AI · Productivity · Calculator
How much routine work does AI save? A calculator based on studies
The studies show very different numbers – from just under 3% to over 40%. Instead of a single number, the calculator gives three scenarios, each with its source, in hours and euros for your team.
The calculator
Enter team size and routine hours and choose your sector – the hourly costs come from the Destatis statistics and can be overwritten. Tool and rollout costs are my own assumptions: please replace them with your own values. Everything is calculated in your browser, no data is sent anywhere.
Result
| Scenario | Hours saved / month | Value / month | Net / month (after tool costs) | Net in year 1 | Payback |
|---|---|---|---|---|---|
| Cautious (2.8 % · Humlum & Vestergaard) | – | – | – | – | – |
| Typical (14 % · Brynjolfsson et al.) | – | – | – | – | – |
| Optimistic (25 % · Dell’Acqua et al.) | – | – | – | – | – |
A model, not a forecast. Basis: 46 working weeks per year; the share saved is applied to your routine hours.
How to read the result
- Cautious (2.8%): the measured average over all working hours in Denmark. Here it is applied to your routine hours – so the estimate is deliberately very conservative.
- Typical (14%): productivity gain in the customer service of a real company, measured across more than 5,000 employees.
- Optimistic (25%): speed gain on tasks that suit AI well (Harvard/BCG experiment).
The studies at a glance
| Source | What was measured | Result |
|---|---|---|
| Brynjolfsson, Li, Raymond (NBER) | Customer support, 5,179 agents | +14% issues resolved per hour; +34% for novices, minimal effect for experienced agents |
| Noy & Zhang (Science 2023) | Writing tasks, 453 professionals | Time −40%, quality +18% |
| Dell’Acqua et al. (Harvard/BCG) | 758 knowledge workers, 18 tasks | +12.2% tasks, 25.1% faster; on a task outside the AI frontier 19% fewer correct solutions |
| Peng et al. (GitHub Copilot) | One programming task in an experiment | 55.8% faster (lab result) |
| METR (2025) | 16 experienced developers in their own large projects | 19% slower with AI – while they felt 20% faster |
| Humlum & Vestergaard (Dänemark, 2025) | 25,000 workers, 11 occupations | On average 2.8% of working time saved, no significant effects on earnings or hours |
| Destatis (29.04.2026) | Labour cost per hour in 2025 | €45.00 on average (estimate), €27.40 to €69.70 depending on sector |
Looking ahead: what these figures could mean for IT careers in 10–20 years is in The world in 10–20 years: three scenarios. An example of AI under human control is AI content with human-in-the-loop.
Limits of the calculation
- The lab is not everyday work: the high figures come from experiments with selected tasks. In real operations the numbers are much lower.
- Not everyone benefits equally: novices gain more than experienced people, and on tasks outside AI’s strengths quality drops. Whoever only measures speed overlooks review and correction time.
- Feeling is not a measurement: in the METR study the developers believed they were faster, but were not. Measure your own tasks before extrapolating.
- The hourly costs are an average: Destatis extrapolates the 2025 values from the 2020 labour cost survey. They include non-wage labour costs.
What I do with it myself
At PflegeOasis an AI creates the drafts of the guide articles offline, and a human reviews every text before publication. The AI speeds up the work; responsibility stays with the human – which is why review and correction time belong in the calculation. More in the note AI content with human-in-the-loop.
Keep reading
AI · LLM · Content pipeline
AI content with human-in-the-loop: 200+ guide articles from LLM drafts
With care law, AI mistakes are not acceptable. This is how I use a language model as a writing aid without handing it the responsibility.
Case study · Health · Guides
PflegeOasis: care-level calculator, objection-letter generator and AI content with human review
A free portal that helps relatives find their way through the German care system – with calculators, checklists and more than 200 guide articles.
AI · Labour market · DevOps
The world in 10–20 years: three scenarios instead of a forecast
Nobody knows what the world looks like after AI, robots and drones. What one can do: collect measured numbers, derive scenarios from them and ask what stays the same in all of them – from the view of someone working in IT and DevOps.
More from practice
Find more notes, case studies and projects on the homepage.
All notes Case study MeisterMeld Case study PflegeOasis Case study Kernbohr-Expert Back to the homepage
Comments
Join the discussion: comments run on GitHub Discussions (sign in with a GitHub account). Content from GitHub/giscus is only loaded after you click. Privacy