Although companies generate large amounts of data in production, this often remains unused. Data collection has so far been rudimentary or non-existent, meaning that valuable information is not systematically recorded. As a result, the potential for optimizing production processes is barely being exploited. Many companies rely on external cloud providers to evaluate their data, but the transfer of sensitive information to third-party providers is problematic for business and data protection reasons and also slows down efficient use.

In my presentation, I will provide insights into answers to these challenges from a past project. In the context of forming processes, we have developed a system for quality assurance in production that ensures both data quality and data sovereignty. Dedicated AI models are used not only to detect and repair outliers in the production data itself, but also to detect deterioration in production quality in order to enable predictive maintenance. All evaluations can be carried out locally in the Edge to save resources.

AI in software quality assurance and software testing is on everyone's lips. Tools, procedures and integrative processing are being promoted everywhere. But how useful are these approaches and ideas really for companies in the German market? Together with André Blunert from DB InfraGO, we will offer our perspective on this topic. We will discuss the governance goals of software quality assurance, its applicability for companies, the challenges of licence management and its usability for users. With insights from various projects, the presentation offers the opportunity to evaluate AI not from the perspective of a tool manufacturer or service provider, but from the perspective of a customer or user. Afterwards, we hope for a stimulating and, if you like, controversial discussion on the topic of ‘AI – hype or real added value’.

Artificial intelligence (AI) is fundamentally changing quality assurance - including in the area of digital accessibility. With the entry into force of the Barrierefreiheitsstärkungsgesetz (BFSG) on June 28, 2025, which transposes the EU Accessibility Directive (European Accessibility Act) into German law, companies are legally obliged for the first time to design digital products and services to be accessible. In future, websites, apps, vending machines and digital interfaces must be accessible and usable for all people, regardless of their limitations. But how can accessibility be tested efficiently and reliably? And what role can AI play in this? In this presentation, we will shed light on how AI-based methods can automatically check the accessibility of software - for example by analyzing contrasts, font sizes, semantic structure, alternative texts or heading hierarchies. At the same time, we show how AI can help to better understand testing processes and interpret results more easily - a contribution to self-help that enables less experienced testers in particular to test and evaluate accessibility independently. We show where the current limits lie - for example in the evaluation of context, logical sequence or emotional comprehensibility. We also discuss the added value: scalability, repeatability and the possibility of anchoring accessibility early on in the development process. The distinction between two fields of application is particularly relevant here:

• On the one hand, accessibility can be incorporated directly into the development of new digital applications and continuously checked using AI-supported tests.
• On the other hand, AI also offers the possibility of systematically identifying and removing barriers in existing websites and applications. Particularly in view of the legal requirements coming into force in mid-2025, this retrospective use is crucial in order to adapt existing systems in a timely manner and make them legally compliant.

The presentation is aimed at QA managers, developers, product managers and anyone involved in digital inclusion. Using concrete examples and lessons learned from projects, we will show how AI and QA can work together to implement accessibility in a way that is not only legally compliant but also user-centred – and why 2025 is the right time to talk about it.

How well does the AI we want to use for testing work, how reliable is it, and how do we actually measure its quality for testing? The answers to these questions determine how much effort we need to put into post-processing, for example in subsequent audits of the proposed tests, and how we should set up an AI-supported testing process. And this brings us to the fundamental question: How do we test an AI system, measure and evaluate its quality?

In this presentation, we do not want to address these questions for the highly complex generative networks, such as those behind Copilot, but rather focus on the significantly simpler feedforward networks that are relevant for autonomous driving, for example. However, our considerations are also particularly significant for generative networks. We will use an obstacle detection system consisting of a camera and AI as an example. Its purpose is to identify fallen trees or broken-down vehicles on the road. The target environment of this system is open and rich, and the possible situations in which we want to trust its classifications are too diverse for complex, rule-based systems to handle, which is why we use AI. Its disadvantage is that it works statistically, misjudgements are always to be expected, and a malfunction, a “bug” in the AI, cannot be easily corrected. And so the quality of an AI can only be assessed statistically: does it work more reliably than a human being?

It is generally known that training data for AI must be complete and balanced, otherwise its learning results will be distorted. And the same applies to test data, of course, if you don't want to preserve blind spots or misleading statistics about its quality. A very good understanding of the target environment must be developed in advance! And it is not enough to collect and test many examples of typical and critical situations, even in a combinatorial manner; for the representativeness of the training and testing, their probabilities of occurrence must also be taken into account. Only these provide the statistical framework on which a sound assessment of the completeness and balance of the data is based! We need a probabilistic model of the environment that captures as completely as possible all relevant situations, including their approximate probabilities of occurrence. We can then reference our statistical evaluation to such a model to assess the quality of the data and the system in the target environment in a transparent and comprehensible manner.

Creating such a model may seem very ambitious. We have developed various techniques and implemented tooling for this purpose, and in this presentation we will model a target environment with different routes, forests, peripheral buildings, and various weather and lighting conditions. This presentation will show that even more complex environments can be described in this way with reasonable effort. We will then show that such a probabilistic model not only allows us to make objective evaluations of training and test data, but also test end criteria: How many test cases do we need to verify specified statistical quality criteria with significance levels, and even to generate abstract test cases that conform to the statistical model, which can then be enriched to form concrete test cases, as we were able to show in a research project.

Are you a QA tester, developer, or just someone passionate about technology and automation? Join us for hands-onevent which will guide you through creating intelligent agents that automate UI testing like never before. No prior AIexperience is required — just bring your technical curiosity and enthusiasm!

In my presentation, I will give an honest account of my experiences from an ongoing project: How does my daily work with Selenium change when generative AI is added as support?

I will use concrete examples to show how AI-based tools help with writing, revising and understanding test code – for example, by suggesting test cases, refactorings or debugging. This is not about technological promises, but about real-world experience: where does AI save time? Where does it lead us astray? And how does this change the role of the test automation engineer?

The presentation is aimed at anyone interested in the use of AI in quality assurance, with a focus on practical insights, stumbling blocks and specific lessons learned from everyday project work.