Calculating the ROI of AI projects: method, studies, worked example

How do you calculate the ROI of an AI project?

The formula is the same as for any investment: ROI = (benefit − costs) ÷ costs. The hard part is not the formula but what goes into it. A period of three to five years has proven useful, because rollout and training costs fall mainly in the first year.

Side	What goes in
Benefit	working time saved × staff costs per hour; avoided spending, for example on external translation or typing services
Costs, one-off	rollout, connecting your systems, training, hardware (when running it on your premises)
Costs, ongoing	licence, operation, electricity, service contract, internal support (contact person, maintaining the apps)

It is important to distinguish between benefits that affect the budget and benefits that do not. Minutes saved only reduce staff costs if, as a result, posts are not refilled, overtime is reduced or external services are no longer bought in. Otherwise they free up time for work that was left undone. That is a real benefit, but it should be shown separately in the funding application.

How much time does AI save according to sound studies?

The range is wide, depending on whether whole working days or individual tasks are measured:

Study	What was measured	Result
Humlum, Vestergaard (Denmark, around 25,000 employees in 11 occupations)	working time saved in everyday work, survey and register data	on average 3% of working time
Federal Reserve Bank of St. Louis (USA, November 2024)	self-reports by people who use AI	5.4% of working hours, 1.4% across all employees
UK Government Digital Service (20,000 employees, 12 government bodies)	self-reports, three months	on average 26 minutes per day; 17% saved nothing
Noy, Zhang (Science, 453 professionals)	individual writing tasks in an experiment	40% less time, quality rated 18% higher
Brynjolfsson, Li, Raymond (5,172 customer service employees)	enquiries resolved per hour	15% more, especially among new employees
UCLA Health (238 physicians, randomised)	writing time per clinical note	with one of the two tools around 41 seconds less per note, almost 10% more saving than in the control group

Two lessons follow from this: for individual, well-suited tasks (drafts, summaries, replies to recurring questions), large savings are proven. Calculated over the whole working day, the figure remains a few per cent, because only part of the work consists of such tasks.

Why time saved is not automatically money

The UK Department for Business and Trade equipped 1,000 employees with an AI assistant. Result: most saved time, most of all on written tasks; for some tasks, however, it took longer than without AI. The evaluation found no robust evidence that the time saved had increased productivity. The Danish study, too, has so far found no measurable effect on working hours and wages.

For your calculation, this means:

• Discount self-reports: surveys usually produce higher values than measurements.
• Count only suitable tasks: calculate with the tasks for which ready-made apps exist, not with the whole working day.
• Apply a usage rate: not every person uses AI every day. Training and ready-made apps decide how many will.
• Deduct review effort: AI results have to be read and checked.

Worked example: public administration with 200 people

The following values are assumptions for illustration, not a commitment. Insert your own figures.

Item	Assumption	Amount per year
People who use AI	200	–
Time saved per person	1 hour per week (around 2.6% of 39 hours, i.e. below the average of the studies)	–
Working weeks	42	–
Staff costs per hour	€45 (average of the German economy in 2025)	–
Benefit	200 × 1 h × 42 × €45	€378,000
basebox licence, cloud, 200 people	list price, net, 1-year term	€47,932
Internal support	assumption: a quarter of a full-time post	€22,500
Ongoing costs		€70,432
ROI in the current year	(378,000 − 70,432) ÷ 70,432	around 4.4

In the first year, one-off costs come on top, for example the optional setup in your IT (€24,900) and connectors to systems that are not yet in our library. If you halve the time saved to 30 minutes per week, the benefit of €189,000 is still well above the costs.

The break-even calculation is revealing: with a licence of around €240 per person per year and €45 per hour, the licence has paid for itself after just over five hours per person per year. That is about one and a half minutes per working day.

How robust does the calculation become?

The most robust approach is a measurement in your own organisation: select three to five tasks, measure the time per case beforehand, measure again after a few weeks with AI, and assess quality as well. With these values instead of the study values, the estimate becomes a calculation that also stands up to a public audit. Ready-made apps for individual tasks make this easier, because use and result can be attributed to each task.

How does this work with basebox?

With basebox, the cost side is predictable: in the cloud you pay per person per year; with the server you pay once for the hardware plus one licence per server, and everyone may sign in. All features, apps, support and updates are included in the licence. The configurator calculates your own number of people in two minutes, without signing up. Calculate your price →

Note: The worked example is based on assumptions. Study values come from other countries and sectors and cannot be transferred one to one.