In December 1990, Jerry Sternin stepped off a plane in Hanoi with his wife Monique and their ten-year-old son to open an office of Save the Children in a country where about two thirds of children under five were malnourished, and where the authorities expected first results within half a year. An ordinary aid programme would have begun with food deliveries and family subsidies, whereas Sternin began with a scale. In four villages of Quang Xuong district in Thanh Hoa province, residents weighed about two thousand children under three and found that 64 percent of them were malnourished.
The most important discovery, however, was the list of children who contradicted that arithmetic. Among the poorest families lived children of normal weight, even though their mothers had no special resources, so poverty alone could not be the explanation. Sternin and the villagers looked at what those mothers did and found things so small that nobody had seen them as a solution: the children ate more often, and the daily rice was joined by freshwater shrimp and crabs and by sweet potato leaves, foods the neighbours considered inferior and left alone.
The twist was that Sternin lectured nobody. Mothers from neighbouring families cooked together following those practices and learned from women of the same village, who had the same fields and the same poverty. According to the Positive Deviance Collaborative, an organisation that develops this approach, within six months more than two thirds of the children in the villages began gaining weight, and after two years 85 percent were no longer classed as malnourished.
How many solutions like this lie in your company today, unnamed, because nobody has yet asked what that one person whose results look impossible actually does?
Sternin found an answer that had been working in the same village for years and that nobody had written down, and a customer service department looks exactly the same: someone resolves tickets faster and more calmly than the rest, and nobody can say what that person does differently. The thesis of this post is that an assistant based on artificial intelligence delivers its greatest return in customer service when it draws on the conversations of the best agents and reaches those who are only learning the job.
What researchers measured at a company handling thousands of chats
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of an assistant based on a language model at a large American Fortune 500 company that sells business software. A language model is a program that learned from enormous collections of text to predict which words fit a given conversation, and it composes replies on that basis. In the study the assistant, during a chat with a customer, suggested wording to the agent and pointed to internal documents about technical problems, while the company gave it to successive groups of employees from November 2020 to February 2021, which made it possible to compare people with and without the tool at the same time.
The results appeared in 2025 in the Quarterly Journal of Economics and cover 5,172 agents. The number of tickets resolved per hour rose by 15 percent on average, and less experienced, less skilled employees improved both the speed and the quality of their work. The most experienced gained little in speed, and the quality of their conversations declined slightly. In the earlier version of the study, published in 2023 by the research centre National Bureau of Economic Research, there are 5,179 agents, an average gain of 14 percent and a gain of 34 percent among beginners.
Why beginners gain more than veterans
The authors call their explanation suggestive: the assistant seems to pass the practices of better agents on to new employees, who then cover in a few months the road that veterans travelled over years. This matches the idea of tacit knowledge, meaning knowledge applied out of habit that an expert usually cannot describe because they do not regard it as knowledge. In Quang Xuong the neighbours had the same shrimp and the same leaves in front of their eyes and still did not reach for them, because they counted as lesser food, so the solution existed and waited for someone to name it and show it.
The researchers also noted better customer sentiment in conversations and lower staff turnover, which matters to a department that retrains a large share of its team every year: with lower turnover, the cost of bringing one person up to speed, meaning pay during the learning period and the mentor's time, has to be paid anew less often.
The mechanism has a human side as well, which I read in my own way. A beginner fears a question that would reveal ignorance, so they stay silent or improvise, and a suggestion appearing on the screen avoids that social cost because nobody sees that it was used. The quality of the suggestions depends on the material the assistant was built on: a tool fed average phrases will spread mediocrity, and a tool fed the conversations of the best agents will spread their way of conducting a conversation.
What the assistant cannot do, what it costs and what happens to the data
A model can state a false fact about a product with complete confidence, so the agent remains responsible for what is sent to the customer and must read it before approving it. The study concerns one company, written chats and a tool from 2020 and 2021, so in a department taking calls by phone or in another industry the result may be weaker, and newer models behave differently from that one. Experienced agents gained little and lost a little in quality, and given their standing in the team resistance comes easily: they see that the tool learns from their conversations and may regard it as a threat to their uniqueness. It is sensible to invite them into the role of authors of the patterns, with their name on the material they produce.
The licence cost is low compared with labour rates. According to the manufacturer's help page, a standard seat in the Claude Team plan costs 25 dollars a month per person billed monthly and 20 dollars billed annually (prices for customers in the United States, excluding taxes, so check the current offer for Poland). For a team of fifty agents that comes to 12 to 15 thousand dollars a year, while a productivity gain of 15 percent corresponds to the work of seven and a half full-time positions. This is an illustrative calculation based on a result from someone else's company, which is why your own measurement is needed here.
Customer conversations contain personal data governed by the General Data Protection Regulation. Before a pilot you therefore need to sign a data processing agreement with the vendor, check in the terms whether conversations are used to train models, and remove first names, order numbers and addresses from test materials.
First step for Monday morning: your own list of differences
Before you buy any licence, do what Sternin did in Quang Xuong: find the people who succeed under identical conditions and check what they do. It takes two people and one afternoon.
- Pick from the last quarter 20 conversations of agents with the highest number of tickets resolved per hour and the best customer ratings, and 20 average conversations on the same types of cases.
- Remove customer data from them.
- Read them in pairs and write down the differences: how the agent opens the conversation, what questions they ask in the first two minutes, when they apologise and how they close the case.
- Ask the authors of the best conversations why they write the way they do, because some differences will be explained at once and others will surprise even them.
- Turn the findings into a half-page brief for the assistant and for two weeks test it on a few new agents, comparing tickets resolved per hour with a group working the old way.
In the pilot, measure four things: tickets resolved per hour, the customer's rating after the conversation, the share of replies corrected by a supervisor, and the number of weeks after which a new agent reaches the team's level. The last measure is the most interesting in this approach, because it shows whether the knowledge of the best really reaches beginners.
The list of differences has value even if you never deploy the tool, because it can go into onboarding materials for new employees. And since you will be reading those conversations anyway: what will you do if it turns out that the difference between the best and an average person fits into a single sentence at the start of the conversation?