Does supervising AI free up worker attention for other tasks?

Beyond the split between technical and emotional escalations, the same Alibaba field experiment finds a spillover effect on the chats workers never handed to the AI at all. The paper states it directly: it documents "a positive spillover effect on AI-ineligible chats, as treated workers adapted their multitasking workflow to devote greater attention to these chats." The consequence shows up in the headline quality metric: customer ratings "diverges sharply across chat types, declining in AI-eligible chats but improving in AI-ineligible chats" — even though "AI-eligible chats accounted for less than 10 percent of total chat volume," meaning most of a treated worker's load was still fully human-handled, non-AI work.

The mechanism the paper points to is reallocation of attention under an otherwise unchanged workload, not a shift in incentives. Compensation for AI-ineligible chats stayed "based on chat volume... multiplied by a quality-based bonus coefficient," the pre-existing scheme, and "AI-eligible and AI-ineligible chats also contributed equally to workload assessments," so nothing in the pay or workload structure pushed workers toward the human-only queue. What changed was multitasking: freed from fully carrying the small AI-eligible slice themselves, treated workers appear to have redirected that attentional slack into the larger AI-ineligible slice they continued to own end-to-end.

This pairs directly with When human workers take over from AI, does it actually fix the problem? from the same study: together, the two findings show that deploying agentic AI does not uniformly help or hurt a worker's output — it redistributes quality, down on the AI-supervised slice when escalations are emotional, up on the human-only slice via spillover attention. It also complicates the autonomy-share framing in Why do AI agents fail at workplace social interaction?: here the automatable share is smaller still (under 10 percent of volume), and the finding is less about what fraction AI can do alone than about what a human does with the capacity that automating a small share frees up. It also sits against Do humans and AI divide labor differently in knowledge work?: rather than settling into one stable collaboration mode, Alibaba's treated workers appear to shift effort fluidly between an AI-supervised queue and a human-only queue within the same shift.

The excerpt does not give an effect size or statistical significance for the AI-ineligible rating improvement, and the experiment ran for a 17-day treatment window on one platform, so whether the attention-reallocation effect persists, grows, or fades as the AI-eligible share expands after full rollout (which began the day after the experiment ended) is not established here. The implication, held to that scope: measuring an agentic AI deployment only by what happens to the chats the AI touches will understate, or mismeasure, its full effect on a worker's output, since freed attention does not disappear — it moves to adjacent, un-automated work.