Every automation proposal ends at the same figure: so many hours per week, times your hourly rate, equals so much per year. The calculation is seductive because it is correct - as long as nobody asks what happens to those hours afterwards. A business case for automation falls apart at exactly that point, not because of the hourly rate but because of an assumption nobody writes down: that an hour saved is automatically an hour less paid for. This page shows what an hour of work really costs, where the saved time measurably goes, and which three routes exist for it to reach your accounts.
What does an hour of work really cost?
Most calculations start with the gross monthly salary divided by 173 monthly hours. That result is wrong in two directions: it leaves out the costs that arise alongside the salary, and it divides by hours that were partly never worked.
The clean figure already exists. Eurostat reports labour costs per hour actually worked, meaning after holidays, public holidays and sick leave have been taken out. For 2025, Austria sits at EUR 46.30 across the whole economy, Germany at EUR 45.00, the EU average at EUR 34.90. Of the Austrian EUR 46.30, EUR 33.80 are wages and salaries and EUR 12.50 are non-wage costs. That is 27.1 per cent, and it is the transferable part: whatever you pay in wages per hour worked, multiply it by roughly 1.37.
In the sectors where automation happens, the figure runs higher. Professional, scientific and technical activities cost EUR 54.90 per hour worked in Austria and EUR 60.40 in Germany. Information and communication sits at EUR 61.60 in Austria and EUR 58.50 in Germany.
Why the distinction between hours paid and hours worked matters so much is visible in the calendar. A full-time employee in Austria has roughly 250 working days a year. Of those, 25 days are paid annual leave under section 2 of the Urlaubsgesetz (30 working days in the six-day count) and 13 are statutory public holidays under section 7 of the Arbeitsruhegesetz. That is 38 paid days without work before the first day of sick leave. Dividing by paid hours therefore produces an hourly price that does not exist.
That is the smaller source of error. The bigger one comes now.
Where does the saved time actually go?
There is one study that answers this question not with self-assessment but with administrative records. For "Large Language Models, Small Labor Market Effects" (version of 15 July 2025), Anders Humlum and Emilie Vestergaard linked two large adoption surveys covering eleven heavily exposed occupations to Danish register data, meaning actual reported hours and earnings.
Reported time savings average 2.8 per cent of working hours. They range from 6.8 per cent among marketing professionals whose employers actively encourage use, down to 0.6 per cent among teachers without such encouragement. For comparison, the same paper cites controlled experiments "often exceeding 15%". The reason for the gap is mundane: people who use these tools typically use them only every third or fourth day, and even on days of use only 5 to 6 per cent of working hours involve actually using them.
The finding that decides a business case, though, sits in a footnote to Table D.2: 80 per cent of users say they redirect the saved time into other job tasks. Fewer than 10 per cent take additional breaks, and around 25 per cent spend more time on the very task where they had just saved some. Part of the time also flows into work that did not exist before: of the newly created tasks, only 41 per cent are productive use, while the remaining 59 per cent are integration, quality review and oversight.
The result on the money side follows. Asked directly whether the tools had changed their earnings, 97.7 per cent of users said no. The authors' difference-in-differences estimates find no effect on earnings or on recorded hours, with confidence intervals ruling out effects above 1 per cent.
Then there is a cost item that appears in no proposal: reworking output that was produced automatically. In September 2025, BetterUp Labs and the Stanford Social Media Lab surveyed 1,150 desk workers in the United States. Forty per cent had received material in the previous month that looked professional and did not hold up. The effort per incident averaged around two hours, extrapolated to USD 186 per employee per month. Those are self-reported figures from a different labour market, not measured times, and the extrapolation comes from the vendor of a coaching platform. As an order of magnitude for a line item usually entered as zero, it still earns its place.
Which three doors turn time into money?
Saved time shows up in your accounts by three routes only. If your business case names none of them, it is converting comfort into euros.
First: you sell more. The freed time goes into work somebody pays for. That requires two things which are rarely checked - that the demand exists at all, and that the time is freed at the point which limits throughput. An hour less spent checking invoices wins no additional order if the bottleneck sits in production.
Second: paid hours disappear. Overtime, agency staff, an external supplier, a move from part-time to full-time that does not happen. This is the only route on which the saving becomes visible in the result immediately and without detour.
Third: a planned hire does not happen. The team does not grow while the volume does. This door can only be used honestly if the position was genuinely planned beforehand; otherwise it is an invented saving.
The difference is substantial. Take an automation that saves twelve hours a week in an agency with four people involved. At the Austrian sector figure of EUR 54.90 per hour worked, that comes to EUR 658.80 per week and roughly EUR 2,850 per month. That is the number in the proposal. If those twelve hours are spread evenly across four people, it is three hours per person per week, or 36 minutes a day. Those 36 minutes make nobody redundant and write no invoice. If the same twelve hours land with one person, they are just under a third of a full-time position - and for the first time a quantity you can discuss in a staffing plan.
What belongs on the cost side that nobody writes down?
Implementation is always in the proposal. The rest rarely is. A complete list adds running costs including licences and usage-based charges, maintenance whenever a form, a contract or an interface changes, checking the output - those 59 per cent of new tasks in the Danish study are precisely this - and rework when something passes through wrongly. Anyone who enters these four items arrives at a different process cost calculation than the proposal, and at one that still holds twelve months later.
For the result of that calculation there are two formats, and only one of them can be verified. An ROI in per cent without a period is not a statement: 300 per cent over five years and 300 per cent over nine months are entirely different things. The payback period, by contrast, names a month in which the accumulated savings overtake the accumulated costs. You can write that month in the calendar and check afterwards whether it held. This is exactly where many initiatives come apart, as we showed in When AI projects fail: nobody measures, because nobody set a measuring point beforehand.
How do you spot a calculation that will not hold?
Five markers are enough for a first check, and none of them requires expertise.
Percentage ranges without a reference point. "30 to 70 per cent time saved" and "200 to 400 per cent ROI" are not results, they are spans wide enough to fit any outcome.
Studies that do not publicly exist. One example from the research for this article: a German-language consultancy page dated 2 April 2026 states that "7 out of 10 SMEs report an average revenue increase of 14.8% after systematic automation" and credits a vendor name with no link. We searched for it on 16.08.2026 and found no such survey - neither a publication nor a description of the method. That does not prove it does not exist; it proves nobody can read it, and for an investment decision that amounts to the same thing.
Misdated evidence. The same page files the "95 per cent of AI pilots" claim under the year 2024. The report that put that figure into circulation is from 2025, and what it actually says is something we took apart in We wanted to check the most quoted AI figure.
An hourly rate without provenance. Ask whether it rests on hours paid or hours worked, and whether non-wage costs are included. Both answers can be checked against the Eurostat figures above.
No afterwards. If the proposal contains no date on which the result gets measured, the business case is a sales document. That is fine, as long as everyone calls it that.
What to do today
Three levers, in this order:
First, take a number instead of a range. Time one single process for a week with a stopwatch rather than estimating it. Everything else builds on that, and the measurement costs less than any argument about estimates.
Second, name the door before you buy. More work sold, fewer hours paid, or a hire that does not happen. If none of the three fits, the automation is not therefore wrong - it then pays off in reliability and in nerves, and that belongs in the record as such rather than in the profit and loss.
Third, book the follow-up measurement now. One date, one metric, one person responsible. Without that date you never learn whether your calculation held, and the next business case gets estimated all over again.
If you like, we will work through one of your processes together - with your real numbers and the honest result, even when it is zero. 🙂