Skip to content
Createfirm

How to measure the ROI of an AI project, step by step

A simple method to measure the ROI of an AI project: set a baseline, total the build, run and oversight costs, value the benefits and work out payback.

By the Createfirm teamPublished 6 min read

Illustration of measuring AI return on investment: rising steel bars, a stack of gold coins and a calculator

To measure the ROI of an AI project, record how the process performs today, add up what the project costs to build, run and oversee, value the time, revenue and error savings it produces, then compare the two. ROI shows the return as a percentage. Payback period shows how long until the project has paid for itself.

What do ROI and payback period mean?

Both are standard finance measures, and both are simple.

Return on investment (ROI). Harvard Business School Online’s guide to project ROI describes ROI as a metric that shows how much profit an investment has generated. For a project, it gives this formula:

ROI = [(financial value - project cost) / project cost] x 100

The same guide separates anticipated ROI, estimated before a project starts, from actual ROI, calculated after it ends. You want both. The first helps you decide. The second tells you whether you were right.

Payback period. OpenStax’s managerial accounting textbook explains that the payback method “computes the length of time it takes a company to recover their initial investment.” When cash flows are even, the formula is:

Payback period = initial investment / net annual cash flow

You can use months instead of years as long as both numbers use the same period.

How do you measure the ROI of an AI project?

Work through four steps in order.

Step 1: Record a baseline

Pick one process and measure it for two to four weeks before you change anything. Useful baseline numbers include:

  • how many times the task happens per week
  • how many minutes each one takes, and who does it
  • how often it goes wrong and what a mistake costs to fix
  • how long customers wait
  • how many leads, bookings or payments the process produces

Write the numbers down with the dates. This is your “before.” Without it, any later claim about results is a guess.

If you are still choosing a process, our AI readiness assessment questions can help you narrow the list.

Step 2: Add up the costs

AI project costs come in three groups.

  1. Build. The one-time cost of design, setup, integration with your systems, data cleanup, testing and staff training.
  2. Run. The ongoing cost of software and usage. Many platforms bill by usage. For example, Zapier measures usage in tasks, where a task is a successful action step. Make bills in credits, and for some AI features the credits used depend on how much text the model processes. Your run cost grows with volume.
  3. Oversight. The staff time spent reviewing outputs, handling the cases the AI passes to a person and keeping the system up to date. This is the cost most often left out.

Step 3: Value the benefits

Benefits also come in three groups. Use your own numbers for each.

  • Time. Hours saved per month multiplied by the loaded hourly cost of the people who did the work. Count this only if the time goes to something useful.
  • Revenue. Extra sales that result from the change, such as more leads answered or more appointments kept. Use gross margin, not sales price.
  • Error reduction. Fewer mistakes multiplied by the average cost of fixing one, including refunds and rework.

Be conservative. If you are unsure about a benefit, leave it out or use the low end of your range.

Step 4: Calculate payback and ROI

Subtract monthly run and oversight costs from monthly benefits to get the net monthly benefit. Divide the build cost by that figure for payback in months. Then apply the ROI formula over a period, such as the first year.

What does a worked example look like?

All numbers below are hypothetical. They are for illustration only and are not results from any real project.

Imagine an automation that follows up on overdue invoices and answers routine billing questions.

Item Hypothetical amount
Build cost (one time) $12,000
Run cost (software and usage) $300 a month
Oversight (4 hours a month at $40 an hour) $160 a month
Total monthly cost $460
Time saved (30 hours a month at $40 an hour) $1,200 a month
Added gross margin $600 a month
Fewer errors $200 a month
Total monthly benefit $2,000

Net monthly benefit: $2,000 - $460 = $1,540

Payback period: $12,000 / $1,540 = about 7.8 months

First-year ROI:

  • Financial value: $2,000 x 12 = $24,000
  • Project cost: $12,000 + ($460 x 12) = $17,520
  • ROI: ($24,000 - $17,520) / $17,520 x 100 = about 37%

In the second year there is no build cost, so the same monthly figures produce a much higher return. That is why the period you choose matters. State it every time you quote an ROI.

What are the limits of these measures?

Payback period is easy to explain, and it has known gaps. The OpenStax text notes that it does not consider the time value of money. Corporate Finance Institute’s explanation of payback period adds that cash flows after the payback point are not part of the decision rule, so a project with large later returns can look the same as one with none.

For a small project with a short payback, that is usually acceptable. For a large one, ask your accountant whether a discounted method is more suitable. This article is general information, not financial advice.

What does research say about AI project results?

Results are real, and they are uneven. Generative AI at work, a National Bureau of Economic Research working paper by Erik Brynjolfsson, Danielle Li and Lindsey Raymond (issued April 2023, revised November 2023), studied an AI assistant used by 5,179 customer support agents. Productivity, measured as issues resolved per hour, rose 14% on average. Novice and low-skilled workers improved by 34%, while experienced and highly skilled workers saw minimal impact.

Two lessons follow for your own measurement.

  • Do not copy a published percentage into your forecast. That study covered one tool in one kind of work. Use your baseline.
  • Measure by group, not only on average. The gain may sit with newer staff, one shift or one type of request.

What mistakes make ROI numbers unreliable?

Use this checklist before you share a result.

  1. The baseline was measured before the change, over a normal period.
  2. Oversight time is counted as a cost.
  3. Time savings are counted only where the time was put to use.
  4. Revenue gains use margin, not sales price.
  5. The comparison period is long enough to smooth out an unusual week.
  6. Quality is tracked alongside speed, so faster work has not created more rework.

Point 6 connects to risk. Our guide to AI risk management for small business covers the checks worth keeping. For project ideas to measure, see these AI automation examples for small business.

The short version

Measure the process first, then count build, run and oversight costs against time, revenue and error savings. Payback period tells you how fast the project returns its cost, and ROI tells you how much it returned over a stated period. Use your own baseline, stay conservative and check the result again after the project has run for a few months.

Frequently asked questions

How do you calculate the ROI of an AI project?
Add up the financial value the project produced over a period, subtract the total project cost for the same period, divide by that cost and multiply by 100. Harvard Business School Online gives the project formula as financial value minus project cost, divided by project cost, times 100. Include build, run and oversight costs.
What is a good payback period for an AI project?
There is no single standard. Payback period is the time it takes to recover the initial investment, and the acceptable length depends on your cash position, the risk of the project and how long the solution will stay useful. Many owners compare it with other uses of the same money and set their own threshold.
What costs should you include in an AI ROI calculation?
Include three groups. Build costs cover design, setup, integration, data cleanup and training. Run costs cover software subscriptions and usage-based fees, such as tasks or credits on an automation platform. Oversight costs cover the staff time spent reviewing outputs, handling exceptions and maintaining the system. Leaving out oversight is the most common gap.
Why is the ROI of AI hard to measure?
Benefits are often spread across many small tasks, and they vary by person. An NBER study of customer support agents found a 14% average productivity gain from an AI assistant, with a 34% gain for novice workers and minimal impact on experienced ones. A baseline and a narrow scope make the result measurable.

Sources

Every fact in this article was checked against these pages on November 14, 2026.

  1. How to Calculate ROI to Justify a Project, Harvard Business School Online
  2. 11.2 Evaluate the Payback and Accounting Rate of Return in Capital Investment Decisions, OpenStax, Principles of Accounting, Volume 2: Managerial Accounting
  3. Payback Period, Corporate Finance Institute
  4. Generative AI at Work, National Bureau of Economic Research
  5. How is task usage measured in Zapier?, Zapier
  6. Credits, Make Help Center

Keep reading