Does AI assistance help less experienced workers most?

The study measures a 15% average productivity gain from access to a generative AI conversational assistant, counted as issues resolved per hour, and finds that the gain is uneven. The abstract states the split directly: "less experienced and lower-skilled workers improve both the speed and quality of their output while the most experienced and highest-skilled workers see small gains in speed and small declines in quality." The data cover 5,172 agents in a staggered rollout at a Fortune 500 software firm (the introduction says 5,000). The abstract also reports that AI assistance "facilitates worker learning and improves English fluency, particularly among international agents," that the gains were largest for "relatively rare problems," and that customers became more polite and less likely to ask for a manager.

The mechanism the excerpt gives concerns what the tool does and what it learns from. The assistant "monitors customer chats and provides agents with realtime suggestions for how to respond," and agents "remain responsible for the conversation and are free to ignore or edit the AI's suggestions." The introduction argues that machine learning infers instructions from examples, and that a model trained on data from human workers "can implicitly learn what specific behaviors and characteristics set high-performing workers apart." On this account the tool passes the techniques of stronger agents to weaker ones, which is why the gains skew toward the less skilled. The excerpt presents this as the expected mechanism. It does not show the analysis that would test it.

Against the neighbors, this paper is the source of the 15% figure that When does AI actually boost worker productivity? reads as skill application. That reading holds that AI speeds up work the worker already understands. The abstract's learning claim complicates it, because this paper says AI assistance also improved learning on the job. The two can only be partly squared. The excerpt does not say how learning was measured, and live suggestions in chats are a different setting from the developers learning a new library in the other note, so this is a qualification of the skill-application reading rather than a refutation. The setting also differs from Can AI narrow the education performance gap?. That study is a randomized experiment with 1,174 adults. This one is a live workplace, which gives it more external weight, and it points the same way: the least skilled gain most. Set against Does generative AI prevent juniors from getting entry-level work?, the two notes pull in different directions about who benefits. Short-run gains for the least experienced do not show whether the struggle that builds expertise is being lost.

What the excerpt does not establish is substantial. It describes one firm's customer-support chat, so it says little about knowledge work in general. It stops before the results, and its limitations passage is cut off mid-sentence, so the paper's own caveats and its quality measure are not visible. The learning and English-fluency findings appear only in the abstract, and the excerpt offers no follow-up measure of whether any skill persists after the assistant is removed. The broader suggestion that generative AI "may be capable of capturing and disseminating the behaviors of the most productive agents" is the authors' conjecture, not a result shown here. The defensible implication is narrow. In this deployment, AI assistance raised short-run output most for the least experienced agents, and whether that carries over once the tool is gone is still open.