Calculating AI ROI on your own figures

Our own page carries a worked example: 1,200 hours of manual work a year, a saving of 68,000 euro, paid back within three months. That number is illustrative, not a promise, and it is exactly usable once you turn it into a model with your own figures in it. This piece is that model.

We wrote earlier about what AI ROI means and about the hidden costs most business cases miss. This piece puts both sides, cost and return, into one calculation model you fill in yourself.

Why benchmarks tell you nothing

A search for "AI ROI benchmark" turns up numbers that have nothing to do with your company. A percentage from an international survey of a thousand companies says nothing about a specific process at a specific hundred-and-fifty-person company in the Netherlands. Every application has a different baseline, and that baseline is exactly what a benchmark averages away.

The only ROI figure with any value is one you calculated yourself on your own process, with your own hours and your own rates.

Working through numbers. Fill in one process at a time, not the whole company.

The model in five lines

Fill this table in for one process, not your whole company at once.

Our own example lands on 68,000 euro saving instead of 57,600, because it counts full value including quality improvement, not just hours saved. Keep your own model simple first: hours times rate. Add quality gains only once you can substantiate them.

The two cost lines you cannot skip

Line 4 above is where most business cases get too optimistic. Two cost items belong in it, and we wrote about them separately already.

The licence or build cost is on the quote. The staff hours for rollout usually are not, and we calculated earlier that for an average implementation those hours can run around 28,800 euro with six people spending six hours a week for two months. Count both, or your payback time is fiction.

A fully worked example

Take a fictional administration team manually checking invoices. Four people spend five hours a week on it each, so a thousand hours a year at forty-eight working weeks. At 48 euro an hour that is 48,000 euro of manual work.

A tool automating three-quarters of that check, with a human still seeing the exceptions, saves 36,000 euro a year. The licence costs 500 euro a month, 6,000 euro a year. Rollout costs, with two people spending four hours a week for two months, 64 hours times 48 euro is 3,072 euro. Total first-year cost: 9,072 euro.

Payback time: 9,072 divided by (36,000 divided by 12) is just over three months. From year two, without rollout costs, the saving is 36,000 euro against 6,000 euro licence, a net 30,000 euro a year.

This is an example, not a promise. It shows how the five lines in the model add up to one number you can discuss with a vendor instead of a feeling.

A calculator over paper. A negative outcome is useful information, not a failure.

Three pitfalls when filling this in

Estimating hours too optimistically. Ask the people doing the work how much time it really takes, not how much the process manual says it should take. Those two numbers often differ by a factor of two.

Counting the full saving when the work only partly disappears. A tool taking over three-quarters of a task saves three-quarters of the hours, not all of them. Calculate with what actually remains as checking work.

Counting quality gains without substantiation. Fewer errors, happier customers, faster delivery are all real benefits, but they only become a number once you can measure them. Add them to your model only once you have a way to check them afterwards.

What a negative outcome means

Not every process returns a positive ROI within a reasonable time, and that is useful information, not a failure. A process with few hours in it, or one where quality is already high, will yield little saving whatever tool you put on it.

That is exactly why we start a scan with the problem, not the system. A model run across several processes points on its own to the one that actually pays off, and stops you spending time on the process that merely looked easiest to automate.

A calculator over paper. A negative outcome is useful information, not a failure.

What to do with a vendor who quotes their own ROI figure

Some vendors present their own ROI percentage, often based on their average client. Do not ask whether the figure is correct, ask how it was built: which hours, which rate, which period. A vendor who cannot break it down into the five lines of this model probably has no clear picture themselves of where the number comes from.

Also ask explicitly about the hours their clients needed for rollout, not just the licence cost. That figure is missing most often, and it is exactly the line that makes most business cases too rosy.

What to do with this model

Fill it in for three processes in your company, not one. Rank them by payback time, not by how interesting the technology sounds. Start with the process combining the shortest payback with the most hours, because that pairs the biggest saving with the fastest proof.

Keep the filled-in model. After the first application you can put the actual hours and costs next to it, and that difference is worth more than any benchmark: it is proof of what actually happens at your company, and it makes the next calculation more accurate.

About this page

The worked example of 1,200 hours and 68,000 euro comes from our own AI Readiness Scan page and is illustrative. The average of 48 euro labour cost per hour comes from the Statistics Netherlands news release on 2024 labour costs. This is the state of play on 19 August 2026.