Does chat delegation actually save time on task completion?

The paper builds a web-based content management system with an LLM agent attached through the Model Context Protocol, so users can do CRUD tasks through the graphical interface, the conversational agent, or both. A between-subjects study (N=73) compared Traditional-Only, AI-First, and Hybrid modes across sixteen scenarios. The abstract reports that "AI-assisted interaction significantly reduced clicks, page navigations, and scrolling, indicating lower interaction effort." It then adds, "Surprisingly, these reductions did not translate into faster task completion, as task duration did not differ significantly." The claim worth keeping is the split itself: the effort measures and the time measure moved independently.

The excerpt gives no mechanism for why the effort savings failed to become time savings; the abstract breaks off mid-sentence at the null result. What it does offer is a caution about what interaction counts can show. In the discussion, chat use did not follow the risk ordering of the operations: DELETE, the most destructive, drew descriptively more chat interactions (μ=0.386) than UPDATE (μ=0.250). The authors suggest that "people may handle risk by using the assistant more carefully, not less often," so a raw count of interactions is a coarse instrument. They also report that chat use "was shaped more by who the user was than by what the task was." Delegation looks like a property of the person more than of the task.

This sits beside the efficiency claims already in the library without contradicting them. Can API-first agents outperform UI-based agent interaction? reports large time and workload gains, but it concerns how an agent reaches an application. This study concerns how a person reaches the agent, and the person's own turns are part of the clock. Do generated interfaces outperform text-based chat for most tasks? rests on preference and load rather than speed, so it does not conflict with a null on duration. The proactivity notes, for example Why can't conversational AI agents take the initiative?, look at the agent's side of delegation. This paper looks at the user's side, where the choice to delegate is already variable before the agent does anything.

The excerpt does not establish much beyond that. It does not say which modes were compared on duration, how large the click, navigation, and scroll reductions were, or how time varied by scenario or operation. It reports nothing on agent accuracy or error recovery, so it cannot say whether delegation time was spent on prompting, waiting, or checking. What follows at this strength is narrow. Effort metrics and completion time should be reported as separate outcomes, and a claim that an agent interface "reduces effort" should not be repeated as "saves time" without a duration measure behind it.