Generative AI in Mid-Market Companies: Why the 78-Percent Figure Is Misleading
7 min read
According to the DIHK digitalization survey, 78 percent of companies use generative AI for text, images, or code. The number suggests a breakthrough. In reality, it’s a trap: usage is being confused with impact. Turning on a tool doesn’t change a process. It’s precisely this gap that will determine which mid-sized company will work more productively in 2027—and which will merely have a more expensive autocomplete.
Key Takeaways
- Usage isn’t value creation. The 78 percent figure measures who opens an AI tool—not who redesigns a workflow with it.
- The hurdles are legal and organizational. According to the DIHK, uncertainty and lack of integration are the brakes—not budget.
- The leverage lies in the handoff. AI only pays off when it’s clear who reviews the output, takes responsibility, and feeds it back into the process.
Related:Three AI defeats and their lessons / Stanford AI Index: Reliability as the bottleneck
What the 78 percent figure really means
The DIHK surveyed nearly 5,000 companies across all sectors for its digitalization study. The headline-grabbing result: 78 percent use generative AI, primarily for text generation, image creation, and coding. Over a third of users expect a significant boost to their productivity.
That’s the good news. The uncomfortable truth is in the fine print. “Use” mostly means someone in marketing drafts text in a chat window, or a developer gets a function suggested. That’s helpful—but it’s not a transformed process. It’s a faster tool on an individual desk. The number measures adoption, not integration.
Anyone who’s guided a transformation knows the difference. Having a tool in-house is easy. Embedding it so the output is reliably processed—that’s the real work. It’s at this handoff, where AI output is passed to the next person or system, that determines whether usage translates into impact.
Why the Gap Isn’t About Money
The German Chamber of Commerce and Industry (DIHK) clearly identifies the biggest hurdle: legal uncertainty. Lack of expertise, limited data access, and costs are cited less frequently than last year. This marks a notable shift. Just two years ago, the standard response to questions about digitalization barriers was: too expensive, no staff. Today, the question is what’s even allowed.
For SMEs, this is a different challenge than drafting a budget request. If you don’t know whether you’re permitted to feed customer data into an AI tool, more money won’t buy you certainty. What’s needed is a clear policy: which data goes into which system, with what approvals, and under whose responsibility. That’s governance work. It can’t be delegated to IT like buying a server—it belongs on the management team’s table.
Those who shy away from this decision end up with the worst of all outcomes: employees use AI anyway, just unofficially and without rules. Shadow AI isn’t the result of too few tools, but of too little clarity. The unspoken ban collides with the ungranted permission in the legal gray zone of everyday work.
Regulatory pressure adds to the challenge. The EU AI Act requires companies to classify and document the risks of their AI applications. If AI is only used informally on individual workstations, compliance simply isn’t possible. Scattered tool usage can’t be audited, but a defined process can. The legal uncertainty the DIHK identifies as a barrier won’t resolve itself by waiting—it requires the very framework that also drives economic benefits. Here, compliance and productivity point in the same direction for once.
What Fails
- AI as a standalone trick without process integration
- No clear rules on which data is permitted
- No one checks if the output is correct
What Works
- A process where AI has a defined role
- Clear data approvals with assigned responsibility
- A designated point of contact to validate results
How Usage Turns into Impact
The path is unspectacular. Perhaps that’s precisely why it gets skipped. Instead of rolling out yet another tool, it pays to take a close look at a single process that currently devours time. Proposal creation, complaint handling, reporting. The question to ask: at what point does AI make a specific step faster? And who then takes over the result?
An example from proposal creation makes the difference tangible. A sales team lets an AI suggest text modules. This saves ten minutes per proposal – a clear gain at the individual level. But impact only materializes when the draft flows automatically into the CRM, a second person checks the prices, and the approval is documented. Without this chain, the time saved remains a private trick that leaves the building the moment the employee resigns.
This second question is the crucial one. An AI-generated draft that no one checks and no one takes responsibility for is not progress – it’s a new risk. Only once the handover is clarified, meaning who approves, corrects, and feeds it in, does a tool become a process building block. That is precisely what the 78-percent figure does not measure. And that is exactly what separates the companies that will be measurably more productive in 2027 from those that are simply typing at greater expense.
Having an AI tool in the building is no achievement. The achievement begins where the result is handed over and someone stands behind it.
Taking an iterative approach does not mean being hesitant. It means starting with one process, setting a clear metric, and honestly checking four weeks later whether that number has moved. If it has, the next process follows. If it hasn’t, it was the wrong tool or the wrong spot. A cheap insight, as long as it comes early.
The reflex to roll out as broadly as possible is understandable, but it produces exactly the 78 percent that prove nothing. Depth beats breadth. A single process that has genuinely been restructured convinces the workforce more than ten licenses that no one brings into their daily routine. Anyone who wants to say next year that AI paid off doesn’t need a higher usage rate. They need a number from the day-to-day business that has demonstrably shifted. And a name behind it that takes responsibility for that shift.
Frequently Asked Questions
Isn’t 78 percent AI adoption a positive sign?
As a signal of uptake, yes. But the figure only shows that a tool is being used—not that a process has been redesigned to become more productive. It’s this second step that determines the economic benefit.
What does the DIHK identify as the biggest hurdle?
Legal uncertainties. Costs, lack of expertise, and limited data access are cited less frequently than last year. The bottleneck is governance, not money.
Where should a mid-sized company start?
With a single process that currently drains time, plus a clear metric and a designated person to sign off on the AI output. Only when that number improves should the next process follow.
What about shadow AI in the company?
It emerges from a lack of clarity, not a lack of tools. If no one defines which data is permitted, employees will continue using AI unofficially. A clear approval rule offers better protection than a silent ban.
Does this require a large budget?
No. Research shows the bottleneck isn’t money—it’s organization and accountability. A clearly anchored process beats an expensive tool with no handoff.
More from the MBF Media Network
Image source: AI-generated (May 2026), C2PA certificate embedded in image
