Customer-support agents often have to diagnose a technical problem while also calming a frustrated customer. The system in this study listened to text chats and suggested possible replies or relevant internal documentation. The agent remained responsible for the conversation and could ignore the advice. ^[1]

The assistant combined a large language model with additional training on the company's earlier support conversations. Those examples included whether a problem was solved, how long the conversation took and whether the agent was considered a strong performer. In effect, the system tried to make useful habits from experienced workers available at the moment a colleague needed them. ^[1]

The researchers studied the staggered rollout of the assistant across 5,179 agents at a large business-software company and its contractors. Their data covered about three million chats. Because different groups received the tool at different times, the researchers compared changes for agents with access against agents who had not yet received it, while accounting for factors such as experience, location and calendar month. ^[1]

The average result came from shorter chats, more chats handled at once and a small rise in the share of problems resolved. Agents with two months of experience and AI assistance performed about as well as untreated agents with more than six months of experience. The study also found signs of better customer sentiment and fewer requests to speak to a manager. ^[1]

The strongest pattern was not simply that the software made work faster. It narrowed a performance gap. Less-experienced agents changed how they communicated and moved toward patterns used by stronger agents. Workers who followed the suggestions more closely also improved more. ^[1]

The most skilled agents did not receive the same productivity gain, and the researchers found some evidence that AI assistance could reduce the quality of their conversations. That matters when deciding where an assistant helps: advice learned from common past cases may be most useful to someone still building experience, and less useful when an expert is handling an unusual situation. ^[1]

This was a real deployment, but it took place in one company's text-based support operation, where most agents handled similar technical questions. The rollout was staggered rather than a company-wide randomized trial. The results may not carry over to other jobs, organizations or newer AI systems. ^[1]

The study could not observe wages, total demand for workers or changes in who the company hired. Higher productivity could lead a company to serve more customers, change job responsibilities or need fewer people. The data cannot choose among those possibilities. The authors also raise an unresolved question: experienced workers supplied many of the examples that trained the system, even though they received relatively little of the measured benefit. ^[1]