AI in the workplace makes many workers faster, and on writing and routine analysis it often lifts the quality of their work too. The gains are uneven. They are largest for less experienced people and on tasks the tools handle well, and they can shrink or reverse on unfamiliar tasks, for seasoned experts, and for skills people stop practicing.
Three assumptions travel with almost every conversation about these tools at work. The first is that a chatbot or coding assistant makes everyone more productive by roughly the same amount, so the only question is how fast to roll it out. The second is that the gains are free: you get speed without paying for it in quality, in skill, or in how people feel about their jobs. The third is that workers can tell whether the tool is helping them, so their own reports are a good guide to whether it works.
Each of these assumptions is partly wrong, and the ways they are wrong matter for anyone who manages people. The controlled studies now available show large average gains that hide wide differences between people and between tasks. They show a hard edge where help turns into harm, and that edge is invisible to the person using the tool. They show at least one group of highly experienced professionals who became slower while believing they were faster. They raise a real worry that skills fade when a machine does the noticing for us. And they show that working alongside software, rather than alongside colleagues, can change how connected people feel at work.
This article walks through that evidence, older studies first and the newest second, and then turns it into decisions you can make on Monday morning.
Does AI in the Workplace Make Everyone Faster?
The popular story goes like this. A widely used chatbot arrives, a study finds that people finish tasks 40% faster, and the number becomes a promise. Leaders multiply it across their headcount and expect the savings to show up in the budget. Vendors repeat it. Conference slides repeat it.
The trouble starts with where the big numbers come from. Most early studies used short tasks, designed by researchers, that fit the strengths of the tools: drafting a press release, writing a small piece of code from scratch, summarizing a report. Tasks like these have clear boundaries and do not depend on years of local knowledge. Real jobs look different. They are full of messy context, half finished work from colleagues, unwritten rules, and problems nobody has seen before.
When researchers pool the evidence, the setting matters a great deal. A 2026 meta analysis combined 23 studies of software programming with tools built on large language models (the text prediction systems behind modern chatbots). It found a moderate average gain in productivity, but the gains were larger in controlled experiments and smaller in open source and company settings, and the results varied widely from study to study. The same review found no reliable effect on how much people learned. It is a preprint that has not yet completed peer review, so treat it as a strong signal rather than a final word.
This is the gap at the heart of the subject. We have good evidence that the tools can speed up well defined tasks. We have much thinner evidence about what happens across a whole job, over months, for people with different levels of skill. The rest of this article fills in what we do know, and it is more interesting than the headline.
What Did the First Controlled Experiments Find?
The study that set the tone was a randomized experiment published in Science in 2023. Researchers gave 453 college educated professionals writing tasks drawn from their own occupations and paid them for quality. Half were randomly given access to a chatbot built on a large language model. In that 2023 experiment, the average time to finish fell by 40% and the quality of the output, judged by people in the same professions, rose by 18%. The gap between stronger and weaker writers shrank. Two weeks later, people who had used the chatbot in the experiment were twice as likely to be using it in their real jobs.
A second early study looked at programmers. Developers were asked to build a simple web server as quickly as possible, and some were randomly given an assistant that suggests lines of code as you type. In one controlled trial, the group with the assistant finished 55.8% faster. This is a preprint that has not completed peer review, and the task was a small, self contained job of the kind these assistants handle best.
Law gives a useful contrast because legal work is judged harshly on quality. In a randomized trial with law students, 59 students at one law school completed four realistic tasks: drafting a complaint, drafting a contract, writing a section of an employee handbook, and writing a memo to a client. Each student used a language model on two tasks and worked unaided on the other two, and graders did not know which was which. Time fell on every task, by between 12% and 32%. Quality was another matter. Only contract drafting showed a reliable improvement. The other gains were small enough to be chance, and the client memo came out slightly worse on average. The study appeared in a law review, which is edited by students rather than reviewed by other scholars, so it sits a step below journal research in the evidence hierarchy.
Put these three studies side by side and a pattern appears that runs through the rest of the research. Speed is the most dependable benefit. Quality gains are real on some tasks, absent on others, and occasionally negative. A leader who hears "40% faster" and assumes "40% better" has confused two different outcomes.
Who Gains the Most, and Why Do Top Performers Gain Less?
The most widely discussed field experiment on this question took place inside a global management consulting firm. Researchers recruited 758 consultants, about 7% of the firm's consultants at that level, and randomly assigned them to work with or without a large language model on 18 realistic consulting tasks such as generating product ideas and writing a launch plan. In this field experiment, consultants with the tool completed 12.2% more tasks, finished 25.1% faster, and produced work rated more than 40% higher in quality. The paper is a working paper that has been widely cited but has not been published in a peer reviewed journal.
The distribution of the gains matters more than the average. Consultants who scored below average on a baseline task improved their scores by 43%. Those above average improved by 17%. The tool lifted everyone, but it lifted the bottom much more than the top, and so the gap between the weakest and strongest consultants narrowed.
The same pattern shows up with programmers in the field. Three field experiments, published in Management Science in 2026, gave a random subset of 4,867 software developers at three large companies access to a coding assistant. Pooled across all three, developers with access completed about 26% more tasks. Each experiment on its own was noisy, and the results varied between companies, which is a reason for caution. But one finding held: less experienced developers took up the tool more often and gained more from it.
Creative work tells a similar story with a twist. In an online writing experiment, some participants received story ideas from a language model before writing a short story. Readers rated those stories as more creative, better written, and more enjoyable, and the boost was largest for writers who had scored low on a creativity test beforehand. Yet the stories written with the tool's ideas were more similar to one another than stories written without it. Each writer gained; the group as a whole produced a narrower range of work.
Why would the tool help weaker performers so much more? The simplest explanation is that these systems are trained on enormous amounts of competent work. They produce output close to the typical standard of a field. For a person working below that standard, the tool pulls them up. For a person already working above it, the tool offers less, and its suggestions can even pull them back toward the average.
For managers, this has two sides. A tool that narrows the gap between your weakest and strongest people is good news for consistency and for new hires who need to reach a basic standard quickly. It is less good news if the tool hides real differences in ability, or if everyone's work starts to sound the same. A team of writers whose drafts all converge on the same safe middle is not a creative team, however fast it works.
Where Does the Help Turn Into Harm?
The consulting experiment contained a second test that deserves more attention than it gets. Researchers designed one task to sit just outside what the tool could do well: a business problem in which the right answer depended on spotting details in interview notes that the tool tended to miss. On this task, consultants using the tool were 19 percentage points less likely to reach the correct answer than consultants working alone. The researchers described the boundary as a jagged frontier. Inside it, the tool helps. Just outside it, the tool produces confident, fluent, wrong answers, and people accept them.
The frontier is jagged because it does not follow human intuitions about difficulty. A tool may draft a complex strategy memo well and then fail at a simple counting problem. Nothing in its tone tells you which side of the line you are on.
The broadest test of this idea comes from a 2024 meta analysis in a leading behavioral science journal. Researchers pooled 106 experiments, reporting 370 results, from 74 papers published between 2020 and mid 2023. On average, a person working with an artificial intelligence system did better than the person working alone. But the combination did worse than the better of the two working alone. Put plainly, if the software was better than the human at a task, adding the human usually dragged performance down, and if the human was better, adding the software helped. The combination did worse on decision tasks, such as classifying cases, and better on creation tasks, such as writing. Explanations from the system and displays of its confidence did not change these results. The studies varied widely, which is a reason to treat the averages as a guide rather than a law.
Medicine shows how this plays out in skilled work. In one experiment with radiologists, 27 doctors at three German university hospitals read 50 mammograms with suggestions from what they were told was an artificial intelligence system. In fact, the suggestions were set in advance, and 12 were deliberately wrong. When the suggestion was correct, every group of readers was right about 80% of the time. When it was wrong, inexperienced readers got the case right only about 20% of the time, and even very experienced readers only 45.5% of the time. Psychologists call this automation bias: the tendency to go along with a machine's suggestion even when your own judgment would have done better. Experience protected readers somewhat. It did not protect them fully.
So the worker who sits beside the tool can fail in two directions. The radiology experiment shows people trusting it too much, especially on cases just beyond their reach. The meta analysis shows the reverse failure as well: people overriding a tool that was better than they were. Neither group can easily tell which mistake they are making.
Why Were Seasoned Developers Slower With Help?
The most surprising study in this field was published as a preprint in 2025 by a nonprofit research group. It is worth slowing down for.
Researchers recruited 16 experienced developers who maintain large, mature open source software projects, the kind of public code bases that thousands of people depend on. On average, the developers had about five years of experience on their own projects. They supplied 246 real tasks from their normal work, such as fixing bugs and adding features. Each task was randomly assigned to allow or forbid the use of the best coding tools available in early 2025.
Before starting, the developers predicted that the tools would cut their time by 24%. Economists and machine learning specialists asked to forecast the result predicted savings of 38% to 39%. The measured result in the 2025 randomized trial was the opposite: with the tools, tasks took 19% longer. Most striking of all, after finishing, the developers still believed the tools had saved them about 20% of their time.
The study is small and has not completed peer review, and the authors themselves list many features of the setting that could have contributed. These were experts working on code they knew intimately, held to high standards, on projects with many unwritten rules. They are the opposite of the novices in the field experiments above. The likely explanation is that the time saved writing code was eaten up, and more, by writing instructions to the tool, waiting for it, reading its output, and fixing suggestions that did not fit the project. For someone who already knows exactly what to type, a fluent assistant that is sometimes wrong can be a net cost.
Two lessons follow. First, the developer evidence is consistent once you look at who was studied: the large company trials found the biggest gains among less experienced developers, and this trial found losses among the most experienced. Second, and more troubling, people's sense of being helped was badly out of line with reality. That should make any manager wary of judging a rollout by asking staff whether they feel more productive.
Do Workers Lose the Skills They Hand Over?
Every tool that takes over a task raises the question of what happens to the person's ability to do it. Pilots and navigation systems are the classic case. Large language models raise the question more sharply because they take over thinking work: noticing, judging, drafting, deciding.
The strongest real world signal so far comes from medicine. In a trial in Poland, four endoscopy centers introduced a tool that highlights possible polyps during colonoscopy. A later analysis compared colonoscopies done without the tool in the three months before and the three months after it arrived. Across 1,443 patients, this observational study, published in The Lancet Gastroenterology and Hepatology in 2025, found that the rate at which doctors detected adenomas (small growths that can turn into cancer) fell from 28.4% of colonoscopies to 22.4% once they had grown used to the tool. In other words, when the assistant was switched off, the doctors appeared to look less carefully than they had before it arrived.
Caution is needed. This was an observational study over a short period, not an experiment, and other doctors challenged it in published letters. Changes in the patients seen or in other routines could account for part of the drop. But the direction of the finding is exactly what psychologists would predict.
A 2024 review in Cognitive Research: Principles and Implications sets out that prediction. Because these tools imitate mental work such as pattern recognition and reasoning, the authors argue, frequent use may wear down experts' skills faster than older forms of automation did. Two further risks follow. Learners who train with a tool may perform well while it is present but never build the ability to work without it, and the gap would only appear when the tool is removed. And both experts and learners may not notice the decline at all, because doing well with the tool feels like being good at the task. This is a theoretical review rather than new data, but it explains the endoscopy finding and the developers' mistaken self assessment in the same breath.
Motivation may fade along with skill. Four online experiments published in Scientific Reports in 2025 involved 3,562 participants. People who did one writing or idea task with a chatbot and then a second task on their own did better on the first task, but the boost did not reliably carry into the second. Their interest in the work dropped after switching to solo work, and in most of the studies they reported more boredom than people who had worked alone throughout. These were short online tasks with paid participants, not employees in real jobs, so the size of the effect in an office is unknown. Still, the finding is a warning: a tool that takes over the interesting parts of a task may leave people less engaged with the parts that remain.
How Does AI in the Workplace Affect Wellbeing and Connection?
Work is not only output. It is also where many adults get most of their daily contact with other people. When a worker asks a chatbot instead of a colleague, something social is lost.
The best evidence on this comes from a set of four studies published in the Journal of Applied Psychology in 2023. Researchers studied 794 employees in Taiwan, Indonesia, the United States, and Malaysia, using surveys, a field experiment, and a simulation. The more people interacted with artificial intelligence systems to get their work done, the more they felt a need to connect with other people. That need had a good side and a bad side. On the good side, it pushed people to help their coworkers more. On the bad side, it came with loneliness, and loneliness was linked to more insomnia and more drinking after work. The effects were strongest for people high in attachment anxiety, that is, people who worry a lot about being rejected or left out.
Large population data give a calmer picture, at least for older forms of the technology. A study of German workers published in Scientific Reports in 2025 followed about 18,500 people from 2000 to 2020. Workers whose occupations were more exposed to artificial intelligence showed no meaningful decline in life satisfaction, job satisfaction, or mental health, and they worked about 30 minutes less each week with no loss of pay. When the researchers used workers' own reports of using artificial intelligence instead of occupation averages, they did find small drops in life and job satisfaction. The data end before modern chatbots spread, and the sample leans toward middle aged workers who stayed in their occupations, so the findings may not hold for today's tools.
The effect on wellbeing also depends on what the tool takes away. Doctors spend hours each day writing notes about patient visits, and this paperwork is a major cause of burnout. In a randomized trial published in 2025 in a medical journal devoted to artificial intelligence, 238 outpatient physicians across 14 specialties were assigned to one of two scribe tools that listen to the visit and draft the note, or to their usual routine. One tool reduced time spent writing notes by 9.5% compared with the usual routine; the other made no reliable difference. Both groups of tool users reported signs of lower burnout and lighter workload, though the authors say these secondary findings need confirming in larger trials. Doctors also reported that the tools occasionally produced clinically important errors, so the notes still had to be checked.
Teamwork offers a final angle. In a 2024 field experiment at a large consumer goods company, published in Organization Science in 2026, 791 professionals worked on real product development problems either alone or in pairs, with or without a large language model. Individuals working with the tool produced solutions of similar quality to pairs working without it. People using the tool also reported feeling more enthusiastic and less anxious and frustrated. Those feelings were self reported over a single day, so they say little about how people feel after months of working mostly with software.
Read together, these studies suggest a simple rule. When the tool removes drudgery that people resent, such as paperwork, it can lift wellbeing. When it replaces the colleague people used to turn to, it can leave them lonelier, and the cost lands hardest on those who most need belonging.
Is Faster Work Also Better Work?
It is worth gathering the evidence on quality in one place, because speed and quality are so often treated as the same thing.
In the writing experiment, both improved. In the consulting experiment, both improved inside the frontier, and accuracy collapsed just outside it. In the law student trial, speed improved every time while quality barely moved and slipped on one task. In the story writing study, quality rose for each writer while variety across writers fell. In the mammography study, quality depended entirely on whether the machine's suggestion happened to be right. In the scribe trial, time savings were modest and the notes still needed checking for errors.
Three conclusions follow. First, speed is the safest bet; quality has to be checked task by task. Second, quality problems are often hidden. A wrong answer written fluently looks like a right answer until someone with expertise examines it, and as the endoscopy and skill decay evidence suggests, that expertise may be the very thing that fades with heavy use. Third, some quality losses are collective rather than individual. If everyone on a team writes with the same assistant, the team's output may grow more uniform even while each person's work looks better.
What Does This Mean for You?
If you manage people, the evidence asks you to stop thinking about these tools as a single upgrade and start thinking about them task by task and person by person.
Your newer and less experienced staff are likely to gain the most. Give them access early, but pair it with ways to build real skill, because a person who only ever works with the tool may never learn to spot its mistakes. Your most experienced experts may gain little on the work they know best, and could even slow down. Do not force a tool on them because a headline says it saves 40% of time; let them test it on their own tasks and measure the result.
Do not rely on how productive people feel. The developers who were 19% slower believed they were 20% faster. Measure time, output, and error rates directly, and compare like with like.
Treat tasks just outside the tool's strengths as dangerous ground. Your people need to know where the frontier lies for their own work, and that knowledge only comes from deliberate testing against cases where the right answer is known.
Watch what happens to skills and to people. If a skill matters when the tool is unavailable, make sure people still practice it. And notice who is now asking a chatbot rather than a colleague. The loneliness evidence suggests that the people most at risk are the ones least likely to say so.
Key Takeaways
- Speed is the most reliable benefit of these tools; quality rises on some tasks, stays flat on others, and can fall on tasks just beyond the tool's reach.
- Less experienced and lower performing workers gain the most, which narrows performance gaps but can also make everyone's work sound alike.
- Highly experienced experts working on familiar, complex problems may gain nothing or even slow down, and they may not realize it.
- Workers' own sense of how much the tool helps them is a poor measure, so judge any rollout on measured time, output, and errors.
- Skills that a tool takes over can fade without the person noticing, as early evidence from doctors suggests, so critical skills need regular unaided practice.
- Replacing human contact with software can increase loneliness for some workers, while removing hated paperwork can ease burnout; the design of the work decides which happens.
Implications for Practice
Start with an inventory of tasks, not jobs. The research shows gains that vary sharply from one task to the next within the same role, so the useful unit of analysis is the task. For each role you plan to support with these tools, list the main tasks and sort them into three groups: tasks where the tool has been tested and works well, tasks where it has been tested and fails, and tasks nobody has tested. Only the first group should be rolled out at scale. The third group needs a trial before anyone relies on the tool. This follows directly from the consulting experiment, where the same consultants gained a great deal on some tasks and lost accuracy on another.
Run your own small experiments before a wide rollout. The developer studies show that results from one setting do not transfer neatly to another, and the self assessment evidence shows that you cannot rely on people's impressions. A simple design works: give the tool to a randomly chosen half of a team for a month, keep the other half on their usual routine, and compare measured output, time, and error rates. Include your most experienced people and your newest people in both groups, so you can see whether the gains differ by experience as they did in the research.
Build a verification step into every high stakes task. Automation bias was strong even among experienced radiologists. For any output that affects customers, money, safety, or decisions about people, require a check by someone with real expertise, and make that check specific: what exactly should be verified, against what source. A quick glance at fluent text is not verification.
Protect the skills you cannot afford to lose. If your organization would be in trouble when the tool is unavailable, wrong, or withdrawn, those skills need a maintenance plan. Practical options include scheduled periods of unaided work, periodic assessments done without the tool, and training in which new staff learn the task by hand before they learn it with assistance. The endoscopy finding and the skill decay review both point to the same risk, and both suggest that the decline may be invisible until it matters.
Design jobs so that people still talk to people. The loneliness research suggests that the social cost of these tools comes from replacing contact with colleagues. Where a tool now answers questions that used to go to a coworker, build in other reasons to connect, such as regular team problem solving, peer review of work produced with the tool, and pairing arrangements. Pay particular attention to remote workers and to people who already seem isolated.
Aim the tools at the work people resent. The scribe trial suggests that removing paperwork can ease burnout, while the motivation experiments suggest that handing over the interesting parts of a task can drain engagement. When choosing where to deploy, ask employees which parts of their work they would happily give away and which parts they find meaningful, and start with the former.
Finally, revisit performance management. If the tool narrows the gap between weaker and stronger performers on routine output, then measures based on volume will tell you less about ability than they used to. Shift more weight toward work the tool does poorly, such as judgment on unusual cases, quality of checking, and the ability to explain why an answer is right, so that your assessments still separate real skill from borrowed fluency.
