AI in industry: 5 key takeaways from the KI im Mittelstand Conference

Key takeaways

1. Use cases, not business model transformation: and that is fine for now

For most industrial companies, AI remains a portfolio of specific use cases rather than the core of the business model — and that is still the right starting point. At TRUMPF, early machine learning work began as far back as 2014 and 2015, with image recognition applied to laser weld seams in metal sheets. It was a concrete quality assurance problem that digital methods solved more reliably than conventional approaches. The business model, however, remains selling machines and services. AI makes those machines easier to use and the surrounding services more efficient. It does not replace the machine.
Vorwerk tells a similar story. With both a hardware product (the Thermomix) and a growing digital ecosystem in Cookidoo, the company launched its first AI assistant for customers at the end of 2025. The assistant helps users identify recipes based on available ingredients or dietary restrictions, rolled out first to the community of independent advisors before being extended more broadly.
Most industrial companies follow a predictable three-stage progression: AI first lifts individual productivity in back-office functions, then begins to generate value in core operations and production, and eventually reshapes the business model itself. Most are currently somewhere between the first and second stage. Recognizing where you are in that journey, and not skipping ahead, is itself a strategic choice.

2. Customers do not ask for AI: they simply leave

Customer demand for AI rarely arrives as an explicit request. Companies that fail to innovate find that customers quietly migrate to competitors who have. The market does not demand AI by name; it demands better outcomes, faster responses, and simpler experiences.
In B2B manufacturing, this dynamic has played out clearly over the past decade. In 2016, there was still genuine debate about whether smartphones belonged on the factory floor. Today, between 50 and 80 percent of service cases at TRUMPF are handled digitally through a smartphone app. Customers photograph a potentially defective part, receive an automated diagnosis, identify the replacement component, and place an order, all within the same workflow. The AI is invisible. The improvement in experience is not.
The implication is straightforward: adoption is not a lagging indicator of success. It is a leading one. Measuring usage rates and active engagement early, well before an ROI can be demonstrated, is essential, because without broad adoption, there is no economic outcome to measure. If you track only ROI but not adoption, you will eventually get neither.

3. Data must be a foundation, not a bottleneck: run use cases in parallel

Poor data infrastructure is one of the most common obstacles to AI progress. Many companies carry legacy databases, disconnected CRM and ERP systems across sites and international locations, and data that has never been structured for machine consumption. Before AI can do meaningful work, this foundation often needs attention.

As Daniel Soujon put it during the discussion: “We often start by talking about AI, but then need to take two steps back and look at data infrastructure, governance, and quality. AI can be a catalyst here because concrete use cases make it much clearer where better data and better data infrastructure creates business value.”

The critical mistake is treating data as a precondition rather than a parallel workstream. Companies that insist on completing a comprehensive data transformation project before launching any AI use cases frequently find themselves stuck. The perfect data infrastructure never arrives, and neither does the AI.
OMMAX's own research confirms this pattern. In the AI Trend Survey 2026/27, data quality was cited as the leading reason AI initiatives fail (28% of respondents), followed closely by integration complexity (27%). More telling still: AI initiatives most often break down not at the ideation stage but during or after the pilot phase, with 35 percent failing during the pilot and 44 percent after it. Scaling, not starting, is where companies lose momentum.
The answer is to build individual data pipelines for specific use cases in parallel with broader data governance work. A functioning pipeline that enables one well-defined use case generates more organizational momentum than a multi-year data lake programme that never ships.

4. AI in hardware: a long-term architectural challenge

For companies embedding AI in physical machines, there is a structural tension that is easy to underestimate. Industrial machines have operational lifespans of 15 to 20 years. AI algorithms evolve on a timescale measured in months. That mismatch requires deliberate architectural thinking from the outset, not as a retrofit.

The consequences of getting this wrong are tangible. Early connected machines at TRUMPF encountered issues as seemingly mundane as log files filling hard drives and causing system slowdowns, requiring costly on-site service visits. The original system specifications did not anticipate such large volumes of data, but real-world usage proved otherwise. This illustrates why systems must be designed with enough flexibility to accommodate future demands, not just current requirements. As AI logic becomes more sophisticated, the challenge intensifies. The computational requirements of modern AI models are evolving faster than the semiconductor industry can track, which means control systems embedded in today's machines need to be designed with modularity in mind, so that compute components can be upgraded without replacing the machine itself.

The structural response is to push AI and business logic to the edge or to the cloud, keeping the software layer current even as physical hardware remains in place for its full intended lifespan. This preserves the value of the installed base while enabling continuous software-driven improvement over the machine's lifetime. 

5. Culture and leadership: the deciding factor in whether AI scales

Technology is rarely the limiting factor in AI adoption. Organization and culture are. OMMAX's AI Trend Survey found that AI ownership still sits predominantly in IT and engineering functions (48%), while only 7 percent of organizations have placed it in business units. That imbalance helps explain why many companies have AI capabilities that are technically sound but strategically disconnected from where business decisions actually get made.

Three organizational patterns consistently separate companies making progress from those accumulating pilots without scale.

First, AI cannot be fully delegated to IT. Business leaders, up to and including the board, need a working understanding of what AI can and cannot do, and need to be actively involved in identifying and prioritizing use cases. Executive masterclasses on prompting and agent-building are not a distraction from strategy. They are a prerequisite for it. You cannot make good decisions about what you do not understand.

Second, internal champions matter as much as central coordination. Vorwerk created a Head of AI role reporting directly to the executive team, a deliberate signal of organizational commitment. A four-day internal hackathon in February 2026 brought employees from every function together to build their own AI agents for everyday work tasks. Results were shared across the organization, and contributors were publicly credited. Pride of authorship turned out to be a genuine motivator for broader adoption.

Third, the framing of AI's purpose needs to evolve. Many organizations are currently stuck in a mode where AI conversations are dominated by cost reduction and headcount displacement. That framing will not sustain momentum or unlock AI's real potential. The competitive advantage from AI comes not from doing existing things more cheaply, but from doing things faster, building new capabilities, and creating value that was not previously possible. Productivity gain (41%) and internal operations improvement (35%) are today's dominant value drivers according to our OMMAX's AI Trend Survey 2026/27 data, but they are the floor, not the ceiling.