Can workplace AI risks emerge from interactions alone?

The paper's framework, built from a literature review on AI agents, "models risk as emerging from the properties of three core components (agents, goals, environments) and from the interactions between them and humans." Together with the humans they work alongside, these form what the authors call an agentic AI system. The discussion states why the decomposition matters: "interaction-driven risks can arise even when no single component is malfunctioning, a pattern our empirical findings confirm." The alternative it rejects is treating an AI system "as a black box."

The reasoning has two parts. First, the risk pathways are wider than agent-to-agent communication. The paper places itself as extending Hammond et al. 2025, which focused on agent–agent communication, by showing that "goal and environment mediation are equally important risk pathways." Second, human interaction is not one category. The framework splits it into five relationship types, of which the excerpt names agent–human, human–human and human–environment before an "etc." The motivation given in the introduction is practical: organizations red-team agents, run impact assessments and audit incidents, and existing taxonomies "focus on broad risks and do not capture the job-specific risks introduced by agents." The framework was then embedded in a structured prompt, applied to 2,078 O*NET tasks, and the resulting 8,356 scenarios were used to extend an existing taxonomy into 15 categories.

Against the neighbors, the closest match is How do agent security layers connect across the stack?, which reaches a similar conclusion from the security side: layers have to be considered together. The workplace paper slices the system along different lines, into agents, goals, environments and humans, rather than along stack layers. Who actually bears the risk when multi-agent workflows fail? separates requester, observer and affected party, and that split looks like a case of the human–human relationship types this framework counts separately. The excerpt itself does not discuss third parties. Does a single benchmark score actually predict agent readiness? argues for decomposing capability, and this paper makes the risk-side companion point: passing checks on each component would not show that the interactions are safe. That companion reading is the vault's, not the paper's. The sibling note Does AI augmentation protect workers from skill erosion? reads as one instance, since overreliance concerns the worker–agent relationship and does not require the agent to fail.

What the excerpt does not establish. It gives no count or example of interaction-driven scenarios, so the claim that the findings "confirm" the pattern cannot be checked here. It names only three of the five human relationship types and none of the 15 taxonomy categories. The taxonomy was extended to cover all of the scenarios, and those scenarios were generated with the framework, so coverage alone does not show the decomposition beats an undecomposed taxonomy. The excerpt reports no such comparison. What follows at this strength is a design stance for risk assessment: list the goals, environment and human relationships alongside the agent, because the excerpt's claim is that a component-by-component review can come back clean while a risk remains.