
| Horizon: 1 year instead of 3 | −30% | Not within 1 year | −$196,296 |
| Horizon: 5 years | 116% | 16 months | $1,576,305 |
Users, adoption, hours saved, cost per hour and the value share multiply each other, so an error in any one moves the time benefit in proportion. Halving the hours saved, or the value share, turns 79% into −5%; the two rows match because each halves the same product. Run costs matter more than the build over several years: a 25% overrun on recurring costs takes ROI to 51%, the same overrun on one-off costs to 68%. The discount rate moves NPV only, and a one-year horizon shows −30% because the build is not yet recovered.
Rerun the case with the low end of every uncertain input. If it works only when every input is optimistic, it is not yet a business case.
Using it as an AI agent, automation or GenAI ROI calculator
The three presets are illustrative assumptions for common project shapes, not benchmarks. Use them to see how a case changes shape, then replace every number.
| Input or result | Internal knowledge assistant | Customer-support agent | Document processing |
|---|---|---|---|
| Target users | 500 | 120 | 40 |
| Steady-state adoption | 60% | 85% | 90% |
| Hours saved per active user per week | 2 | 5 | 10 |
| Cost per hour | $60 | $40 | $45 |
| Share of saved time that becomes value | 50% | 60% | 70% |
| Other benefits a year | $50,000 | $85,000 | $120,000 |
| One-off costs | $325,000 | $300,000 | $275,000 |
| Recurring costs a year | $300,000 | $290,000 | $160,000 |
| ROI | 79% | 44% | 112% |
| Payback | 16 months | 19 months | 15 months |
| NPV at 8% | $758,081 | $382,689 | $669,177 |
- GenAI assistant. Small savings each, across many people, so adoption is the input to test hardest.
- Customer-support agent. Larger savings per person plus retained revenue; it returns less here because model usage is its largest running cost.
- Document processing automation. A small team, large savings per person and low running costs, hence the highest ROI of the three.
For an agent that completes work on its own, count the work it absorbs (tickets resolved, documents processed) as the hours people would otherwise spend, and keep the time spent checking its output in the support, operations and governance line where reviewers can see it. One agent run can involve many model calls, so forecast its usage from expected volume. The guide to automation business cases covers process-level inputs, and the value, risk and effort method tells you which candidates are worth modelling.
What an ROI figure leaves out
- Downside quality. Wrong answers, AI-caused rework and incidents are costs; measure them in the pilot and price them in.
- Options. A platform that serves later use cases changes their economics, which one project’s ROI cannot show.
- Overlap. Replacing a tool you pay for is a benefit; duplicating one is a cost.
- Drift. Model prices, deprecations and volume growth move recurring costs, which the calculator holds flat.
- Distribution. An average can hide one team that gains a lot and three that gain nothing.
Real projects tend to take longer than models suggest. In Deloitte’s survey of 1,854 executives in Europe and the Middle East, published in October 2025, most said a typical AI use case reached satisfactory ROI within two to four years, against seven to twelve months expected of technology investments; only 6% reported payback inside a year.
After launch, measured results replace the estimate: the guide to measuring AI ROI once a system is live covers baselines and control groups, and the AI KPI guide lists the leading indicators to watch meanwhile.
AI ROI calculator questions
How do you calculate the ROI of an AI project?
Add up the benefits over a fixed horizon, subtract everything the project costs over the same horizon and divide by the total cost. Benefits are usually adoption-weighted time savings, meaning active users × hours saved × cost per hour × the share of that time that becomes value, plus gains such as retained revenue or avoided rework. Costs are one-off build costs plus recurring licences, model usage, infrastructure and operations.
What is a good payback period for an AI investment?
Compare it with the rule your finance team applies to other investments; there is no universal threshold. Deloitte’s 2025 survey of 1,854 executives in Europe and the Middle East put the usual time to a satisfactory return on an AI use case at two to four years, against seven to twelve months expected for technology in general, and only 6% saw payback within a year.
How do you calculate ROI for an AI agent?
Use the same structure, but count the work the agent absorbs, such as tickets resolved or documents processed, converted to hours and multiplied by the cost of the people who would otherwise do it. Keep the time people spend reviewing its output in the support, operations and governance line, and forecast model usage from expected volume, because one agent run can involve many model calls.
What costs should an AI ROI calculation include?
One-off costs for implementation and integration, data preparation, and change management and training, plus recurring costs for the platform or licences, model usage, infrastructure, and the people who support, evaluate and govern the system. Leaving out those people is an easy way to overstate ROI. Gartner’s 2026 survey of 2,002 CIOs found that 51% expected AI to increase the total cost of ownership of technology solutions.
How many hours a week does AI save an employee?
It depends on the tool and the role, so measure it in a pilot. Published figures are self-reported: UK civil servants in a 2024 Microsoft 365 Copilot experiment reported saving 26 minutes a day, and the St. Louis Fed’s tracker put savings at 2.2% of all US work hours in the second quarter of 2026, about two hours a week per user.
Why calculate NPV as well as ROI for AI?
ROI treats a dollar of benefit in year three like a dollar today. NPV discounts each year’s net cash flow at your discount rate and subtracts the up-front cost, so late benefits count for less. A positive NPV at your finance team’s rate means the project earns more than the cost of the capital it ties up.
Related guides and templates
Costing a governed AI platform
VDF AI is one platform that would sit in the recurring lines above: governed AI agents, Networks to orchestrate several agents, retrieval over your own documents in VDF AI Chat, role-based access control and audit logs on every plan, and VDF AI Router, which routes between approved cloud models and registered on-premises model endpoints. Deployment options are VDF’s managed cloud, your own data centre, a private cloud or an air-gapped network.
This calculator holds no VDF AI prices; cloud plans are on the pricing page. Audit logs and usage records can supply the adoption figures an ROI review needs, while business outcomes still come from your own systems.
Sources
Each statistic and reference rate quoted above was re-read at its publisher on 4 November 2026.
- BLS news release USDL-26-1494: employer compensation costs per hour worked, June 2026
- BLS ECEC table 4: private industry workers by occupational group, June 2026
- Eurostat: hourly labour costs in the EU in 2025 (published 31 March 2026)
- FRED Blog, St. Louis Fed: generative AI time savings at work (27 August 2026)
- GOV.UK: cross-government findings from the Microsoft 365 Copilot experiment (2 June 2025)
- Gartner: 2027 CIO and Technology Executive Agenda survey results (19 October 2026)
- Deloitte: AI ROI, the paradox of rising investment and elusive returns (22 October 2025)