The gap is trust, not technology

There is a sentence about Germany's mid-sized companies that has been repeated so often it sounds like a measurement: the Mittelstand is sleeping through AI. The sentence was convenient. It flattered everyone who felt faster. And it has broken against the current numbers. What the data actually shows is more precise and less comfortable: adoption is accelerating sharply, and what slows it is not missing technology but missing certainty. The Mittelstand's AI gap is not a tech problem. It is a trust problem. And trust problems have a different cure than tech problems.

This is an analysis, not a survey of our own. We read the primary sources (the Bitkom study report page by page, the two OECD reports each on its own terms), and we state the window every figure was measured in. That is the point of this piece: certainty comes from verifiable claims, not from mood.

Section OneThe cliché died on contact with the numbers

The most reliable running series on AI use in the German economy comes from the digital-industry association Bitkom: an annual survey of companies with 20 or more employees, most recently 604 in the sample. The verified series reads: 9 percent (2022), 15 percent (2023), 20 percent (2024), 36 percent (2025), nearly doubling within a single year. And the latest wave, fieldwork in early 2026: 41 percent of companies use AI.

Bitkom, "Studienbericht Künstliche Intelligenz 2025," Figure 15; survey of 604 companies with 20+ employees, fieldwork calendar weeks 27–32/2025.

Bitkom, press release "Digitalisierung der Wirtschaft: Fast jedes Unternehmen beschäftigt sich mit KI," March 11, 2026; fieldwork early 2026, n = 604.

The rest of the distribution tells the same story. Only 17 percent of companies now say AI is not a topic for them at all. Another 47 percent are planning or discussing its use. The still-circulating claim that the vast majority of the Mittelstand uses no AI whatsoever is incompatible with every current measurement. We could not find a solid source for it. Anyone still painting that picture is working from outdated numbers.

But if the adoption curve points steeply upward, where does the persistent sense of hesitation come from? The answer sits in the same study, one figure further on. You just have to read the list the right way up.

Section TwoThe barrier list, read from the bottom

The same Bitkom survey asks companies what keeps them from using AI. Start at the bottom of the list, because that is where the real news is: only 23 percent name missing use cases. Ethical concerns sit in last place at 17 percent. The Mittelstand knows what it would use AI for, and it does not reject the technology on principle. This is not the profile of an economy that doesn't want to.

Now the top of the list. Legal hurdles and uncertainty: 53 percent. Missing technical know-how: also 53 percent. Missing staff capacity: 51 percent. Data-protection requirements: 48 percent. The fear that company data ends up in the wrong hands: 39 percent. In between: results that can't be traced (38 percent), poor output quality (36 percent), missing budget (36 percent), worry about future legal restrictions (35 percent).

Bitkom, "Studienbericht Künstliche Intelligenz 2025," Figure 20; all companies, multiple answers allowed.

At the bottom of the barrier list sits the missing use case. At the top sits the missing certainty.

That is the actual diagnosis, and it changes the treatment. You fix a tech problem with better tools. You fix a trust problem with clear rules, verifiable commitments, and results that can be traced. The market has enough tools.

Section ThreeThe size gap is real, and it is a trust gap

Inside the economy, the divide runs along company size. Bitkom asks which companies give their employees access to generative AI: 21 percent of companies with 20 to 49 employees do, 31 percent of mid-sized ones (50 to 499), and 43 percent of large companies with 500 or more. At the other end of the spectrum: 82 percent of startups use AI, 87 percent of those generative.

Bitkom, "Studienbericht Künstliche Intelligenz 2025," Figure 21 (genAI access by size class) and the startup finding from the same study.

The OECD measures the same pattern with a different panel and a different definition. The figures therefore cannot be netted against the Bitkom series, but the direction is identical. In the official statistics of OECD countries, the share of businesses with ten or more employees using AI rose from 5.6 percent (2020) to 14 percent (2024); small firms with 10 to 49 employees sit at 11.9 percent, mid-sized firms with 50 to 249 at 20.4 percent.

OECD, "AI adoption by small and medium-sized enterprises," December 2025; official statistics with its own survey definition, not to be blended with the Bitkom series.

Why does it hit the small firms harder? Not because the tools are worse for them: the same models sit in every browser, and the startup number proves small organisations can move very fast. The difference is the reassurance infrastructure. A large corporation has a legal department, a data-protection officer, a committee that can issue a sign-off. A business with 30 employees has an owner who is personally on the hook if the judgment call was wrong. Under equal uncertainty, hesitating is the rational answer for her, until someone lowers the uncertainty.

Startups have nothing to lose and run at 82 percent. The Mittelstand built everything that is at stake, and waits for certainty.

Section FourUsing is not yet integrating

A second OECD study (a separate survey, not the official statistics) shifts the view from breadth to depth. In the April 2026 D4SME survey (2,018 responses from twelve countries, fieldwork Q4 2025 to Q1 2026), 61 percent of small and medium-sized enterprises report using at least one AI-enabled application. But 76 percent of those users are classified by the OECD as "AI novices": beginners who operate a tool here and there without having integrated it into their processes.

OECD, "Empowering SMEs in the age of AI," April 2026; D4SME survey, 2,018 responses, 12 countries, Q4 2025–Q1 2026.

A chatbot in a browser is not AI in the business. Between "someone in the building uses a language model" and "our processes are built on it" lies exactly the distance the barrier list describes: whoever wants to integrate needs answers on law, data protection, and data sovereignty. Those are the items 53, 48, and 39 percent of companies name as brakes. For trying things out, you don't need them. Which is why experimentation is everywhere and integration is rare.

So the two measurements don't contradict each other; they describe two floors of the same building: use has grown broad, integration has stayed thin. And the threshold between them is built out of trust, not out of technology.

Section FiveWhat closes a trust gap

If the diagnosis holds, something unfashionable follows: clear rules accelerate. The common narrative treats regulation as innovation's brake block. For the companies at the top of the barrier list, the opposite is true: 53 percent name legal uncertainty as an obstacle, another 35 percent fear future restrictions. What slows these firms is not the existence of rules: it is their blur. A rule that says plainly what is allowed is no shackle to a cautious business. It is a permission slip.

And here the analysis tips from observation into opportunity. A trust gap this size is also an open lead: the majority is waiting. Whoever writes down their own rules first (which tools, which data, which sign-offs, which controls) converts the uncertainty that paralyses everyone else into a documented process. That firm can explain to customers, employees, and partners how it works with AI while the competition is still wondering whether it may. The lead does not come from the boldest model. It comes from the clearest record.

Regulation, done right, does not close doors. It closes a confidence gap.

In ClosingLegibility is the antidote

The through-line of everything we publish is a single claim: machines can only act on what they can clearly read, and people only trust what they can clearly follow. Outside, that means: a language model recommends the companies whose offer the web has made unambiguously legible. Inside, it means: a business integrates AI only once its own rules, data, and responsibilities are unambiguously legible to the workforce, to the regulator, to the tools themselves. The Mittelstand's barrier list is the invoice for missing legibility in both directions.

The numbers in this piece say the race has long since started, and the starting line is not the technology. 41 percent use AI, 76 percent of users are at the beginning, and the top of the barrier list is addressable with rules and documentation, not with a bigger model. The companies that make themselves and their tools legible first take the lead that everyone else's hesitation leaves open.

You can measure how legible your company already is to the machines at /signal-index/, or write to us at /contact/.