AI in medium-sized businesses: loud on LinkedIn, secret in operation.

Why both pictures are right and what that means for your company: Between corporate headlines and shadow AI in medium-sized businesses lies the gap between individual use and operational added value.

By Martin Salwiczek · 2026-08-14

CEOs of medium-sized companies are facing a puzzle. They read on LinkedIn that AI is replacing entire departments. In their own companies, ... actually nothing is happening. And yet: when they ask, it turns out that an employee is already using ChatGPT for emails. Another is using Copilot for their minutes, and a third is using Gemini for image creation. Management is caught in the middle, wondering if the AI topic is just a hype or if they are missing out on something.

The following text provides insight into some current figures and shows why the introduction of AI in companies is not only a technical but above all an organizational issue. It also outlines possible first steps for entrepreneurs.

Even the statistics disagree

How many German companies use AI? The ifo Institute estimates 54.5 percent for May 2026, up from 40.9 percent a year earlier. Bitkom reports 41 percent active usage for 2026. Both surveys are current and reputable. They only question different companies and count every form of usage – from occasional chatbots to AI in production control.

The breakdown by size is more interesting. According to the ifo Institute, large companies are at 67.2 percent, small firms at 51.2, and medium-sized ones at only 47.2. The middle segment is lagging not only behind corporations but also behind the small ones. Corporations have their own digital departments, micro-businesses try things out at a low threshold, and in between lies the established medium-sized company with entrenched processes, no spare time, and no in-house IT.

The percentages show who is at what stage. They don't explain why the perception differs so much. For that, you don't need to look at company sizes, but at what is happening publicly and what is not.

The loud and the quiet

One side is visible: corporations that are downsizing. In the US, where most of these reports originate, there were around 55,000 announced job cuts in 2025 around 55,000 announced job cuts attributed to AI, a multiple of previous years. The most well-known example is Klarna: Its AI assistant was said to be doing the work of about 700 full-time employees. Later, the founder admitted quality problems in customer service and shifted back to more human interaction. Not an isolated case: in a survey of 600 HR managers with AI-related job cuts, about two-thirds reported having refilled some of these roles. Those who replace staff without paying attention to the quality of work will bring it back later.

The other side remains invisible: medium-sized businesses. Institutionally, little is happening, but the workforce has long since started. According to a study by Software AG, about half of the 6,000 knowledge workers surveyed use AI tools that the company neither knows about nor has approved, i.e., shadow AI. The counter-check from the company side confirms this: in a representative Bitkom survey Under 604 companies with 20 or more employees, 42 percent suspect private AI use, only 29 percent can definitively rule it out, and only 23 percent have rules for it.

This shows both a real need and a risk, as company data flows into tools that no one has vetted. As long as nothing is decided at the management level, each employee decides for themselves. And because no one is aware of these decisions, the company appears from the outside as if nothing is happening.

Usage is not yet value creation

Between these two extremes lies the question that arises in every other discussion: Is AI overestimated? An MIT study from 2025 went through the press as "95 percent fail": So many of the investigated AI pilot projects in companies showed no measurable economic effect. The reason is more interesting than the number. There is no lack of infrastructure, budget, or skilled workers - the systems themselves did not learn. They did not retain feedback, did not adapt to the context, and could not be integrated into actual processes.

A study commissioned by the North Rhine-Westphalian Ministry of Labour starts even earlier for Germany: AI is hardly used strategically in value creation in practice, and its introduction is particularly slow in small and medium-sized enterprises. So the MIT explains why projects fail - the NRW study shows that it doesn't even get that far in many places.

Taken together, this results in an uncomfortable diagnosis: The problem for German companies is neither too little AI nor too little enthusiasm, but the gap between individual use and corporate value creation. Those who have their emails written faster privately save their own time. Nothing changes for the company as long as the work is not organized differently.

The World Economic Forum has researched with PwCwhat companies that actually make money with AI do differently. They change their workflows twice as often as the rest. So they don't just buy AI, but reorganize the work around it.

From a macroeconomic perspective, the restructuring pays off: The Institute for Employment Research expects 0.8 percentage points more growth per year over 15 years and around 4.5 trillion euros in additional value creation - with roughly the same number of jobs, but around 1.6 million of which are created or eliminated. This time, experts and specialists are primarily affected, fewer auxiliary and skilled workers.

What actually works

AI works best where it enables a person to do more, rather than replacing them. Two examples from our work.

For a client, we generated a drone flight video of the planned buildings from a construction sketch, which replaced a physical architectural model. Compared to external model building, the savings were up to 5,000 euros. Here, AI does not automate a job away. It makes something possible that was previously beyond the budget.

In individual coaching with the marketing director of a public employer, we built an AI assistant as a persona. Its task was to challenge the company's own advertising materials. It provided impulses that helped to sharpen the texts to be closer to the customer. The marketing director still made the decisions. The AI provided him with the critical perspective that is often lacking in everyday life.

Two isolated cases prove nothing. However, the pattern is also found in large studies. A Stanford study with over 5,000 service employees showed 14 percent more productivity through AI assistance, and even 34 percent among inexperienced employees: The system passed on to the newcomers what the experienced employees had acquired over years.

In both cases, a human assessed the result. This remains necessary, because a wrong AI result sounds just as convincing as a correct one.

What you can do tomorrow

Based on the findings of this text, there are four steps:

1. Make shadow AI visible A ban only shifts usage to places where no one can see it anymore. Ask openly who uses which tool for what. This takes an hour and shows where the need in the company really lies.

2. Provide an official tool and regulate its use in writing Which data is allowed in, which is not, who decides in case of doubt. This doesn't have to be a rulebook, but it must be binding and findable. Those who are given a usable alternative will no longer resort to private ones.

3. Start with a real task, not the biggest one Something recurring that eats up time and where a mistake causes no damage: pre-drafting an offer, summarizing minutes, having a text challenged. A team will learn more from this than from any presentation on digitalization.

4. Turn the experiment into a process When a use case works, the real work begins: At which point will AI be located in the future, who will use it there in a binding manner, who will check the result? Only here does personal time savings become a business advantage. This step is the most uncomfortable because it affects established processes – that's why most people skip it.

Why this depends on further training

Of technology, process, and competence, competence is the factor that cannot be bought. A tool is ordered, a process is described, but the ability to judge only develops within the company itself. The World Economic Forum expects that by 2030, almost 40 percent of the skills currently in demand will change; 63 percent of employers cite the skills gap as the biggest hurdle to their transformation. For Germany, it names Bitkom lack of know-how as one of the biggest obstacles.

In addition, there is an obligation that many do not have on their radar: Artikel 4 der KI-Verordnung verlangt seit Februar 2025 Maßnahmen für die KI-Kompetenz der Mitarbeiter, seit dem 27. Juli 2026 in entschärfter Form, überwacht seit dem 2. August 2026. Die Rechtslage ist aber der schwächste Grund.

Wer ein Werkzeug nicht versteht, benutzt es entweder gar nicht oder vertraut ihm blind – beides macht Menschen kleiner, als sie sind. Wer weiß, was die Maschine kann und wo sie irrt, bleibt derjenige, der entscheidet. Darum geht es uns: nicht Menschen an eine Technologie anzupassen, sondern ihnen die Sicherheit zu geben, souverän mit ihr umzugehen.

Wie Solanius unterstützt

Wir im Ruhrgebiet kennen Strukturwandel. Wir wissen, dass er selten an der Technik scheitert und fast immer daran, ob die Menschen mitgenommen werden. Genau da setzen wir an. Menschenzentriert heißt für uns: Der Einzelne wird gesehen und in die Lage versetzt, KI als Werkzeug zu nutzen, während er die Verantwortung behält.

Es geht nicht darum, Ihnen ein Werkzeug zu verkaufen. Es geht darum, Ihre Leute so weit zu bringen, dass sie es beherrschen. Wenn Sie wissen wollen, wo Ihr Unternehmen zwischen den beiden Polen steht, sprechen Sie mit uns. Das erste Gespräch kostet nichts.

Heute lernen, was morgen zählt.

Bei der Erstellung dieses Beitrags hat Künstliche Intelligenz unterstützt – etwa beim Titelbild (Midjourney), Recherche, Textoptimierung und Formulierungshilfe. Idee, Struktur und Argumentation stammen von Martin.