One of the largest real-world studies of generative AI at work found that the biggest productivity gains did not go to the best employees. They went to people with less experience and weaker prior performance. The result suggests AI may be especially powerful when it helps workers access knowledge they have not had time to build themselves.
Most discussions about AI productivity begin with an average.
How much faster does AI make a worker? How many more tasks can someone complete? How much time does a company save?
A large workplace study shows why that average can hide the most important part of the story.
Researchers examining the introduction of a generative AI assistant across thousands of customer-support workers found that productivity increased about 15% overall, but rose roughly 30% among less experienced and lower-skilled employees. The peer-reviewed study of 5,172 customer-support agents, published in The Quarterly Journal of Economics in 2025, found little productivity improvement among the most experienced and highest-performing workers.
The difference was large enough to change how quickly new workers appeared to gain experience. Employees with only two months of tenure who used the AI assistant performed about as well as workers with more than six months of experience who did not have access to it.
An earlier NBER version of the research, based on 5,179 agents, reported a 14% average productivity gain and a 34% increase among novice and lower-skilled workers. The final peer-reviewed paper revised the sample and estimates slightly, which is why the current published figures are 15% overall and approximately 30% for the workers who benefited most.
The broader finding remained the same.
AI did not lift everyone equally. It disproportionately helped the people who had the most experience left to acquire.
AI appears to compress part of the experience curve
Experienced workers usually have an advantage that is difficult to reproduce quickly.
They have seen more problems.
A senior customer-support employee knows which questions to ask because they have handled similar cases before. They recognize patterns in vague customer descriptions, understand which solutions tend to work and know how to communicate when a customer is frustrated.
A new employee cannot download that experience during onboarding.
They have to accumulate it through repeated exposure to real situations.
The AI system in this study appears to have shortened part of that process.
It monitored customer conversations and provided agents with real-time suggestions about how to respond. Workers remained responsible for the conversation and could follow, edit or ignore the recommendations.
The important feature was not simply that the system could generate text. Its recommendations reflected patterns learned from previous customer-service interactions, allowing less experienced employees to access behavior that resembled the practices of stronger agents.
Researchers found that newer and lower-performing workers changed their communication patterns more substantially after receiving AI assistance than experienced workers did.
That provides a plausible explanation for the uneven productivity gains.
Senior employees already possessed much of the knowledge the system was offering. Newer employees did not.
AI had more to give them.
The productivity gain was not simply faster typing
The study measured productivity through the number of customer problems successfully resolved per hour.
The improvement came from several places.
Workers completed individual conversations more quickly, handled more conversations per hour and achieved a small improvement in the share of customer problems they successfully resolved.
That combination matters because speed alone can be a misleading measure of productivity.
A customer-support employee who ends conversations quickly but leaves customers with unresolved problems is not necessarily more productive.
In this case, workers using the tool were able to handle more customer interactions without an equivalent deterioration in successful resolution.
The benefits also extended beyond output volume. The research found improvements in customer sentiment, fewer requests to escalate conversations to managers and evidence of lower employee turnover among workers with access to the assistant.
The technology was therefore changing more than how quickly employees could type responses.
It was helping some workers navigate the interaction itself.
A beginner can benefit more because an expert already knows the answer
The weak effect among top performers is one of the most useful findings in the research.
The most experienced workers saw relatively little improvement in productivity. The final study also found evidence of a small decline in conversation quality among some of the highest-skilled workers when using AI assistance.
That result should not be interpreted as evidence that experts generally become worse when they use AI.
It highlights a basic limitation of assistance systems.
A recommendation is valuable only when it contains something the worker does not already know.
If an experienced employee can diagnose a customer’s problem faster and more accurately than the model’s suggestion, following that suggestion can add little value. In some situations it can even pull the worker away from a better answer.
The same technology can therefore function very differently depending on who receives it.
For a beginner, AI can supply missing knowledge.
For an expert, it may simply repeat familiar information.
In the worst case, it can introduce a weaker alternative to the expert’s own judgment.
That distinction is important for companies measuring AI adoption. An organization should not expect the same return from giving the same tool to every employee.
The largest opportunity may exist precisely where the organization’s performance gaps are largest.
Other experiments have found the same leveling effect
The customer-support study is unusually valuable because it examined AI inside a real workplace at substantial scale.
It is not the only research to find larger gains among weaker performers.
A randomized experiment involving 453 college-educated professionals asked participants to complete realistic professional writing tasks. Access to ChatGPT reduced the time required to complete the work by 40% while increasing assessed quality by 18%.
The researchers also found that the technology reduced performance inequality. Participants who began with weaker writing performance benefited more, narrowing some of the difference between stronger and weaker workers.
A more complex field experiment involving 758 management consultants produced a similar result.
The peer-reviewed research on consultants working with GPT-4 found that, on tasks within the technology’s capabilities, people using AI completed 12.2% more tasks and worked 25.1% faster while producing substantially higher-quality results.
The improvement was not evenly distributed. Consultants who had performed below the median on an initial assessment improved their performance by 43%, compared with 17% among above-median performers.
These studies involve different occupations, technologies and productivity measures, so their percentages should not be directly compared.
The repeated pattern is more important than the precise numbers.
When AI performs a task well, people starting with weaker performance often have more room to benefit.
This is different from the way many earlier technologies rewarded skill
New workplace technologies have not always reduced performance gaps.
Earlier waves of computing often complemented highly educated or highly skilled employees particularly well. Workers who already possessed valuable expertise could use technology to become even more productive, contributing to what economists often describe as skill-biased technological change.
Generative AI can behave differently because it is able to provide something resembling expertise directly inside the workflow.
A spreadsheet does not usually tell an inexperienced analyst which assumption is unreasonable.
Traditional software does not tell a new employee how a strong colleague would phrase a difficult response.
A language model can sometimes provide that guidance.
This makes generative AI unusual as a productivity technology. It can automate parts of a task, but it can also provide a less experienced worker with examples, explanations and recommendations at the moment they are needed.
The customer-support study provides real-world evidence that this second function can be economically important.
The researchers’ interpretation is that AI was helping distribute some of the practices embodied by better workers to employees who had not yet learned them.
That is closer to knowledge transfer than simple automation.
AI may be useful because human coaching is expensive and irregular
Training inside most organizations is constrained by time.
A manager can review an employee’s work, explain mistakes and demonstrate better approaches, but continuous one-to-one coaching is expensive.
Feedback therefore tends to arrive periodically.
An employee completes work. Someone reviews it. A meeting happens later. The employee tries to remember the lesson when a similar situation appears again.
AI can operate at a different frequency.
In the customer-support setting, the assistant could provide suggestions during individual conversations. New employees were able to receive guidance repeatedly while the problem was happening rather than waiting for a manager to review a sample of interactions later.
That does not make AI a replacement for a manager.
A manager understands broader context, employee development, organizational priorities and exceptions that may not exist in the system’s training data.
The difference is scale.
A human manager cannot personally sit beside every employee during every customer interaction. Software can.
For relatively structured work, that creates the possibility of providing frequent guidance that would previously have been economically difficult.
The value may be greatest for workers early in their learning curve, precisely because they encounter more situations where guidance is useful.
The evidence suggests workers learned rather than merely copied
One of the strongest parts of the research concerns what happened when the AI was temporarily unavailable.
If employees were simply becoming dependent on automated suggestions, their performance should have returned toward its previous level when access disappeared.
The researchers found something different.
During software outages, workers who had previously been exposed to the AI assistant continued to perform better than their own pre-AI baseline. The effect was strongest among people who had received more exposure to the system and had followed its recommendations more closely.
The analysis of performance during AI outages provides suggestive evidence that some of the gains persisted after the immediate assistance disappeared.
That distinction matters enormously for workforce development.
An AI system that improves performance only while it is present is a productivity tool.
A system that also leaves the worker more capable afterward begins to function as a learning tool.
The study cannot tell us how durable that learning remains over many years, but it provides evidence that employees were absorbing at least some of what the system was showing them.
International workers experienced another kind of benefit
The research also found evidence that AI assistance improved English-language fluency, particularly among international employees.
That result illustrates another way AI can reduce differences between workers.
A customer-support agent may understand the technical problem perfectly well while communicating less fluently in the language used by customers. In a text-based support environment, that communication gap can affect how professional or empathetic the employee appears.
Generative AI can assist with expression without changing the employee’s underlying technical knowledge.
That means part of the productivity effect can come from removing barriers around a worker’s expertise rather than supplying the expertise itself.
The distinction has implications well beyond customer service.
International freelancers, researchers, engineers and other professionals can possess strong technical or domain knowledge while competing in markets where communication quality influences how their work is perceived.
AI can reduce part of that disadvantage.
It does not make language ability irrelevant, but it can reduce the degree to which language fluency determines whether someone can successfully apply another skill.
The greatest gains appeared where workers had limited experience but AI had enough data
The study also examined the types of customer problems where assistance created the largest productivity improvements.
The strongest benefits did not necessarily occur on the most common problems.
Workers already had considerable experience with those.
The gains were larger for moderately uncommon issues where human employees had less personal experience but the AI system still had enough historical data to offer useful recommendations.
That is an important condition.
AI can help fill an experience gap only if the system itself has relevant knowledge.
For extremely rare situations, there may not be enough examples for the model to provide reliable guidance. For extremely common situations, the worker may already know what to do.
The strongest opportunity exists between those extremes.
This offers companies a more useful way to think about AI deployment than simply asking which tasks consume the most employee time.
AI may be particularly valuable where workers repeatedly encounter situations that are uncommon enough to be difficult for individuals to master but common enough for the organization to have accumulated substantial historical knowledge.
In that setting, AI can potentially turn organizational experience into assistance available to each worker.
There is a major difference between transferring expertise and creating expertise
The finding can easily be pushed too far.
If AI helps a novice behave more like an experienced employee, it is tempting to conclude that experience itself has become less important.
The evidence does not establish that.
The assistant performed well because the work took place inside a relatively stable customer-support environment with recurring problems and substantial historical data.
Expertise often becomes most valuable when the situation is unusual, ambiguous or poorly represented in past examples.
An experienced employee may recognize that the current problem is different from previous cases. They may understand an unstated business constraint, know that a standard procedure should not be followed or recognize that the AI’s confident recommendation is wrong.
Those capabilities are harder to package into a suggestion system.
The distinction becomes especially important outside tightly structured domains.
In the consulting experiment, AI produced impressive gains when tasks fell within its capability frontier. When researchers deliberately gave consultants a task outside that frontier, participants using AI were 19 percentage points less likely to reach the correct conclusion than people completing the task without AI.
The technology made the consultants more capable where it was strong and more vulnerable where its output was misleading.
That is why narrowing a performance gap should not be confused with eliminating the value of expertise.
A beginner can become much more productive while still lacking the judgment required to know when the assistant should be ignored.
Companies may need to rethink where training happens
Traditional workplace development often separates learning from production.
Employees attend training, read documentation or complete a course before returning to the work itself.
Generative AI creates the possibility of moving more learning into the task.
A junior employee can ask for an explanation while solving the problem. They can compare their approach with an AI recommendation, request examples and receive feedback while the context is still fresh.
The customer-support study suggests this can accelerate practical learning when the system has reliable knowledge to share.
That has obvious implications for onboarding.
Organizations may be able to bring some employees to useful productivity faster than before. Experienced employees could spend less time answering routine questions and more time teaching the exceptions, judgment calls and contextual knowledge that AI handles poorly.
The goal would not be to remove mentorship.
It would be to use human mentorship where human expertise is most valuable.
This could also change how companies evaluate beginners. If AI dramatically improves routine performance, output volume alone becomes a weaker signal of whether the employee is actually developing independent expertise.
Managers may need to pay more attention to whether workers understand the recommendations they accept.
The productivity advantage can create a learning risk too
There is an uncomfortable possibility inside every AI-assisted training model.
The same tool that helps a beginner learn can also help a beginner avoid learning.
A worker can accept a correct recommendation because they understand why it is correct.
They can also accept the same recommendation because the system presented it confidently.
Those behaviors look identical when the output is right.
They become very different when the system is wrong.
This is why evidence of learning in the customer-support study is important, but it should not be generalized into an assumption that AI use always builds skill.
How a system is designed matters.
Whether employees receive explanations matters.
Whether they are encouraged to challenge recommendations matters.
Whether performance is ever evaluated without assistance matters.
A company that uses AI only to maximize immediate output could produce employees who are highly effective when the system works but less capable of operating independently when it does not.
The better objective is not simply making beginners look like experts.
It is helping them become experts faster.
This matters for freelancers as much as employees
Independent professionals do not usually have access to a corporate training department or a manager whose job includes developing their skills.
That makes AI potentially valuable as an on-demand learning layer.
A freelancer entering a new but adjacent area can use AI to understand terminology, review unfamiliar concepts, test approaches and accelerate parts of the research process.
That can shorten the distance between existing capability and a new marketable skill.
The opportunity is particularly relevant because Flexable’s research has already examined how 87% of skilled freelancers report learning new skills within six months. Independent professionals carry much of the responsibility for keeping their skills current themselves.
The customer-support findings suggest AI could lower part of that learning cost.
They do not show that AI can replace real project experience.
For a freelancer, using an assistant to understand an unfamiliar technology is very different from having successfully implemented that technology for a client.
AI can make knowledge easier to access.
Clients still pay for someone capable of applying that knowledge when the consequences are real.
That distinction becomes even more important as Flexable’s research on jobs fixing AI-generated work shows growing demand for professionals who can recognize when AI output needs correction.
The same technology that helps beginners perform more advanced work can therefore increase the importance of knowing when its advice is unreliable.
Productivity equality does not automatically become wage equality
The study also raises a larger economic question.
If AI helps weaker workers approach the productivity of stronger workers, what happens to the economic value of experience?
The research does not answer that question.
It measures workplace performance, not long-term wages or career progression.
Several outcomes are possible.
Companies could share productivity gains with workers through higher wages. Employers could hire less experienced people for work that once required greater experience. High performers could move toward more complex tasks while routine work becomes easier for everyone. The market value of some intermediate levels of expertise could also fall if AI makes those capabilities widely available.
Which outcome dominates will vary by occupation.
The important point is that reducing a productivity gap changes more than how much work gets completed.
It can change which workers are qualified for which jobs.
This is particularly relevant as other labor-market evidence shows employers demanding more sophisticated capabilities from people earlier in their careers.
If AI can accelerate some forms of skill development, that could help workers meet those rising expectations.
If employers simply raise expectations without creating better learning systems, the opposite could happen.
What the study does and does not prove
The primary evidence comes from one Fortune 500 software company and one occupation: customer support.
The final published analysis uses data from 5,172 agents during the staggered introduction of a GPT-based conversational assistant.
The system was not the same as giving employees unrestricted access to a general-purpose consumer chatbot. It was designed for customer-support work and provided recommendations inside a specific professional environment.
Productivity was measured primarily through successfully resolved customer issues per hour.
Because access was introduced at different times, researchers could compare workers with access to employees who had not yet received the system while controlling for worker experience and other factors.
That design provides unusually strong real-world evidence, but it does not justify assuming a 30% productivity increase for every beginner using AI.
Customer support contains repeated problems, observable outcomes and large quantities of historical interaction data. Those characteristics make it particularly suitable for an AI assistant.
The effect could be smaller, larger or negative in other occupations.
The research also predates today’s most capable AI systems. The technology has improved considerably, but greater capability does not guarantee that the distribution of productivity gains will remain identical.
The finding worth carrying forward is therefore not a universal productivity percentage.
It is the shape of the effect.
The original 34% figure and the final 30% result tell the same larger story
The first widely circulated version of the research appeared as an NBER working paper.
That version analyzed 5,179 customer-support agents and reported an average productivity improvement of 14%, including a 34% gain among novice and lower-skilled workers. The original working paper remains available with those estimates.
The study was later refined and published in The Quarterly Journal of Economics.
The final peer-reviewed paper analyzes 5,172 agents and reports a 15% average productivity gain, with less experienced and lower-skilled workers improving the number of issues they resolved per hour by approximately 30%.
For current reporting, the final publication is the stronger number to use.
The revision does not change the central result.
Workers with the least experience gained roughly twice as much as the workforce overall, while highly experienced workers gained very little.
That may prove to be one of the most economically important characteristics of generative AI.
The technology is often discussed as a substitute for human knowledge. In this workplace, one of its strongest effects was making valuable knowledge available to employees who had not yet accumulated it themselves.
If that pattern holds across more occupations, AI’s biggest productivity contribution may not come from making the best workers dramatically better.
It may come from helping everyone else catch up faster.