A Framework of User Experience Principles for Human-AI Agent Interaction in the Workplace
Citation. Paimann, K., Valarini, E., & Juhl, S. (2026). A framework of user experience principles for human-AI agent interaction in the workplace [Preprint]. arXiv. https://arxiv.org/abs/2607.19941
Summary. This preprint derives and prioritizes eight user-experience principles for workplace human-AI-agent interaction through participatory design, expert review, a qualitative meta-analysis, a paper-and-pencil survey, and interviews. In its enterprise-oriented sample, human control, reliability/safety/robustness, context awareness, transparency, data governance, ecosystem integration, collaborative partnership, and usability are presented as actionable design criteria rather than generic ethical aspirations. The framework translates these criteria into features such as override and stop controls, critical-decision confirmation, audit trails, uncertainty indication, role-based access control, and visibility into specialized agents.
Key claims
- In the 22-person survey, human control was the highest-prioritized principle: 65% selected it among their top three, 40% ranked it first, and 95% required human confirmation before an agent makes a critical business decision.
- Reliability, safety, and robustness were selected among the top three by 60% of survey participants; the associated criteria include accurate output, up-to-date resources, uncertainty or knowledge-limit signals, repair strategies, and audit trails.
- Transparency was selected among the top three by 40% of participants; 75% required agent actions and outputs to be easily verifiable, while 60% required clear explanations of how and why results were reached.
- The proposed framework treats context awareness, data privacy/governance, ecosystem integration, collaborative orchestration, and responsive intuitive interaction as distinct but overlapping design principles for enterprise agent workflows.
Evidence & limitations
- The multi-method exploratory framework combines a 21-person participatory workshop, five-expert review, qualitative meta-analysis of selected literature/frameworks, a 22-person prioritization survey, and 12 semi-structured interviews with AI-experienced software, data-science, and product professionals.
- It is a public arXiv v1 preprint, submitted 22 July 2026 and described as forthcoming in Mensch und Computer 2026; it is not yet evidence of peer-reviewed publication.
- The small, self-selected and AI-experienced samples, controlled research setting, forced ranking, conceptual overlap among principles, and lack of real-world longitudinal deployment tests limit generalizability and causal claims about adoption, trust, performance, or accountability.
- Public provenance was verified on the arXiv landing page; the arXiv API/PDF endpoint was rate-limited during preflight, so a Drive-to-arXiv binary hash match was not independently obtained.