AI is rewriting work, and we have to relearn what performance means
The workplace impact of AI is decided at the level of tasks, and can be measured cleanly only if the cost of checking, correcting and learning is included in the calculation. Research shows real gains, and also how easily the feeling of being helped can be mistaken for actual improvement.
Corporate language calls a faster-made presentation, a sharper customer service reply and the work of an autonomous AI agent all by the same name, performance, but I see three quite different changes in them. A working day consists of a surprisingly wide range of activities: searching for information, interpreting uncertain situations, coordinating, calculating, writing, checking, deciding, and taking responsibility for the consequences of decisions. Generative AI can offer useful help with several of these, but the value of a whole job is created by how the activities connect.
An HR employee, for example, writes job advertisements, handles applications, organizes interviews, understands a manager's unspoken expectations, and mediates between the candidate and the organization. Speeding up the writing of an advertisement is a real result, but without further evidence it does not tell us much about whether the whole role can be automated. As I see it, the first managerial task in introducing AI is precisely to carry out this breakdown, because a useful picture emerges only when the transformation of tasks, the quality of the outcome, the cost of oversight and people's learning are weighed together.
The task is the basic unit of analysis
Daron Acemoglu and Pascual Restrepo's 2019 paper, Automation and New Tasks, offers a task-based approach: the effect of technology can be understood through how the division of labour between humans and capital changes. Automation takes over tasks previously performed by people, while new human tasks may also emerge. Productivity growth and the trajectory of employment are therefore two separate questions, since better performance may increase demand, whereas the displacement of labour directly reduces the human input needed for certain tasks. The end result depends on the combined effect of these processes 1. Valerio Capraro and colleagues' 2024 review, The impact of generative artificial intelligence on socioeconomic inequalities and policy making, urges the same caution: generative AI in the workplace may raise productivity and create new jobs, but the benefits are expected to be distributed unequally 2.
For an organization, this means that analyzing workflows says more than looking at job titles. Comparing a contract is a different task from judging the business significance of a missing clause, and drafting a reply that can be sent to a client is different from determining which promise the company can make. The same model may work excellently for one of these and, for another, may require more checking than the time it saves.
The task inventory must also include invisible work. Coordination, handling exceptions, clarifying a poorly defined request and assessing the consequences often become visible when someone takes over the work after an automated step. If this is left out of the measurement, one team's efficiency gain can easily become another team's extra burden, and at the organization level part of the reported savings is in fact displaced work.
The novice and the experienced colleague get different things from the tool
The 2025 study Generative AI at Work by Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined the staged rollout of a generative AI assistant using data on 5172 customer support agents. The number of issues resolved per hour rose by 15 percent on average. Less experienced and lower-performing workers showed larger improvements, while the strongest workers saw a smaller speed gain and a slight decline in some quality measures. The study looked at a single company, and the authors explicitly note that their results do not support conclusions about aggregate employment or wages 3. A similar pattern is echoed in a 2025 paper by Matthias Söllner and colleagues, ChatGPT and Beyond: Exploring the Responsible Use of Generative AI in the Workplace, based on an experiment with professional writing tasks: alongside time savings and quality improvements, the gap between participants narrowed, because those with weaker baseline skills benefited more 4.
From this I conclude that a uniform licence does not mean uniform development needs. A new joiner can learn from good response patterns, while an experienced professional may use the system more for finding exceptions, checking, or preparatory work that would otherwise distract them. The performance of the two users should therefore be examined with different metrics.
Experienced staff have a distinct role in improving the system. If the strongest experts simply adopt a mediocre machine suggestion, the company may look more uniform in the short term while producing fewer new professional solutions. This is a possible organizational risk, also raised in research by Brynjolfsson and co-authors, and it needs to be measured separately at each company. An expert's deviation from the machine suggestion is therefore a valuable signal from which both the team and the system can learn, provided the process has a place where that deviation is recorded.
The capability frontier is uneven
In their pre-registered experiment with 758 consultants, Fabrizio Dell'Acqua and colleagues examined the jagged frontier of AI capabilities. According to Navigating the Jagged Technological Frontier, presented as a working paper in 2023 and later published in Organization Science, AI improved performance on tasks within its capabilities, but on one complex task deliberately chosen to lie outside the frontier it reduced the probability of a correct solution. The researchers call this the "jagged technological frontier": tasks that seem similarly difficult to a human may work very differently for the model 5.
This difference explains why generic AI training that teaches people to write good prompts is not enough. Employees also need to learn how to recognize the type of task, the evidence available and the limits of verifiability. The tool's fluent answer is a weak guide here on its own, because it sounds equally confident on either side of that limit.
Imagine a market research task in which the model organizes the recurring themes of interviews and writes a clear summary. It easily undervalues a rare but strategically important dissenting view if it was given the task purely as a frequency summary. In such a case, the human role lies in refining the question and recognizing the significance of the exception, and good division of labor starts from what the output will be used for.
The feeling of speed and the measurement of speed
In METR's 2025 study, 16 experienced open-source developers worked on 246 real tasks in their own projects. With the early-2025 tools examined, allowing AI use was associated with 19 percent longer completion times on average, even though participants perceived a speedup. The authors regarded this as a snapshot of one specific work environment and cautioned against generalizing 6.
The research group's follow-up in February 2026 already reported estimates pointing toward speedup, but identified substantial selection bias. Several people stayed away precisely because they no longer wanted to work without AI, and some tasks were also not submitted to random assignment. The authors therefore described the size of the recent effect as uncertain and modified the study design 7.
Sida Peng and colleagues' 2023 preprint, The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, draws the other end of the picture. In their controlled experiment, developers completed a well-defined task, writing an HTTP server in JavaScript, 55.8 percent faster with the AI tool than the control group 8. To me, the two figures together mean that the measured effect is a property of the task and the environment: a well-specified task started from a blank slate measures something different from a mature codebase, where most of the work is integration and verification.
Taken together, the results teach the discipline of measurement. Data from an old model must be interpreted together with its year and context. And a favourable new impression becomes reliable management information only when the measurement also addresses who and which tasks were left out of it.
In corporate measurement, it is therefore worth recording not only the time of production but also the time of review, correction and follow-up queries. The speed of the first version and the speed of the accepted result may differ, and the difference matters most when a faulty result enters further workflows. An internal trial could compare cases of similar difficulty, properly classified, with evaluators examining the output against pre-set quality criteria, separately for novice and experienced colleagues. When organizing the trial, it is also necessary to record which cases were left out and who received what support. The plan must be adapted to local operations. Such a study is less spectacular than a few successful demos, but it shows much more precisely which work step is worth investing further time and money in.
Psychological safety is a quality assurance factor
Amy Edmondson and Jean-François Harvey's 2017 book Extreme Teaming shows the importance of speaking up and sharing knowledge in collaboration across different fields of expertise. Psychological safety is an environment in which a colleague can take the social risk of asking questions, admitting mistakes and voicing a different position. Shared knowledge becomes usable when people are able to say what is uncomfortable as well 9.
In an AI setting, this has a direct operational consequence. An employee may spot a fault in a system the manager favours, recognise a misleading summary, or realise that their own automation was built on a faulty assumption. The value of such a signal is determined by the organisation's reaction. If the rollout may only be presented as a success story, information about errors gets stuck, and the feedback loop breaks precisely where the system could be fixed.
Two recent survey studies add nuance to the picture. A 2026 preprint by Aaron Reich and colleagues, Safety First: Psychological Safety as the Key to AI Transformation, found, using data from 2,257 employees of a global consulting firm, that psychological safety goes together with the uptake of AI tools, but does not predict the frequency or duration of use 10. A 2025 study by Byung-Jik Kim and colleagues, The dark side of artificial intelligence adoption, in a three-wave survey of 381 South Korean employees, concluded that organisational AI adoption was associated with lower psychological safety, which in turn was associated with higher levels of depression, and that the first relationship was weaker under ethical leadership 11. Both findings are self-reported, observational data, so I read them as associations; treating them as causal evidence would be premature. They do suggest, however, that the safety to speak up may be both an input and an output of adoption, which means that managers may do well to measure how it changes during rollout.
My own suggestion is a regular, short professional review in which each team presents one successful use, one problematic case and one changed work step. With this I translate the theory of collaboration into AI practice: the aim is for individual experience to become shared knowledge, so that the next colleague does the same task better prepared.
Learning must be redesigned too
Introducing AI raises a delicate question for early-career workers. In many professions, later independent judgment was built precisely on smaller preparatory tasks: the beginner collected data, made comparisons, wrote drafts, made mistakes and then received feedback. If a system does most of the preparation, the organization has to create new learning situations deliberately.
The size of the effect may differ by task, and the sensible response is graduated responsibility. A junior can be given material compiled by AI with the task of showing its strongest and weakest claim, checking the decisive source, and then justifying what they would change in it. This way, professional judgment can be practiced alongside fast production.
For a manager, the question is how much better the decisions the team can make six months from now will be when the usual system gives an incomplete answer. Lasting competence is a reserve that becomes especially valuable when a new situation, missing data or an unusual client problem appears. Measurement that rewards only today's output makes this reserve invisible, and what measurement does not see, an organization will in time stop allocating resources to.
Leadership decides what happens to the time saved
Imagine a team that saves ten hours a week on a recurring analysis. The ten hours alone are the start of several possible business stories. The team can produce more material, check existing material more thoroughly, talk to more customers, or reduce overtime. Each decision creates a different value and a different work culture.
I believe management should state in advance what it expects from the freed-up capacity, because in the absence of a stated expectation, the easiest metric to count, volume, fills the gap. If improving quality is the goal, the number of errors corrected after the fact and of unsupported claims should be measurable. If speeding up the service is the goal, the turnaround time perceived by the customer should be tracked. If uncovering new business opportunities is the goal, the number of hypotheses tested and evaluated can be monitored.
Incentivizing employees is consistent only if it also accepts the sensible bypassing of AI use. An experienced colleague may know that they can do a given small task faster and more reliably directly. An organization that focuses on results treats this as a professional decision and evaluates the whole workflow; the frequency of tool use is at most a side datum there.
The place of human value in the new division of labour
With AI, good problem framing, understanding of interests, verification of results and responsibility for decisions may become more valuable. However, these activities require training, access and time. A mere exhortation that everyone should be more creative and strategic offers little practical help on its own.
A usable organizational response is concrete. It shows what AI is suited for in which task, who checks the output, in what cases a human decision must take over, and what feedback improves the system. At the same time, it preserves the ability of staff to understand the result they hand over.
The future of work is thus shaped in part by a management decision about what kind of work is built from faster output. At the end of a good AI deployment, the organisation knows more about how it operates, the employee understands their responsibilities more precisely, and the customer gets a better result. The measure of performance should make these three changes visible, and I consider a deployment successful if all three appear in its productivity table.
Recommended reading
- Automation and New Tasks: How Technology Displaces and Reinstates Labor (Acemoglu, D. és Restrepo, P). The foundational text of the task-based framework: it shows why productivity and employment are separate questions, and how displacement and the emergence of new tasks act at the same time.
- Generative AI at Work (Brynjolfsson, E., Li, D. és Raymond, L). One of the most widely cited corporate studies on how the same AI assistant gives different results to a novice and to an experienced colleague.
- Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality (Dell’Acqua, F. és munkatársai). The experiment that gave rise to the concept of the jagged capability frontier: performance improves inside the frontier and worsens outside it.
- Extreme Teaming: Lessons in Complex, Cross-Sector Leadership (Edmondson, A. C. és Harvey, J.-F). Through case studies, it shows how speaking up and knowledge sharing become preconditions for cooperation across disciplines.
- Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity (Daron Acemoglu, Simon Johnson). A broader historical framework for seeing the distribution of technology's benefits as a matter of decisions and institutions.
Sources
- [1] Acemoglu, D. és Restrepo, P (2019): Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2), 3–30.
- [2] Valerio Capraro, Austin Lentsch, Daron Acemoğlu, Selin Akgün és munkatársai (2024): The impact of generative artificial intelligence on socioeconomic inequalities and policy making. PNAS Nexus.
- [3] Brynjolfsson, E., Li, D. és Raymond, L (2025): Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942.
- [4] Matthias Söllner, Thomas Arnold, Alexander Benlian, Ulrich Bretschneider és munkatársai (2025): ChatGPT and Beyond: Exploring the Responsible Use of Generative AI in the Workplace. Business & Information Systems Engineering.
- [5] Dell’Acqua, F. és munkatársai (2026): Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403–423. A kutatás első munkatanulmánya 2023-ban jelent meg.
- [6] Becker, J., Rush, N., Barnes, B. és Rein, D (2025): Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, 2025. július 10.
- [7] Becker, J., Rush, N., Cunningham, T., Rein, D. és Mahamud, K (2026): We are Changing our Developer Productivity Experiment Design. METR, 2026. február 24.
- [8] Sida Peng, Eirini Kalliamvakou, Peter Cihon, Mert Demirer (2023): The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv (Cornell University).
- [9] Edmondson, A. C. és Harvey, J.-F (2017): Extreme Teaming: Lessons in Complex, Cross-Sector Leadership. Emerald Publishing. Különösen a 3. és 5. fejezet. ISBN 978-1-78635-450-1.
- [10] Aaron Reich, Diana Wolfe, Matt Price, Alice Choe és munkatársai (2026): Safety First: Psychological Safety as the Key to AI Transformation. arXiv (Cornell University).
- [11] Byung‐Jik Kim, Min‐Jik Kim, Julak Lee (2025): The dark side of artificial intelligence adoption: linking artificial intelligence adoption to employee depression via psychological safety and ethical leadership. Humanities and Social Sciences Communications.
- [12] Daron Acemoglu, Simon Johnson (2023): Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity.