How to Measure the ROI of AI Projects (With a Worked Formula)

Your team has been using an AI tool for 3 months. Everyone says it “saves time.” Your CFO asks how much, and what it is worth. If you did not measure anything before the tool arrived, that question is very hard to answer.

This guide gives you a simple AI ROI formula, a worked example with real numbers, and the baseline metrics to capture before any pilot starts, so you can show what AI is actually delivering.

Why AI ROI is hard to measure

AI benefits are real but slippery. Time savings are spread across many people in small chunks. Quality gains, like fewer errors or happier customers, take time to show up. And costs hide in places like staff review time and integration work.

There is evidence that the time savings are real. In the Census Bureau’s March 2026 Household Trends and Outlook Pulse Survey, about 56% of US workers said they used AI on the job for at least 1 of 11 tasks. Among those who used it in the prior week, 31% said it saved 1 to 2 hours, while 10% said it saved no time and 4% said it took extra time. That spread is exactly why you need to measure your own results instead of assuming.

The AI ROI formula

The standard ROI formula works for AI projects:

ROI (%) = (Total benefits − Total costs) ÷ Total costs × 100

Measure it over a set period, usually the first 12 months. Two related measures help too:

• Net benefit = Total benefits − Total costs
• Payback period = Upfront costs ÷ Net monthly benefit

The hard part is not the math. It is getting honest numbers for benefits and costs.

What counts as benefits

Most AI benefits fall into 4 groups:

1. Time saved. Hours saved per task × number of tasks × loaded hourly labor cost. “Loaded” means wages plus benefits. The Bureau of Labor Statistics Employer Costs for Employee Compensation report put average private industry compensation at $46.89 per hour worked in June 2026, with wages at 70% and benefits at 30%. Use your own payroll figures where you can.
2. Revenue gains. More leads, higher conversion, larger orders or faster sales cycles that you can tie to the AI use case.
3. Cost avoided. Fewer contractor hours, lower overtime, reduced software spend or avoided hires as volume grows.
4. Quality and risk. Fewer errors, rework, refunds or compliance issues. Put a dollar value on these only when you can measure them, such as the average cost of a returned order.

Apply a realization factor. Saved time only creates value if it is reused for productive work. If a support agent saves 30 minutes but spends it waiting for the next ticket, the company gains little. Many teams apply a conservative factor, such as 50%, to time savings unless they can show the time went to revenue-producing or backlog work.

What counts as costs

AI costs go well beyond the subscription price:

Count staff time at the same loaded hourly rate you use for benefits so the comparison is fair.

A worked example

Here is an illustrative example for a 10-person customer support team using an AI assistant to draft replies. The license price and time savings are assumptions for the example, not market figures; prices vary by vendor and plan.

Baseline and assumptions

• 10 agents, 48 working weeks a year
• Loaded labor cost: $46.89 per hour (the BLS private industry average)
• Measured time saved in the pilot: 2 hours per agent per week
• Realization factor: 50%
• License: assumed $30 per user per month
• Setup and training: 40 staff hours
• Ongoing review and admin: 2 hours per week in total

Benefits (year 1)

• Hours saved: 10 agents × 2 hours × 48 weeks = 960 hours
• Value of hours: 960 × $46.89 = $45,014
• After 50% realization: $22,507

Costs (year 1)

• Licenses: 10 × $30 × 12 = $3,600
• Setup and training: 40 × $46.89 = $1,876
• Review and admin: 2 × 48 × $46.89 = $4,501
• Total: $9,977

Results

• Net benefit: $22,507 − $9,977 = $12,530
• ROI: $12,530 ÷ $9,977 × 100 = about 126%
• Payback: monthly benefit of $1,876 minus monthly running costs of $675 gives a net $1,200 a month, so the $1,876 setup cost pays back in about 1.6 months

Notice how much the review time and realization factor matter. Without them, the ROI would look several times higher, and a CFO would rightly question it.

Key takeaway: A credible AI ROI number depends on 3 things: a baseline measured before the pilot, every cost including staff review time, and an honest realization factor for time saved.

Baseline metrics to capture before a pilot starts

You cannot measure improvement without a starting point. Before the pilot, record 2 to 4 weeks of data on the metrics tied to your goal:

Efficiency - Average time per task (for example, minutes per support reply or per proposal) - Volume handled per person per week - Backlog size and turnaround time

Quality - Error or rework rate - Customer satisfaction or quality review scores - Escalation or refund rate

Business outcomes - Conversion rate, revenue per customer or deal cycle length - Overtime and contractor spend

Adoption - Number of active users and how often they use the tool (you will track this during the pilot)

Use the same measurement method before and after. If possible, keep a comparison group that does not use the tool during the pilot, which helps separate the AI effect from seasonal changes.

How to run an AI pilot you can measure: 6 steps

1. Pick one use case and one goal. For example, “reduce average reply time for support tickets.” See our guide on how to create an AI strategy for ranking use cases.
2. Define success up front. Write down the metric, the target and the decision rule, such as “scale if time per reply drops 20% with no fall in quality scores.”
3. Capture the baseline. 2 to 4 weeks of data, measured the same way you will measure the pilot.
4. Run the pilot for 30 to 90 days. Give users training and a champion to ask. Expect a learning curve in the first weeks.
5. Track costs as you go. Log setup hours, review time and any integration work, not just the invoice.
6. Compare, calculate and decide. Apply the formula, check quality metrics and choose to scale, adjust or stop.

Before you start, make sure the tool meets your data rules. Our guides on whether it is safe to put company data into AI and writing an AI acceptable use policy cover the basics.

Common mistakes that inflate AI ROI

• No baseline. Estimates made after the fact are usually optimistic.
• Counting all saved time as money. Use a realization factor unless the time clearly went to valuable work.
• Ignoring review time. Checking AI output is a real cost, especially for customer-facing content.
• Measuring only the enthusiasts. Early adopters often save more than the average user. Measure the whole pilot group.
• Forgetting quality. Faster but less accurate work can cost more in rework and lost customers.
• Short time frames. Some benefits, like faster onboarding or lower churn, take months to appear.

The voluntary NIST AI Risk Management Framework also stresses ongoing measurement of AI system performance, which supports tracking quality alongside savings. For more practical first use cases, see our generative AI for small business guide.

Conclusion

Measuring AI ROI is straightforward once you do the groundwork. Record a baseline before the pilot, count every cost, value time savings at a loaded labor rate with an honest realization factor, and track quality next to speed. Then apply the formula and decide based on evidence. That discipline turns “AI saves us time” into a number your leadership can trust.

Sources

• US Bureau of Labor Statistics, Employer Costs for Employee Compensation (June 2026) — average private industry compensation of $46.89 per hour; wages and benefits shares
• US Census Bureau, Does Using Artificial Intelligence Save Time at Work? — worker AI use and reported time savings from the March 2026 HTOPS
• NIST AI Risk Management Framework — measuring and monitoring AI system performance