Context: Scrum management includes knowledge-intensive activities such as backlog refinement, event preparation, and stakeholder alignment. Large language model (LLM) chat assistants may support these activities, but empirical evidence remains limited. Objective: This study investigates practitioners’ self-reported LLM use in Scrum management, perceived benefits and risks, and expectations for human–AI collaboration. Methods: We conducted a practitioner survey, and the results are based on a final analytic sample of 79 respondents who reported using LLM chat assistants in Scrum management. Closed-ended items were analyzed descriptively; open-ended responses underwent thematic coding. Results: In this sample, 48.1% reported daily or almost daily use, and 79.7% rated their LLM knowledge as at least qualified. Reported use centered on backlog refinement, event summarization, and Scrum-practice exploration. Perceived benefits included productivity (77.2%), reduced repetitive-task time (75.9%), and artifact quality (74.7%). Reported problems included almost-correct outputs (65.8%), hallucinations (54.4%), privacy concerns (50.6%), and content-validation difficulty (46.8%). Most respondents envisioned human-led or balanced human–AI collaboration (81.0%); 65.8% selected the no-role-replacement option. Conclusion: Respondents describe LLM chat assistants as complementary tools in Scrum management, particularly for text-intensive and analytical tasks, while emphasizing the continued need for human judgment, validation, and oversight.
Systems operating in safety-critical industries increasingly adopt Artificial Intelligence (AI). Satisfaction of system quality attributes in these industries is mandatory and rigorously verified. To influence system quality, a crucial development practice is the use of architectural tactics. However, selecting and balancing these attributes is often based on experience rather than on conscious decisions. This can unexpectedly impact system quality and might be detected too late in the testing phase. For development projects, it is desirable to identify and plan the selection of architectural tactics in a quality-driven manner. This article presents an approach that calculates a tactic’s relative influence between 24 linked quality attributes and 150 tactics as part of a semantic model that supports the quality-driven development of AI-based systems. Attribute selection and filtering create recommendations of relevant tactics for a specific design and development task. In an exploratory multi-case study, the suitability of the recommended tactics was analyzed using source code. Results show that the model can compensate missing quality attribute information input and can provide recommendations. Despite an average input of only 39.98% from 30 case-relevant quality attributes for the model, the average recommendation completeness reaches 49.26% for 20.67 known tactics in source code across cases, leading the model to suffer from gaps between the literature and cases. Yet, 28 new design tactics linked to 30 known quality attributes were identified. Our work thus lays the foundation to support quality-driven design and development by recommending architectural tactics based on quality requirements.
