How to measure AI ROI: a simple formula for AI agents
By the Codivated team · Updated
Measure AI ROI the way you'd measure any hire: the value of the work it does (hours back, errors avoided, revenue gained), minus everything it costs (build, model usage, care and your team's review time), compared against a baseline you measured before it went live. For one workflow, the math fits on one page. Our free AI ROI calculator gives you a first estimate in about a minute.
What is the formula?
ROI = (value gained − total cost) ÷ total cost, over a set period, usually 12 months. The payback period is the total first-year cost divided by the monthly value. The hard part isn't the formula. It's measuring the value honestly.
What counts as value?
| Value | How to measure it | Watch out for |
|---|---|---|
| Hours back | Cases handled × minutes saved per case × loaded hourly cost | Hours saved only count if the time goes to other work |
| Faster response | Time to first response or to order entry, before and after | Tie it to an outcome, like win rate or fewer "where is my order" calls |
| Fewer errors | Credits, rework, wrong shipments per month | Measure the baseline; most teams underestimate their error rate |
| Revenue | Leads followed up, upsells, recovered carts | Compare against a holdout group if you can |
| Avoided cost | A software contract or a hire you no longer need | Count only what you actually cancel |
What counts as cost?
- Build. The project to design, connect and test the agent.
- Model usage. What the AI models charge per request. It grows with volume and with how much the agent reads per case.
- Care. Monitoring, tuning and updates when your process or software changes.
- Your team's time. Reviewing and approving the agent's work, especially in the first months.
- Tools. Any platform or hosting fees.
A worked example
An operations team enters 1,500 emailed purchase orders a month at about 8 minutes each, with a loaded cost of $40 an hour. That is 200 hours a month, about $8,000. An order-entry agent handles 70% of orders with a 1-minute review, and people still handle the rest fully.
| Before | After | |
|---|---|---|
| Orders a month | 1,500 | 1,500 |
| Minutes of people's time | 12,000 | 3,600 + 1,050 review = 4,650 |
| Hours a month | 200 | about 78 |
| Monthly labor value | $8,000 | about $3,100 |
That's about $4,900 a month in time back before counting faster orders or fewer errors. Subtract monthly model usage and care, divide the build cost by what's left, and you have the payback period. These numbers are illustrative; use your own volumes.
How do you get a trustworthy baseline?
Measure two to four weeks of the current process before the agent goes live: volume, time per case (sample it, don't guess), error rate and response time. Write the target down before the build starts. Without a baseline, the result will be argued about forever.
When should you measure?
Three times: during practice mode (what it would have done), at 30 days live, and at 90 days. Expect value to grow as you widen what the agent does on its own, and review time to fall.
Why do AI ROI estimates go wrong?
- Counting hours that never turn into other work or lower cost.
- Using vendor benchmarks instead of your own numbers.
- Forgetting review time and ongoing care.
- Measuring a pilot on hand-picked examples. See why AI pilots fail.
Common questions
What's a good ROI for an AI agent?
A first agent on a high-volume workflow should usually pay back within its first year. If the math needs three years, pick a different first job.
Should model usage be included in ROI?
Yes. It's usually smaller than labor savings, but it grows with volume, so model it at your real volume.
How do we measure ROI on a knowledge assistant?
Track questions answered, time saved for the experts who used to answer them, and faster ramp-up for new hires. See AI knowledge bases for employees.
Sources
- The state of AI, McKinsey: how companies report value from AI.
- MIT report: 95% of generative AI pilots at companies are failing, Fortune, August 2025.