AI Assistant Boosted Customer Support Productivity by 14%

The 14 Percent That Changes Everything

In 2022, a company called [redacted] gave 5,179 customer support agents access to a new tool. It was a generative AI assistant, the kind that could listen to a customer complaint and suggest a response in real time. The agents did not ask for this. Some of them did not want it. But within months, something strange happened: the agents who used it started resolving 14 percent more issues per hour. Not the veterans. Not the ones who had been doing this for years. The novices. The ones who were still figuring out how to say "I understand your frustration" without sounding like a robot. They got faster. They got better. And they stayed longer.

This is not a story about automation replacing workers. It is a story about something more interesting. It is about what happens when a machine can speak the tacit knowledge of the best humans and hand it to everyone else.

How Do You Measure a 14 Percent Productivity Gain?

Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, economists at MIT, Stanford, and the National Bureau of Economic Research, designed a study that is unusually clean for the messy world of real workplaces. They looked at a large U.S. software firm that rolled out a generative AI assistant to its customer support team in a staggered way. Some agents got the tool in early 2022. Others got it later. This timing difference allowed the researchers to compare agents who had the AI with those who did not, while controlling for everything else: seasonality, agent skill, call difficulty.

The AI assistant did not take over calls. It listened to the conversation between the customer and the agent, then suggested responses in a sidebar. The agent could accept, edit, or ignore them. The tool was built on a large language model, the same kind of technology behind ChatGPT but fine-tuned for customer support.

The result: a 14 percent increase in issues resolved per hour (Brynjolfsson, Li, and Raymond, 2023). That is not a small number. In customer support, where margins are thin and turnover is high, a 14 percent productivity boost can mean millions in savings.

But the number itself is not the story. The distribution of that number is.

Why Novices Got the Biggest Boost, and Veterans Barely Budged

The researchers broke down the productivity gains by agent skill level. They classified agents into quartiles based on their historical performance before the AI arrived. The bottom quartile, the least skilled agents, saw a 35 percent increase in issues resolved per hour. The top quartile, the most skilled agents, saw essentially zero.

Think about that. The best agents, the ones who already knew how to de-escalate a frustrated customer, how to navigate the knowledge base, how to type a response in 30 seconds, got almost nothing from the AI. They were already operating near the ceiling of what a human can do with a keyboard and a script.

The novices, the ones who still hesitated, who fumbled for the right phrase, who took too long to find the refund policy, got everything. The AI closed the gap between the worst and the best.

This is not how most workplace technologies work. Usually, new tools benefit the most skilled workers first. They are the ones who can figure out the new software, who can adapt quickly. But generative AI is different. It does not require skill to use. It gives skill. It takes what the best agents know instinctively and makes it explicit.

The authors describe this as "disseminating the potentially tacit knowledge of more able workers" (Brynjolfsson et al., 2023). In plain language: the AI learned from the best agents and then taught the worst ones.

The Experience Curve, Compressed

There is a concept in economics called the "experience curve." The more you do something, the faster and better you get. New workers start slow, then improve rapidly, then plateau. It usually takes months or years to move down that curve.

The AI compressed that curve. Novice agents with the tool performed as though they had months of additional experience. They did not just become faster. They became better at handling difficult calls. The researchers measured customer sentiment, requests for managerial intervention, and employee retention. All three improved.

Customers were happier. Managers got fewer escalations. And agents quit less often.

This last point matters. Customer support has one of the highest turnover rates in any industry. The work is repetitive, emotionally draining, and poorly paid. If an AI can make it less frustrating for new agents, if it can reduce the number of times they get yelled at by a customer because they did not know the right answer, then it might keep them in their jobs longer. The study found that AI assistance reduced employee turnover, though the authors note the effect was concentrated among lower skilled workers (Brynjolfsson et al., 2023).

What the AI Actually Did

The AI assistant did not write entire responses from scratch. It suggested them. The agent could choose to use the suggestion, edit it, or ignore it. The researchers tracked these choices. They found that agents accepted about 30 percent of the AI's suggestions on average. But the acceptance rate varied wildly by skill level. Low-skill agents accepted suggestions more often. High-skill agents accepted them less.

This is intuitive. If you already know the right thing to say, you do not need the AI to tell you. But if you are unsure, the AI gives you a template. It reduces cognitive load. It lets you focus on listening to the customer instead of searching for the right words.

The authors also found that the AI's suggestions became more useful over time. The model was fine-tuned on the conversations it observed. It learned which responses worked and which did not. The more agents used it, the better it got.

What This Research Does Not Prove

This is a single study at a single company. The results might not generalize to every industry or every type of work. Customer support is a relatively structured task. There are scripts, policies, and known solutions to common problems. A generative AI assistant might not work as well for a software engineer debugging a novel bug or a doctor diagnosing a rare disease.

The study also does not tell us what happens when the AI is wrong. The authors do not measure the error rate of the AI's suggestions. If the AI gives bad advice, and a novice agent follows it, the result could be a worse customer experience. The study found that customer sentiment improved on average, but that does not mean every interaction improved.

There is also a question of long-term skill development. If novices rely on the AI too much, do they ever learn the tacit knowledge themselves? Or do they become dependent on the tool, unable to function without it? The study only covers a few months. We do not know what happens after a year.

Finally, the researchers note that the AI's impact on productivity was zero for the most skilled workers. That raises a different question. If the best agents get nothing from the AI, is the tool worth the cost? The answer depends on how many novices you employ. If your workforce is mostly experienced, the AI might not pay for itself. If it is mostly new hires, the ROI could be enormous.

What This Actually Means

• ▸The biggest productivity gains from generative AI will come from helping the least skilled workers, not the most skilled. Companies should target AI tools at new hires and underperformers first, not their star employees.

• ▸Tacit knowledge can be extracted and distributed. The best workers know things they cannot easily explain. Generative AI can learn from their behavior and make that knowledge available to everyone. This changes how we think about training and onboarding.

• ▸Employee retention improves when the AI reduces frustration. If a tool makes a hard job easier, people are less likely to quit. That alone can save more money than the productivity gains.

• ▸The AI is not a replacement for human judgment. It is a suggestion engine. The agent still decides whether to use the suggestion. The human in the loop matters.

• ▸The value of AI depends on your workforce composition. If you employ mostly experts, the tool may not help much. If you employ mostly novices, it could be transformative. Companies should measure their own distribution of skill before investing.