What is inside
- Six measured figures – five on AI use in Switzerland, one international on its effect
- Why many AI promises do not hold – what experiments with control groups found
- The five things that are established – including data protection and liability
- The test: four questions for any AI promise, before you buy licences
- The metric that misleads – and the one to track instead
- A template for measuring one task over two weeks, and three actions for Monday
- Length
- 10 pages
- Languages
- German, English
- Version
- September 2026
- Sources
- 22
Download the guide
The link arrives by email and stays valid – so you still have the guide to hand later.
A look inside
Four questions every AI promise has to answer
The test from Part 4 of the guide. It works for any figure anyone shows you – ours included.
Measured or estimated?
Many time-saving figures are users' own estimates. In a randomised experiment, developers estimated they had needed 20 % less time with AI – measured, they needed 19 % more.
Is there a control group?
Without a comparison group, any improvement can also be explained by a quiet month. "Rolled out, then faster" is a sequence of events, not an effect.
Inside or outside the frontier?
AI helps with individual tasks, not whole professions. A study of 758 consultants found both: inside the frontier they got faster and better with AI; on a task outside it, they were right less often with AI than without it.
Who checks the result?
An AI output is a draft until a person has checked it. Estimated time-saving figures usually leave the checking time out. The liability for the result is likely to be yours.
The sources
Where the figures come from
The Swiss Federal Statistical Office and KOF at ETH Zurich on use in Switzerland, AXA/Sotomo on SMEs, KPMG and the University of Melbourne on behaviour at work. Plus experiments and register studies from the Quarterly Journal of Economics, Science and Organization Science, from NBER and METR, the UK government's Copilot evaluations, the European Investment Bank, the FDPIC, the Federal Office of Justice and the Moffatt v. Air Canada decision. The full list with dates and references is on the last page of the guide.
Who it is for
For managing directors and decision-makers in Swiss SMEs who are deciding on AI licences or an AI project – or who have noticed that people in the firm have been using it for some time. No prior knowledge needed, just a willingness to ask for the measurement.
All Academy materialThe test, applied to your firm
In 30 minutes we walk through one task from your day-to-day work – whether it sits inside the frontier, how to measure it and which data rule it needs. With numbers, and no sales pressure.