How do you measure the return on AI in a small business?

An owner told me their AI tools had "definitely paid for themselves." I asked by how much. After a long pause, the answer was that everyone felt faster. Feeling faster is real, and it is also how a business ends up with a growing software bill and the same payroll, the same hours and the same owner working weekends.

What should you measure?

Four numbers for each job, before and after: the time each instance takes, how many instances happen a month, the loaded hourly cost of whoever does it, and how often the work has to be redone. Then add the monthly running cost of the AI and the hours a person spends checking it. Anything vaguer becomes an opinion.

In that example the job costs $1,200 a month less, before rework. If the rework rate rises, the saving shrinks, and if the thirty freed hours drift into nothing in particular, the saving never reaches the bank. Both are measurable, which is why they belong on the same sheet.

What does an hour of staff time actually cost?

More than the wage. The Bureau of Labor Statistics reported that US private employers paid an average of $46.89 per hour worked in total compensation in June 2026. Wages and salaries were $32.82 of that, or 70%, and benefits were $14.07, or 30%. Use your own payroll numbers, and treat the average only as a check.

Most owners underestimate this by using the hourly wage alone, which makes every saving look about 30% smaller than it is. They also tend to forget the other side: the hours a manager spends reviewing agent output are paid hours too, usually at a higher rate than the work being reviewed.

Why don’t saved hours show up in the bank account?

Because an hour saved is money only when it goes somewhere. Economists Humlum and Vestergaard studied Denmark and found precise null effects on earnings and recorded hours, ruling out effects larger than 2% two years after ChatGPT launched. Jobs changed, with new tasks such as overseeing AI, while pay and hours stayed where they were.

Decide where the hours go before the agent starts. Name the work that has been waiting for time, such as sales calls, a delayed project or cutting a contractor’s monthly bill, and check at the end of the month whether it happened. If nothing on that list moved, the return was on paper only.

What returns does the research support?

Real ones, in narrow settings. In an NBER study of 5,179 customer support agents, access to an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with little effect on the most experienced. That was one company, one kind of work and a tool built for it.

Use that finding to choose where to look, and your own baseline to decide. It suggests the largest gains come where less experienced people do work that experienced people already know how to do well. It doesn’t tell you what your business will see, and a vendor quoting a percentage without your baseline is guessing.

How do you run the measurement?

1. Pick one job that happens at least weekly. Measuring a whole department at once hides everything.
2. Record the four baseline numbers for two weeks, using real timestamps where you have them.
3. Write down what counts as an error or rework before you start, so the definition doesn’t move later.
4. Run the agent for four weeks with a named reviewer, and log review time and corrections.
5. Recalculate the same four numbers, add running and review costs, and check whether the freed hours went where you said they would.
6. Decide: keep it, fix the instructions and rerun, or stop. Write the decision down with the numbers.

Which mistakes make the numbers lie?

• Measuring the baseline from memory. People remember how long a task felt, and feelings run long.
• Leaving out review time because the reviewer is a manager whose hours nobody tracks.
• Changing the job while measuring it, such as cutting a section from the report, and then crediting the AI for the time.
• Counting setup as free. The hours spent writing instructions and fixing early mistakes are part of the first quarter’s cost.
• Averaging several jobs together, so one clear win hides one quiet loss.

None of these take bad intent. They happen because the person measuring also wants the project to work, which is why the numbers should sit in one shared sheet that someone other than the project’s champion can read and question.

What returns are hard to count?

Faster replies to new leads, more consistent client updates, and your own hours back. Count them, and count them separately from the cash figure, with a note of how you measured each one. Founder hours in particular are easy to overvalue in a spreadsheet and easy to lose in practice. If an agent frees six hours of your week and those hours go into more email, the business didn’t gain anything. Put the freed hours on the calendar, by name, before the agent goes live.