"The main challenge is data quality and availability"
Developing a company-specific AI solution can help businesses optimize their processes and improve efficiency. However, success depends on reliable data, clearly defined processes, and a well-structured data environment, explains Prof. Dr. Evangelos Xevelonakis, Managing Director of Swiss Valuenet and Professor of Business and Technology at HWZ University of Applied Sciences in Business Administration Zurich.
According to a survey published by the consulting firm EY in May 2026, 89% of Swiss employees use artificial intelligence daily. At the same time, a growing number of small and medium-sized enterprises (SMEs) are developing their own AI tools to address specific business needs with greater precision, speed and efficiency. This approach can help reduce the cost of certain processes and limit the need for expensive outsourcing. However, effective data management remains essential for developing reliable and secure in-house AI models, says Evangelos Xevelonakis, Program Director of the CAS Agile Business Engineering at HWZ University of Applied Sciences in Business Administration Zurich.
How can SMEs integrate AI into their business processes, and what should they expect?
Evangelos Xevelonakis: Most SMEs begin by using existing foundation models such as Gemini, Claude, DeepSeek, or models from Mistral. These can already support tasks such as drafting emails and reports, searching and summarizing documents, and creating marketing content. Companies can then enrich these models with company-specific knowledge using retrieval-augmented generation (RAG) and integrate them into existing workflows. With an initial investment of around CHF 10,000, an SME can already develop a useful AI solution or prototype for a clearly defined use case. Dedicated or customized AI models become an option when there is a clear business case and the company’s scale justifies the additional investment.
How can companies develop effective proprietary AI models?
Xevelonakis: The main challenge is the quality and availability of data. Before developing a proprietary AI solution, SMEs should ensure that the data it will rely on is accurate, consistent, and up to date. They then need to consolidate information that is often spread across multiple systems, such as Excel spreadsheets, accounting software, marketing databases, and customer feedback, into a structured and accessible data environment. This could include, for example, a data mart or centralized database designed for analytics. Open-source solutions of this type are now available at relatively modest cost.
Process automation often raises concerns about losing control over business operations. Is that a legitimate risk?
Xevelonakis: Yes. Automating a poorly designed process simply means automating errors more quickly and on a larger scale. Before introducing automation, companies should ensure that their processes are clearly defined, stable, and well controlled. This begins with understanding how a process actually operates, for example through process mapping or, where sufficient data is available, process mining. This allows inefficiencies, risks, and potential errors to be identified and addressed before automation is introduced.
You lead a continuing education program that combines business, data, and technology. How does it work?
Xevelonakis: The Certificate of Advanced Studies is designed for experienced professionals working in business, information technology, marketing, controlling, and operations. It provides participants with a solid understanding of how data analytics and AI can be applied to real business challenges. A key element is requirements engineering: participants learn how to identify business needs, translate them into clear requirements, and evaluate where data and AI can create measurable value. They also learn how to prepare and analyse data, develop AI-based solutions, and interpret the results. Case studies cover a wide range of applications, including banking and insurance, targeted marketing, predictive maintenance, and process mining. Several projects developed during the program have subsequently been implemented by participating companies.
Can you give an example of a recent application?
Xevelonakis: We recently carried out an analytics project for a company that received a large volume of unstructured customer feedback. The challenge was to extract useful insights from large amounts of free-text data that could not easily be analysed using conventional methods. We combined traditional machine learning techniques with natural language processing (NLP) to structure and analyse the data, and then used generative AI to interpret the results. This approach made the company’s business processes significantly more efficient.
We use a range of tools, including Python, low-code platforms, and generative AI models such as DeepSeek, models from Mistral, and ChatGPT to review code and identify potential errors. At this stage, we use only non-confidential or anonymized code with external AI services, which minimizes data protection risks.
Data protection remains a major issue in AI. How can companies ensure it?
Xevelonakis: Companies must comply with Swiss regulations on data transparency and confidentiality. One option is to use cloud service providers while ensuring contractually that relevant data remains hosted in Switzerland. However, each additional guarantee comes at a cost because it reduces the provider's available capacity.
Another option is to host pre-trained models locally, such as models from Mistral, Llama, or Apertus, and securely combine them with the company’s internal data (e.g. through a RAG architecture). The appropriate level of protection depends primarily on the sensitivity of the data and the sector involved. A bank or insurance company naturally has stricter confidentiality requirements than a small local retailer.
