How to Measure AI ROI: A CEO's Guide
To measure AI ROI, measure one workflow at a time, not your AI budget as a whole. Write down the before number, run the AI version for six to eight weeks, and compare measured results, not how people say it feels. Count a gain only when the time or output it frees up shows up somewhere you can point to on the P&L.
Most companies skip at least one of those steps, and it shows. In PwC's 2026 Global CEO Survey of 4,454 CEOs, 56% said they had seen no significant financial benefit from AI to date, and only 12% said it had delivered both cost and revenue benefits. A year earlier, IBM's survey of 2,000 CEOs found that only 25% of AI initiatives had delivered the expected ROI, and only 16% had scaled enterprise-wide. Some of that is real failure. Some of it is AI that works but was never measured in a way anyone could defend.
This post covers why AI ROI is hard to prove, what to count as the return, a six-step method you can run on any AI project, and how the right measure changes as your company gets better at AI.
Why is AI ROI so hard to prove?
Three mistakes account for most of it.
The wrong unit. Companies try to measure the return on "our AI tools" or on a stack of licenses. Licenses do not produce value. Workflows do. A seat that one person uses to rewrite emails and another uses to cut a two-day report to two hours has no single ROI. The workflow does.
No baseline. If nobody wrote down how long quoting, month-end close or ticket triage took before AI, every after number is a guess. And it is hard to rebuild a baseline once the old way of working is gone.
Feelings instead of measures. People are poor judges of their own speed with AI. In a 2025 METR study, 16 experienced software developers worked on 246 real tasks, some with AI and some without. They expected AI to make them 24% faster and, afterward, believed it had made them 20% faster. Measured, they took 19% longer with AI. METR has since said tools have improved and those early results are out of date, but the lesson holds: survey answers about time saved are not evidence. Timestamps are.
What should you count as the return?
AI ROI is the measured gain from an AI-enabled workflow, minus its full cost, divided by that full cost. The gain usually comes in one of four forms:
- Time: hours per unit of work, such as minutes per quote or hours per close.
- Throughput: more units with the same people, such as tickets resolved per hour.
- Quality: fewer errors, less rework, fewer escalations.
- Revenue: faster response, higher win rate, better retention.
Time is where most AI projects start, and it is where most ROI claims fall apart. An hour saved is not money until it goes somewhere. Here is the test:
- Freed hours absorb growth. You handle more volume without the hire you had planned. The return is the avoided salary.
- Freed hours move to higher-value work. A salesperson who spent most of the day quoting now spends it on outbound. The return shows up in pipeline and win rate, so measure those.
- Freed hours shorten a cycle. Faster quotes or a faster close. The return is in conversion or cash timing.
- Freed hours go nowhere in particular. People are a bit less busy. The return on your P&L is close to zero, even if everyone likes the tool.
This is the one decision most ROI guides skip, and it is a CEO decision: where the time goes has to be chosen before launch, not discovered after.
Real gains can be large when the measure is right. A study of 5,179 customer support agents found that a generative AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled workers. Note what the researchers measured: resolved issues per hour, from the company's own systems. Not a survey. Not a feeling.
How do you measure AI ROI, step by step?
Run this on one workflow at a time. It takes about two months end to end, and the paperwork fits on one page.
- Pick the workflow and name an owner. One process, such as quoting, invoice matching or first-line support, with a beginning and an end. One person owns the number. If you cannot name the workflow in a sentence, it is not ready to measure.
- Capture a two-week baseline. Record four numbers: volume (units per week), time per unit, error or rework rate, and current cost (people time plus any tools). Pull them from systems where you can: ticket timestamps, CRM stages, file histories. Where you cannot, have people log real time for two weeks rather than estimate it.
- Decide where freed capacity goes. Absorb growth, move to named higher-value work, or shorten the cycle. Write it down. This step turns hours into money.
- Set a target, a kill line and a date. For example: time per unit down by a stated percentage within eight weeks, with quality no worse than baseline. Below the kill line on the review date, you stop or redesign. No extensions without a new hypothesis.
- Run it and compare like for like. Use the same metrics as the baseline. Where you can, keep a holdout: some people or some volume still on the old way. For output quality, do a blind comparison: have a reviewer rate AI-assisted and human-only work without knowing which is which.
- Do the math with the full cost. Full cost means licenses or build cost, setup time, training time and the review time people now spend checking AI output. That last one is the cost most teams forget. Then: (measured gain minus full cost) divided by full cost, over a year.
The result is an ROI card: workflow, owner, baseline, target, kill line, where the capacity goes, measured result. One page per project. If a project cannot fill in the card, it is not ready for budget.
How does the right AI ROI measure change as you mature?
The right measure depends on where you are. We use four stages of AI maturity, which we explain in our AI maturity model, and each one calls for a different kind of number.
- Scattered: people use AI on their own, with no shared approach. Do not try to compute a company ROI yet. Measure who uses what, and pick the first workflow to run the six steps on.
- Experimenting: pilots are running. Every pilot gets an ROI card with a kill line. The goal at this stage is not a big number. It is learning which pilots to stop.
- Operating: AI is part of how several workflows run. ROI cards roll up to P&L lines, and workflow owners report them in the same review as the rest of the business.
- Compounding: AI gains feed the next project. The measure shifts to unit economics, such as cost per quote, cost per ticket or revenue per employee, tracked quarter over quarter.
The most common mistake is using a Compounding measure at the Experimenting stage. A leadership team asks "what is our AI ROI?" before any workflow has a baseline, gets no clear answer, and concludes AI does not pay. The fix is to ask a smaller question first: "What did this one workflow return?"
Who should own AI ROI numbers?
Not IT, and not finance alone. The owner of the workflow owns the result, because only they can make the freed time go somewhere. Finance checks the math and the full cost. Between them sits what we call the product function: the person who defines the workflow, sets the baseline and runs the review. We cover that role in our post on why IT should not own AI on its own.
Your part as CEO is short and firm. Ask for the ROI card before approving any AI spend. Decide where freed capacity goes. And keep the kill lines: a project that misses its line twice should stop, however much the team likes it.
What should you do this quarter?
Take the two or three AI bets you care most about. If you have not ranked them yet, our guide on how to prioritize AI projects walks through it. For each one, fill in the ROI card this week, even if the baseline row is still empty. An empty baseline row is a finding: it tells you what to start measuring on Monday.
Then put a monthly review on the calendar, 30 minutes, cards only. Within one quarter you will know which bets are paying, which need redesign and which to stop, with numbers you can defend to your board.
If you would like a second pair of eyes on which of your AI projects can show a real return, Journey offers a free CEO AI audit to help you see where you stand and what to do next.
Frequently Asked Questions
Measure one workflow at a time. Record a two-week baseline of volume, time per unit, error rate and cost. Decide where freed time will go, set a target and a kill line, then run the AI version for six to eight weeks and compare the same metrics. ROI is the measured gain minus the full cost, divided by the full cost.
Usually for three reasons. They measure tools or licenses instead of workflows, they never recorded a baseline before AI arrived, and they rely on how much time people feel they save rather than measured time. Research shows self-reported time savings can be badly wrong, so use system data such as timestamps wherever possible.
Only when the hours go somewhere. Freed time counts if it absorbs growth without a planned hire, moves people to higher-value work you can measure, or shortens a cycle such as quoting or month-end close. If the time simply makes people a little less busy, the return on your P&L is close to zero, however popular the tool is.
Include everything the workflow now costs: licenses or build costs, setup and integration time, training time, ongoing maintenance and the time people spend reviewing AI output. Review time is the cost teams most often forget. Leaving it out can make a project look profitable when the real gain is small or even negative.
For a single workflow, about two months. Spend two weeks capturing a baseline, then run the AI-enabled version for six to eight weeks before the first review. Company-wide AI ROI takes longer, because it only becomes meaningful once several workflows have their own measured results that roll up to lines on your P&L.
