Measuring AI ROI once the tool is running comes down to one task, before and after: the hours it takes, how long it takes to finish, and how often something has to be fixed. Then subtract what the tool costs you each month, including the time spent checking its work. Do the math at 30 days and again at 90, because the first weeks include learning time and the real number shows up later.
TL;DR
- Measure one task, not "AI" in general.
- Count checking and fixing time as a cost.
- Judge at 90 days, not in week two.
What should you measure?
Four things, on one task:
- Hours. How long the task takes a person each week.
- Turnaround. How long from start to done, such as request received to quote sent.
- Errors and rework. How often the output has to be corrected or redone.
- Checking time. How long someone spends reviewing what the AI produced.
Pick the task the tool was bought for and measure it the same way each time, with the same person doing it, in a normal week rather than your busiest or slowest.
Checking time is the one people forget. A tool that saves four hours of writing and adds three hours of checking saved one hour.
What if you never measured the before?
You can rebuild it. Ask the person who did the task how long it took, then check it against what's on record: email timestamps, the dates quotes went out, how many invoices were sent late. If the old way still exists, run it by hand for one week and time it. A rough baseline beats none, as long as you write it down now and use the same one at 30 and 90 days.
What does the tool really cost each month?
- The subscription or plan.
- Usage charges, if you pay per run.
- Checking time, from the list above.
- Upkeep: fixing it when an app or your process changes.
Subscriptions can hide the real number. Rick Kranz at AI Marketing Labs added a live dashboard to one of his own agents because he wanted to know "what is this thing actually costing?" and found the cost per job still unclear, because the agent ran on a subscription rather than paying per use.
How do you turn hours into dollars?
Use what an hour actually costs you. If you don't know your number, the U.S. Bureau of Labor Statistics reports that "Total employer compensation costs for private industry workers averaged $46.89 per hour worked in June 2026" (BLS). That includes benefits, so it's closer to your real cost than a wage.
A worked example, with our assumptions stated:
| Line | Per month |
|---|
| 3 hours a week saved (13 hours a month at $46.89) | $609 |
| Subscription | minus $40 |
| 1 hour a week checking output (4.33 hours) | minus $203 |
| 1 hour a month of upkeep | minus $47 |
| Net | $319 |
Swap in your own numbers. Faster turnaround and fewer errors can matter more than the hours, like a quote that goes out the same day, but put a dollar figure on them only if you can trace it, such as jobs won that you'd have lost.
Why check at 30 and 90 days?
Because the first weeks include setup, learning and fixes, and they make the tool look worse than it is. Check at 30 days to catch problems early, and judge at 90, when people have settled into it.
Expect the gain to differ from person to person. In a study of 5,179 customer support agents, an AI assistant "increases productivity, as measured by issues resolved per hour, by 14% on average," "including a 34% improvement for novice and low-skilled workers" (NBER). Measure each person on the task, not just the team average.
What if the numbers don't add up?
You're not alone. In PwC's survey of 4,454 CEOs, "More than half (56%) say their company has seen neither higher revenues nor lower costs from AI" (PwC, 2026). Those are large companies. At a small business, check two things first: was the tool pointed at too much at once, and did anyone measure the task it was meant to fix?
Look at where the time goes. If checking eats the savings, narrow the job or fix the instructions. If the task changed, re-measure the new version. If neither helps by day 90, cancel it and move the money to a task that passes the test.
Sometimes the tool is the problem: it can't do the job the way your process runs. That's when AI Marketing Labs builds Custom AI Automations around the one process, scoped and priced before any building starts, and owned by you outright.
What should you do with the result?
Keep it, fix it or drop it, and write down why. Use the same method on the next tool before you buy it; here's how to estimate AI ROI before you automate anything, and how to tell whether an AI tool is worth what it costs. If you're starting from scratch, automate one weekly task first so you have something clean to measure.
Related articles
Frequently asked questions
What's a good ROI for an AI tool?
One that clears its full monthly cost, including checking time, with room to spare by day 90. If it only breaks even, the risk and upkeep may not be worth it.
Should I count time saved if nobody's workload actually dropped?
Only if the time went somewhere useful. If the saved hours turned into more work done or faster turnaround, count that. If they disappeared, the tool didn't save anything you can bank.
How often should I recheck ROI after 90 days?
Every quarter, or whenever the price, the task or the people doing it change.
.jpg)
