Measure AI ROI one workflow at a time. Time each step of the workflow before you change it, then measure it again, the same way, after the change has run on live work. Count the minutes people spend to check the AI and the cost to run it. Report the hours freed and what those hours are worth, not "savings", and let finance decide how much of that time turns into money.
This sounds modest. It is. It is also the only kind of AI ROI number we have seen hold up in a board meeting.
Why most AI ROI numbers do not survive finance
AI promises big numbers, and most companies cannot show them. In McKinsey's 2025 global survey, 88% of organisations use AI in at least one function, but only 39% report any effect on EBIT (McKinsey). Part of that gap is real: many AI efforts never change how work is done. But part of it is a measurement problem. Companies that did change something often cannot prove it. We see three causes again and again.
They measure usage, not work. Licences, active users and prompts per week are easy to count, so they end up on the dashboard. They show that people open a tool. They do not show that a quote went out faster or that a file needed less rework. A team can use AI every day and still run the same workflow at the same speed.
There is no baseline. Nobody wrote down how long the work took before the change. Afterwards, everyone agrees it "feels faster", and nobody can say by how much. A result without a starting point is an opinion, and finance is right to treat it as one.
The numbers are counted twice, or counted as cash. A team-wide volume gets multiplied by the number of people in the team. Or freed hours get presented as money saved, while the same people are still on the payroll doing other work. Both make the business case look larger, and both fall apart the moment a controller asks one question.
Measure workflows, not tools
The right unit for AI ROI is the workflow: a piece of work with a start, an end, an owner and a volume. Client onboarding is a workflow. Supplier invoices are a workflow. "Copilot" is not, because one tool touches dozens of workflows a little, and you cannot isolate its effect.
A workflow gives you everything a measurement needs. It has steps you can time. It happens a known number of times a week. It has someone who owns the outcome and can confirm what changed. And it has a natural before and after, because a new version of a workflow goes live on a date.
If you want to know which workflow to measure first, start where the hours are. We cover that choice in which processes to automate with AI first.
The six numbers that matter
For each workflow, six numbers are enough. Collect them per step, not only for the whole workflow, because the steps tell you where AI helps.
| Measure | What it tells you | Example |
|---|---|---|
| Volume | How often the work happens | 42 new clients a week |
| Active time | Minutes of hands-on work per case | 15 minutes to check documents |
| Waiting time | Time a case sits between steps | 2 days for a missing payslip |
| Rework | Share of cases that come back | 1 in 5 files |
| Check time | Minutes a person spends to check the AI | 3 minutes per file |
| Running cost | Model use and licences for this workflow | Per month |
Two of these are easy to forget. Waiting time often matters more than active time. A client does not see that a check took 15 minutes; they see that their file sat still for two days. AI that drafts the missing-documents email the moment a file arrives can cut waiting time far more than it cuts work. Check time keeps the result honest. If a person reviews every AI output, that review is part of the new workflow and belongs in the after number.
How to build a baseline in a week
A baseline is not a consulting project. For one workflow it takes days.
- Write down the steps. Who does what, in which tool, and where the work moves to another person or team. Six to ten steps is typical.
- Ask the people who do the work. For each step: how often does it happen, and how many minutes does it take? Ask two or three people, not one, and note where their answers differ. The differences are often the most useful finding.
- Check against real cases. Take 20 recent cases and check the numbers against them, or against system data such as timestamps in the CRM or the mailbox.
- Label every number. Estimated, if it comes from an interview. Observed, if you measured it. Never mix the two without saying so.
- Put a date on it. This is your baseline. Everything afterwards is compared with it.
Do this before you build anything. Once a new workflow is live, nobody remembers how the old one ran.
The arithmetic, step by step
Keep the calculation simple and show every step. A reader should be able to check it with a phone calculator.
| Step | Calculation | Result |
|---|---|---|
| Time per day | 10 times a day × 15 minutes | 2.5 hours |
| Time per year | 2.5 hours × 220 working days | 550 hours |
| Worth of that time | 550 hours × €48 per hour | €26,400 a year |
Now the most common mistake. Suppose a team of five shares those ten cases a day. The total is still ten cases, still 2.5 hours a day. Multiply by five and you claim 12.5 hours a day that never existed. Multiply by the number of people only when each person handles their own volume. This single error makes many AI business cases five times too large, and it is the first thing a careful controller checks.
Keep active time and waiting time in separate columns too. Waiting time does not cost wages, but it costs clients.
Capacity is not cash
Freed time is capacity. It turns into money only when something changes: you hire one person fewer next year, you stop paying overtime, or the same team handles more work without growing. Until one of those happens, the hours are real, but the money is not yet on the books.
That is why we report three numbers per workflow, and keep them apart.
| Number | What it means | Who sets it |
|---|---|---|
| Expected value | What the plan says the new workflow will free | The team, at the plan stage |
| Verified value | What you measured after go-live, the same way as the baseline | Measured on live work |
| Confidence | How sure you are, given how much live data you have | Stated openly |
Then finance decides which freed hours become money, and when. This is not a weakness in the case. It is what makes the case believable. A board that sees "24 hours a week freed, verified over six weeks, confirmed by finance" trusts that line far more than "€300,000 in annual savings".
Count the costs honestly
A one-sided ROI number fails as surely as an inflated one. Put the costs next to the gains:
Check time. Minutes people spend to review AI output. It shrinks over time as the workflow earns trust, but it starts high, and it belongs in the after number.
Running costs. Model use, licences and connections for this workflow, per month.
Building and change effort. The hours your people spend to design, test and learn the new workflow. This is a one-time cost, and often the largest one.
Maintenance. Systems change, a sign-in expires, a form gets a new field. Someone owns the workflow and fixes it. Count that time.
When to report a result
Measure again only after the new workflow has run on live work for a few weeks. Use the same people, the same method and a similar mix of cases. A quiet week in August is not a fair comparison with a busy week in March.
With less than three weeks of live data, call the result provisional. Check it again after twelve weeks. The second measurement is usually more convincing than the first, because the check time has come down and the team has stopped improvising.
A worked example
This example is illustrative: a client onboarding workflow at a mortgage adviser, 42 new clients a week, measured the same way before and after.
| Step | Before (minutes) | After (minutes) | What changed |
|---|---|---|---|
| Read the email, save the documents | 5 | 0 | Done by the workflow |
| Check documents against the checklist | 15 | 3 | AI checks, a person approves |
| Chase missing documents | 6 | 2 | AI drafts, a person sends |
| Enter client details in the CRM | 12 | 1 | Filled by the workflow, a person reviews |
| Sanctions screening | 6 | 6 | Unchanged, stays with Compliance |
| Hand the file to an adviser | 4 | 1 | Summary sent automatically |
| Total per client | 48 | 13 |
At 42 clients a week, the workflow goes from about 33.6 to 9.1 hours a week. That frees about 24.5 hours a week, or roughly 106 hours a month. At €48 per hour, that time is worth about €5,100 a month, before running costs. Finance then decides how much of it becomes money: here, the team handled growth without a new hire.
How to present it to your board
One page per workflow is enough. Show the workflow before and after, step by step. Show the hours freed, verified or provisional, and the period you measured. Show the costs. State your confidence. End with the next workflow you plan to change and why. Boards do not need more detail than that. They need to see that every number has a source.
Frequently asked questions
What is a good ROI for an AI workflow?
Compare the value of the hours freed each month with the running cost each month, and keep the one-time building effort separate. A first workflow that frees many times its running cost is a strong start, even if only part of that time turns into money this year.
Should we track how often people use Copilot or ChatGPT?
Use it as a signal, not as a result. Usage shows access. The result is a workflow that takes less time, waits less or needs less rework. We explain the difference in why Copilot licences don't change how you work.
Who should confirm the numbers?
Finance, or the director who owns the budget. Not the team that built the workflow, and not the vendor.
How soon can we show a result?
A provisional result after three to six weeks of live work. A confirmed one after about twelve.
Want to know what your workflows cost today? Book a 30-minute call. For where to begin, read where to start with AI in a mid-size company.