AI Navigator #13: Artificial intelligence overtakes the daily
When AI models enter sprint planning and daily stand-up, it is not only tools that change, but also team roles and team spirit in practice.
(Image: CoreDESIGN/Shutterstock)
- Dr. Konstantin Hopf
- Dr. Karoline Glaser
- Daniel Dorsch
Welcome to the thirteenth edition of the DOAG AI Community's AI Navigator column!
Generative AI tools help individuals to write texts, program or create images. But they are also changing the way teams work together. In future, intelligent agents will be able to take over entire areas of responsibility in teams and assign tasks according to expertise, derive the next work steps from discussions or generate software tests.
This puts proven methods of project management and software development to the test. AI-based systems differ fundamentally from previous software: instead of clearly defined rules in the program code, probabilistic models sometimes make uncertain predictions. The rules of the models can change over time if the ML processes uncover new correlations in the data.
The results of the systems are already better than those of humans in some cases. This makes the use of AI in software development"outsourcing deluxe". One challenge here is that AI-based systems are complex and in some cases cannot be explained using our human approaches.
(Image:Â DOAG)
The AI Navigator conference on November 19 and 20 in Nuremberg will showcase the application of AI in the areas of IT, business and society. The event, organized by DOAG, heise conferences and de'ge'pol, offers a good 100 sessions. Tickets are available at the discounted early bird price of 990 euros (plus VAT) until October 1.
All of this has an impact on the way teams work. Even agile working approaches such as Scrum are coming under pressure – not because the methods are outdated, but because the teams have to adapt them to the new conditions.
To better understand this change, we brought together experts from the fields of agility and data science to discuss the impact of AI on agility. We present the initial results of this exchange of experiences in this issue of the AI column.
The workshop "Is AI destroying my Scrum? How artificial intelligence is changing agility" took place in Nuremberg in November 2024 at the QualityMinds premises. After two keynote speeches on AI systems and agile methods, the 30 or so participants from the agilist and data scientist communities split into three groups and discussed using the 1-2-4-all method, in which the number of participants in the discussion gradually increases. Our thanks go to Manuel Illi and Ursula Maichen, who planned, moderated and conducted this workshop with us.
Do AI and agile methods go together?
In recent years, agile methods have been seen as a miracle cure for solving complex problems and keeping pace with the rapid changes in the market in terms of customer requirements and technology. Methods such as Scrum operationalize agile principles in teamwork and are intended to enable companies to anticipate changes and react to them in a targeted manner. The remainder of this article will focus on Scrum, or more specifically the agile events and roles within the Scrum framework.
Videos by heise
In theory, it seems that agile methods and AI work wonderfully together. Both AI projects and agile methods require hypothesis-driven, iterative-incremental work and enable a culture of error. Despite this, or perhaps precisely because of this, a lively discussion broke out during our workshop between the agilists and the more explorative data scientists, which developed in two directions.
Integrating AI into agile working methods
The first direction revolved around the question of how AI can help to optimize agile product development. In software development, for example, not every activity is complex, but many activities are easy to assess and plan. In practice, this means that sprint goals are often equated with specific features or tasks, meaning that the focus is often on increasing efficiency: more software in less time. In the AI Navigator column, Semjon Mössinger and Bastian Weinlich described five stages of AI use in software development. For almost every activity in software development, there are already tools to (partially) automate the process. The following table lists some examples of AI applications in the software development process.
| Softwareentwicklungsprozess | KI-UnterstĂĽtzung (Beispiele) |
| Anforderungsanalyse | Generieren von Ideen und Mockups, Transkription von Besprechungen |
| Planung und Analyse | Backlog-Generierung und Analyse |
| Design und Architektur | Generierung von Architekturmodellen, Simulationen und Gap-Analysen |
| Implementierung | Copiloten und Vibe-Coding |
| CI und Testen | Test-Case-Generierung, Testautomation, Generierung von Infrastructure as Code |
| Review und Feedback | KI-Agenten für Feedback (intern) |
| Deployment | Agenten für Infrastructure as Code |
| Monitoring und Wartung | AIOps-Werkzeuge, Incident Prediction, Chatbots fĂĽr First-Level-Support |
| Nutzerfeedback und Auswertung | KI-Agenten fĂĽr Feedback (extern) |
| Kontinuierliche Verbesserung | Empfehlungssysteme |
Developing products with an AI component
The second direction of the discussion revolved around the question of how teams need to change agile methods when developing data-driven, AI-based systems instead of traditional software. These systems process large, often unstructured amounts of data, the quality of which is crucial for the performance of the subsequent application. Development is driven by hypotheses, which the team tests and adapts, and is only rarely geared towards specific requirements. Results are not produced linearly, but through exploration and experimentation. Validating the results is particularly complex, as an AI model can often only be meaningfully tested as a whole and isolated interim results are of little significance. In addition, AI applications need to be regularly retrained and adapted to new data in order to remain up-to-date and efficient.
Agile methods such as Scrum generally rely on clearly structured rhythms: Sprint planning, review, retrospective. They aim to deliver visible results early on that can be iteratively improved. In theory, agile methods are therefore well suited to the development of products that have at least an AI component. In particular, the rapid iterations, the promotion of cross-functional collaboration and the high adaptability to changing requirements through frequent feedback loops seem to create ideal conditions for AI development. This allows teams to react flexibly to new findings and continuously improve the development process.
In practice, however, agile implementations reach their limits in AI and data science projects. Data scientists complain about the excessive effort involved in maintaining or defining issues or feel pressured to show an interim result of the exploratory work at the end of the sprint. Sprint goals are also difficult to define as a precise result. The role of the product owner loses contour, as a continuous gain in knowledge significantly changes the product. Some teams therefore improvise and can lose their methodological clarity, which can lead to an ineffective use of methods.
