How to Calculate AI ROI (Formula + Worked Example)

AI ROI is the net benefit an AI project produces divided by its total cost: ROI = (Benefits − Costs) ÷ Costs. To calculate it honestly, put a dollar value on the hours, errors and revenue the project actually changes, subtract every cost (including licenses, model usage, maintenance and your team's time), and judge the result on payback and a 3-year view, not on year one alone.

Where AI ROI estimates go wrong is predictable. They count saved hours that never turn into useful work, assume everyone adopts the tool on day one, and leave out usage and maintenance. This guide gives you the formula, a fully worked example with every assumption visible, a sensitivity analysis, and a worksheet you can copy into a spreadsheet.

The AI ROI formula

AI ROI uses the same formula as any other investment, applied over a period long enough to include both the build and the payoff:

ROI = (Benefits − Costs) ÷ Costs

If a project costs $40,000 over three years and produces $100,000 in benefits, ROI = ($100,000 − $40,000) ÷ $40,000 = 150%. Two companion numbers make the result useful:

• Payback period is how long it takes cumulative benefits to cover cumulative costs. A quick version is one-time costs ÷ monthly net benefit, where monthly net benefit is monthly benefits minus monthly running costs. Then add the months it takes to build and ramp up.
• 3-year ROI applies the formula to three years of benefits and costs. Year one carries the build and the learning curve; years two and three show the steady state. Three years is long enough to be fair to the investment and short enough that the assumptions still mean something.

For large projects, finance teams discount future years to net present value (NPV). For most small-business automation, a simple 3-year view is enough to make the decision.

What counts as an AI benefit

AI benefits fall into four categories, and each needs its own formula so you can check it after launch.

Hours recovered: use a loaded rate and a capture rate

An hour saved is worth the full cost of that hour, not just the wage. At private-industry establishments with fewer than 50 workers, employers paid $27.88 an hour in wages and $9.76 in benefits per hour worked in June 2026, according to the Bureau of Labor Statistics' Employer Costs for Employee Compensation. That's about 35 cents of benefits on top of every wage dollar. Across all of private industry, benefits were 30% of total compensation.

A quick loaded rate for a small business is wage × 1.35; use your own payroll numbers if you have them. Then apply a capture rate: the share of saved hours that turns into work that matters, like calling accounts, collecting receivables, or absorbing growth without a new hire. Ten minutes saved here and there rarely becomes anything. As a rule of thumb, use 70–90% if you can name where the hours will go, and 30–50% if you can't.

Revenue lift: count margin, not revenue

If an AI agent answers leads in minutes instead of hours and lifts your close rate, the benefit is the added gross profit, not the added revenue. Example: 400 leads a month × a 1-point increase in close rate × a $2,500 average sale × 35% gross margin = $3,500 a month. Revenue lift is the hardest benefit to prove, so give it your most conservative assumption and, if you can, compare against a control group.

Error reduction and cost avoidance

Price each error the way your finance team would: re-shipping, restocking, credits, staff time and, for some errors, the customer you lose. Cost avoidance counts money you would otherwise spend, such as a hire you no longer need as volume grows, overtime, late-payment fees or software you can retire. Count an avoided hire only if you really would have made it.

What counts as an AI cost

Count every cost that wouldn't exist without the project, one-time and recurring.

Human review time is also a cost. The cleanest way to handle it is to subtract it from the hours saved, as the example below does. For typical prices on each line, see how much AI consulting costs.

Worked example: AI ROI for order-entry and invoice automation

This example is illustrative, not a client result. Every assumption is shown so you can swap in your own numbers.

The business: a 15-person distributor. Customers email about 40 purchase orders a day as PDFs and spreadsheets, and two office staff key them into the ERP, along with 500 supplier invoices a month.

The project: AI reads each order and invoice, matches customers, SKUs and purchase orders, and drafts the ERP entry for a person to approve. Anything that doesn't match goes to an exception queue.

At full use, the annual benefit is about $38,980:

• Labor value: 1,405 hours × $31 × 60% = about $26,130
• Error savings: 840 orders × 1.5 percentage points × 12 months × $85 = about $12,850

Running costs are about $7,540 a year at full use. Here's the 3-year view, rounded to the nearest $10:

The results:

• 3-year ROI = ($107,200 − $57,670) ÷ $57,670 = 86%
• Year-one ROI = −31%
• Payback: month 18 from kickoff, about 16 months after go-live

That's a solid, unspectacular and realistic result. A naive version of the same calculation, with every saved hour at full value, no ramp-up and only the build cost counted, claims 88% ROI in year one and payback in about six months. It counts benefits the business won't receive and leaves out costs it will pay.

Sensitivity analysis: pessimistic, expected and optimistic

A single ROI number hides how fragile it is, so run three scenarios and see which assumptions carry the result. Figures are rounded.

Change one assumption at a time, starting from the expected case's 86%, and the ranking becomes clear:

• Capture rate of 40% instead of 60%: 3-year ROI falls to 44%.
• Errors fall only to 1.25%: 55%.
• Build runs 25% over budget: 57%.
• Exception rates double: 64%.
• Adoption takes twice as long: 80%.
• Model usage costs double: 80%.

In this project, what the business does with the freed-up time matters more than anything the AI does. That's why the capture rate deserves an owner and a plan, not a guess. Usage barely moves this result because the volume is modest. For a customer-facing agent handling thousands of conversations a month, run the same test, because usage can become one of the largest lines.

How to set a baseline before you start

A baseline is the measured "before" number for each metric the project should move, captured with the same method you'll use afterward.

1. Pick 3–5 metrics tied to the workflow: volume, minutes per transaction, error rate, cycle time (for example, email received to order entered) and cost per transaction.
2. Measure for 2–4 normal weeks. Avoid holiday peaks and month-end spikes unless they're the point.
3. Use the cheapest reliable method. Time 20–30 transactions with a stopwatch, pull timestamps from email and the ERP, and count credit memos or rework tickets by reason code.
4. Agree on the dollar values, such as the loaded rate, cost per error and gross margin, with whoever owns the P&L.
5. Write it down: the metric, the number, the method, the dates and who measured it.

If you can't measure a metric before the project, you won't be able to prove it moved afterward. The baseline is also the starting point of any forecast, which is why financial analysis and forecasting and ROI work go together. It's also why every recommendation in a NextGen Code AI Opportunity Assessment comes with a measured baseline, a target metric and a payback estimate.

How to measure AI ROI after launch

Measure the same metrics, with the same method, at 30, 60 and 90 days after launch, then quarterly. Track four groups:

• Adoption: the share of transactions flowing through the new process, and how often people work around it.
• Quality: error rate, exception rate and how often reviewers edit the AI's drafts. Rising edits are an early warning.
• Cost: actual usage bills, platform fees and maintenance hours against the plan.
• Realization: where the saved hours went. Ask the manager to name the work, the overtime cut or the hire not made.

Put these on one dashboard so nobody has to assemble a report by hand. It's standard business intelligence work, and it keeps the project honest; if you don't have dashboards yet, start with our guide to business intelligence for small business. At 90 days, recalculate ROI with actual numbers and make a call: scale it, fix it or stop it.

Common AI ROI mistakes

1. Counting hours that never get redeployed. Saved time only has value if it goes somewhere. Apply a capture rate and name the destination.
2. Ignoring adoption. Nobody uses a new tool fully on day one. Model a ramp-up and measure actual use.
3. Ignoring usage costs. Pay-per-task fees grow with volume and with how much text each task sends. Estimate cost per task during testing.
4. Forgetting review time. A person approving AI drafts is part of the new process. Subtract that time from the savings.
5. Counting revenue instead of margin. A $10,000 sale isn't a $10,000 benefit.
6. Leaving out maintenance and internal time. Both are real costs, and both are easy to forget.
7. Judging on year one alone, or on five rosy years. Build costs make year one look bad, while a 5-year view with no maintenance costs flatters almost anything.

AI ROI worksheet

Copy this into Excel or Google Sheets, one worksheet per workflow. Each formula refers to the line letters, so it carries over directly. Enter rates as decimals (2% = 0.02).

Line T counts a full year of running costs in year one, which is slightly conservative. Run the worksheet three times, with pessimistic, expected and optimistic inputs, before you trust any single answer.