How do we measure ROI on an AI project?
Pick one number the project should move, such as hours spent on a task, errors, or time to reply, and record it before the pilot starts. Then compare that same number after the pilot against what the system cost to build and costs to run each month, and you have your payback period. Without that before number there is no ROI to measure, only opinions.
Evidence
I measure my own AI systems the way I ask clients to measure theirs. WASBuilder, the system that has AI agents write, test and review software, recorded its outcomes from the first ticket, so its results are counted rather than remembered. This site's citation tracker records its own AI bill on every run, which is the cost side of the same sum.
- 665 of 681 WASBuilder tickets created were shipped to production
- 17.4% of WASBuilder tickets needed any rework at all
- $0.54 to $1.10 per weekly run of the citation tracker, read from production
Read how those figures are counted in the WASBuilder case study.
Choose the baseline before the pilot
The most common reason an AI project cannot show a return is that nobody wrote down how the work went before it. Once the tool is in place, the old numbers are gone, and the argument about whether it helped becomes a matter of opinion.
- Choose one number. Hours a week spent on the task, invoices processed per day, errors caught after the fact, or time to answer a customer. One that someone already counts is best.
- Record it for a few weeks first. A single good or bad week is not a baseline.
- Write down what counts as success. For example: the same work in half the hours, with no more errors than before.
- Count the whole cost. The build, the monthly running cost, and the time your staff spend checking what the AI does.
Choosing that number is the first thing an AI Audit does.
How long payback should take
Most companies are not seeing a return yet. In PwC's 2026 Global CEO Survey, 56% of CEOs said AI had brought no significant financial benefit to date (PwC 2026 Global CEO Survey). That is mostly a measurement and scope problem, not a sign the tools do not work.
Expectations also differ inside the same company. CEOs tend to expect AI to pay back faster than their CFOs do, and a project judged against the CEO's timeline can look like a failure to the CFO before it has had a fair chance. Agree on the payback horizon in writing before the pilot starts, alongside the baseline.
A pilot that automates one job should show its effect on the baseline number within a few months of going live. If the number has not moved by then, more time rarely fixes it: the process around the tool usually needs changing.
Why pilots don't scale
A pilot is often a few keen people trying a tool. It shows promise, and then nothing changes across the company. The usual reasons:
- No baseline. Nobody can show the pilot worked, so nobody will fund the next step.
- The tool sits beside the work, not in it. If staff have to copy results from a chat window into the real system, the saving disappears at scale.
- No owner. The pilot was an experiment, and nobody's job is to make it the way the work is done.
If your company has already been through this, see We tried ChatGPT and nothing changed. What do we do now?
Next step
Tell me the process you want AI to take on, and I will help you choose the number to measure it by, record the baseline, and estimate what it would cost to build and to run.