Esta publicación aún no está en español. Aquí está la versión original.
Six months after the start, there is a spreadsheet in the controller's office. Licence costs are up. Two people from quality spent weeks going through old test reports and collecting examples. Headcount is unchanged. In the bottom right corner is the phrase every AI project dreads: no measurable effect.
This is where many stop. The project is "re-assessed", next year's budget is cut, and the steering committee concludes that AI is probably not ready for the Mittelstand yet. That is understandable. It is also often exactly wrong, and economic research has a name for it.
The J-curve
General purpose technologies such as AI require large complementary investments: new processes, new products and business models, new skills among staff. These investments are mostly intangible and poorly measured in the national accounts, even when they create real value for the firm. The consequence, as the economists Brynjolfsson, Rock and Syverson model it: in the early years of a new general purpose technology, productivity growth is underestimated, and later, when the intangible investments bear fruit, it is overestimated (NBER). Plot measured productivity over time and it looks like a J: first down, then steeply up.
The model describes a whole economy. We carry it over to a single company, and there it looks like this: The weeks quality spends going through old test reports are not lost time. They build an asset no balance sheet shows: knowledge that can now be found, and processes that are now written down. The controller's spreadsheet just doesn't see any of it.
In concrete terms, these are things like: describing tasks so they can be checked. Collecting good and bad examples from real cases. Cleaning up master data nobody has touched in years. Deciding who may see what. Showing colleagues how to check an answer instead of simply believing it. None of this is a software purchase, and all of it costs working time that shows up in the spreadsheet as expense.
Why patience is short in Germany right now
The problem is that patience is especially short right now. In Germany, labour productivity per hour worked rose only 0.4 per cent in 2025, while wage costs per hour rose 5.0 per cent (Destatis). Under that pressure, people want results next quarter, not the year after next.
Then there is a size gap. In 2025, 26 per cent of German companies used AI: 57 per cent of large, 36 per cent of medium-sized and 23 per cent of small ones (Destatis). So large companies use AI more than twice as often as small ones, and many of them have probably reached the dip earlier and left it behind. Small and medium-sized companies are more often still at the start or in the middle of it, exactly where the numbers look worst.
Across the EU, things are moving fast: the share of enterprises using AI rose from 8.1 per cent in 2023 to 13.5 per cent in 2024 and 20.0 per cent in 2025 (Eurostat). If you give up in the dip now, you start again from scratch in two years, while others are already on the steep side of the J.
This does not mean that every project that looks bad is secretly going well. Some projects really do fail. The difference is whether they build knowledge and processes in the dip or just pay for licences. That can be measured, if you choose the right measures.
Where the first gain really shows
If the first year's savings don't show up in headcount, where do they show up? A field study of 5,179 customer support agents gives a clear hint. An AI assistant raised issues resolved per hour by 14 per cent on average, but by 34 per cent for novices, with little effect on the most experienced (NBER).
So the gain sits where onboarding is most expensive. For an industrial company, that means the first effect is not fewer jobs but how quickly a new colleague can work on her own and how often the experienced experts get interrupted. Hardly any controlling department measures either. Both are expensive.
What to measure instead
If the usual metrics make the first year look bad by design, you need different ones. Three of them can be collected in any company without a new system:
First, time to a correct answer. How long does it take today to answer a technical question from sales, service or production with proof? Measure it before and after, on real questions.
Second, onboarding time. How many weeks does a new employee need before she handles a typical task without help? If that number falls, the project is working, even if licence costs went up.
Third, the load on the experts. How many questions a week land with the three or four people who know everything? If that number falls, the company wins back exactly the time it needs most.
What matters is writing down the starting values before you begin. If you only start measuring after six months, you have nothing to compare with and end up back at the licence bill.
The reasons for hesitating can be solved
The most common reasons for hesitating are not a stance against AI. Of German companies without AI that had already considered using it, 72 per cent named missing knowledge as a reason, 62 per cent uncertainty about the legal consequences (Destatis). Both reasons shrink over time once someone starts. Neither shrinks if you stop after six months.
Our assessment
An assessment, not a fact: the Mittelstand companies that report clear AI gains in 2027 will mostly be the ones that measured onboarding time and expert load in their first year, not the ones that measured licence savings.
The practical advice is simple: budget for the dip. Before you start, write down how long you will hold on and how you will know it is working. Then the spreadsheet after six months is not a verdict, but a progress report.
Fuentes
- National Bureau of Economic Research (Brynjolfsson, Rock, Syverson), oct 2018: The Productivity J-Curve: How Intangibles Complement General Purpose Technologies (NBER Working Paper 25148)
- Statistisches Bundesamt (Destatis), 15/01/2026: Bruttoinlandsprodukt 2025 für Deutschland (Statement zur Pressekonferenz)
- Statistisches Bundesamt (Destatis), 2025: Nutzung von IKT in Unternehmen (Erhebung 2025)
- Eurostat, 11/12/2025: 20% of EU enterprises use AI technologies
- National Bureau of Economic Research (Brynjolfsson, Li, Raymond), abr 2023: Generative AI at Work (NBER Working Paper 31161)
