capinorium · Article · AI
What is missing is not ideas but translation: what really holds mid-sized companies back on AI
Proven methods for finding AI use cases have existed for a long time. What mid-sized companies lack is something else: the translation of a general AI capability into a concrete step in their own process or product. Those who start from the problem rather than the tool, check data early and seriously consider conventional software as an alternative can reach a robust decision in five manageable steps.
KfW reaches the same conclusion in its latest analysis: identifying the first practicable AI applications is "a central barrier" (KfW Research, Fokus No. 554, 29 July 2026). The same data show the pattern: 14 per cent of mid-sized companies use language generation, but only 5 per cent use automation and decision support. KfW calls these applications "very company-specific". Broadly applicable tools are taking hold. Where a solution has to be tailored to a company's own workflow, uptake thins out.
The method is known; applying it is the work
Guides from the Plattform Lernende Systeme (2021), KI-Transfer BW (2022) and Fraunhofer IIS (2025) are freely available and follow almost the same pattern. Knowledge of the method is therefore hardly the issue. The demanding part is applying it to one's own process. KI-Transfer BW describes the core of the problem: data specialists and domain experts often lack a common language about what good data and good results are.
Five steps from process to proof of feasibility
1. Understand the problem and the process. It starts with a bottleneck that has a measurable baseline: lead time, error rate, processing effort. Fraunhofer IIS recommends defining success criteria and metrics at this stage. Without a baseline, it is impossible to show later whether a pilot has improved anything. The Plattform Lernende Systeme also advises looking specifically at the core business, because use cases are easily overlooked there.
2. Assign the right AI capability. The AI use case canvas from Fraunhofer IAO makes the "relevant AI capability" a mandatory field alongside the problem statement – for example image recognition, language processing, forecasting or planning. An illustrative example: a manufacturer has so far classified customer enquiries by hand. The appropriate capability is then text classification, not a chatbot. This assignment forces you to state what the technology is actually supposed to do.
3. Check the data before building. Is there enough data, is it accessible, is it representative, and is the company allowed to use it? KI-Transfer BW quotes a practitioner whose experience is that the algorithm accounts for only about ten per cent and data acquisition and cleansing for the rest. Several guides therefore advise starting where historical data already exist.
4. Assess benefit and feasibility together. The Fraunhofer IAO canvas awards points for added value, data situation, know-how and time to a verified proof of feasibility. Another test question is important: are there already known solutions to comparable problems? A large benefit with low feasibility is a research project, not a quick start. Both can be right, but they require different budgets and expectations. As a guide, the Plattform Lernende Systeme suggests a return-on-investment target of under two years for mid-sized companies.
5. Prove it small and time-limited. A prototype should demonstrate feasibility and, according to Fraunhofer IIS, is "typically built in a few weeks". Acceptance criteria and a termination criterion belong in place beforehand. Just as important is the integration path: who operates the system, who monitors the results, what happens when the data change? In 2024 Gartner predicted that by the end of 2025 at least 30 per cent of generative AI projects would be abandoned after the proof of concept, partly because of unclear business value (Gartner, 29 July 2024). This is a global forecast, not a measurement among mid-sized companies, but it shows where projects typically fail.
When conventional software is the better choice
An honest result of the translation can be: no AI. According to Fraunhofer IIS, rule-based systems continue to work well when domain knowledge can be captured in stable rules. The People + AI Guidebook from Google PAIR names further cases in which AI is probably not better: when predictability matters, errors are very costly, full transparency is required or a quick, inexpensive solution is needed. AI plays to its strengths, by contrast, with variable and unstructured data, such as language in many variants, images, forecasts or changing patterns.
The same logic applies to the current wave of AI agents. Gartner notes that many use cases marketed today as "agentic" do not require an agentic implementation, and predicts that over 40 per cent of these projects will be cancelled by the end of 2027 (Gartner, 25 June 2025). A rule, a workflow automation or better search is often enough.
Knowledge exists; translators are scarce
KfW observes a moderate correlation: companies for which trade journals, trade fairs or consultancy are important sources of knowledge use AI somewhat more often. Whether one causes the other cannot be inferred from this. But it supports the assumption that contact with examples and expertise helps with the first step. According to KfW, science and specialist service providers are the least important sources of knowledge for mid-sized companies. Translation therefore has to start with a low threshold: with examples, a short assessment and a clear test question.
In practice, this means: start with a concrete process, not with a tool, and define what the first prototype is supposed to prove before you build it. If you are looking for a structured assessment with external methodology, a short conversation with capinori can be a sensible starting point.
Frequently asked questions about AI use cases in mid-sized companies
How do mid-sized companies find suitable AI use cases?
Mid-sized companies find suitable AI use cases most reliably when they start from a measurable process problem. They then assign a suitable AI capability, check the data, assess benefit and feasibility and test the idea in a small, time-limited proof of feasibility. Guides from Fraunhofer IIS, Plattform Lernende Systeme and KI-Transfer BW describe this approach consistently.
What does it mean to assign an AI capability to a process?
Assigning an AI capability means naming, for a specific process problem, which technical capability is needed: for example image recognition, text classification, forecasting, planning or content generation. The AI use case canvas from Fraunhofer IAO lists this capability as a mandatory field alongside the problem statement. This makes it clear whether a tool really fits the need.
When is conventional software better than AI?
Conventional, rule-based software is usually the better choice when domain knowledge can be captured in stable rules, results must be predictable and fully traceable, or errors would be very costly. AI is more worthwhile for variable, unstructured data such as language, images or forecasts. Criteria for this are given by Fraunhofer IIS and the Google PAIR Guidebook.
Why should the data check take place before the prototype?
The data check belongs before the prototype because data availability, data quality and usage rights determine whether an AI project is feasible. The KI-Transfer BW white paper quotes a practitioner according to whom data acquisition and cleansing account for most of the effort. If data are checked late, projects often fail because of problems that could have been identified earlier.
How long does an AI proof of feasibility take?
According to a guide for SMEs from Fraunhofer IIS, an AI prototype for a proof of feasibility is typically built in a few weeks. It is not yet integrated into productive operation. What matters are acceptance and termination criteria defined in advance and a planned integration path, so that the proof of feasibility leads to a robust decision on the next budget step.
Why do many AI projects fail after the proof of concept?
Many AI projects fail after the proof of concept because of unclear business value, rising costs, poor data quality or inadequate risk control. In 2024 Gartner predicted that by the end of 2025 at least 30 per cent of generative AI projects would be abandoned after this step (Gartner, 2024). This is a global forecast, not a measurement among German mid-sized companies.
Sources (8)
- KfW Research: Fokus Volkswirtschaft No. 554, 29 July 2026 (Dr Volker Zimmermann) (German). PDF
- Fraunhofer IIS: Schritt für Schritt zur eigenen KI-Lösung. Ein Leitfaden für KMU, April 2025 (German). PDF
- Fraunhofer IAO (KODIS): AI use case canvas, template (German). PDF
- Plattform Lernende Systeme: Künstliche Intelligenz im Mittelstand, June 2021 (German). PDF
- KI-Transfer BW: Wie gelingt KI im Mittelstand?, white paper, 2022 (German). PDF
- Google PAIR: People + AI Guidebook, chapter "User Needs + Defining Success". Website
- Gartner: press release of 29 July 2024 on generative AI. Website
- Gartner: press release of 25 June 2025 on agentic AI. Website
Rudy Winklhofer is Owner and Director R&D of capinori GmbH.
More articles by Rudy Winklhofer
