It is 8:07 a.m., and the first invoice is glowing on the monitor. A PDF sits in the left window, the accounting system in the right. Amount, payment reference, due date, supplier identification number, cost centre. Six fields, several checks, then the “save” button. The next document looks almost the same. So does the one after that.
Yet the worker is not merely entering numbers. They spot a changed bank account, an incorrect tax charge, a duplicate and a purchase order that has not been closed. Routine turns into investigation, communication and a decision, even though from the outside it looks like a single job.
The company then deploys a system that reads standard invoices and fills in the fields. What will it do with the time saved? It can reduce headcount, process more documents, clear the queue or entrust the worker with more complex exceptions.
“A capable person has nothing to fear” is therefore only half true. Someone who recognises a faulty input, understands the consequences and assumes responsibility has more paths available. But even they are not protected from the closure of a department, the transfer of a function or a decision to turn productivity gains solely into cost savings.
This article therefore does not divide people into the valuable and the expendable. It breaks work into steps and asks what happens when the very steps that used to be paid disappear.
One invoice is not one task
The word “work” joins several layers that easily blur together in ordinary conversation. An occupation is a broad social category. A job is a specific agreement between a person and an organisation. A task is an individual activity: read a document, find a figure, compare it with the purchase order, decide what to do with an exception, explain a discrepancy and bear the consequences of a wrong decision.
Technology does not usually enter every layer at once. The spreadsheet did not abolish accounting, but it removed some manual addition. Search engines did not abolish lawyers or journalists; they shortened the path to a document and raised the expected volume of research.
Generative AI is unusual in the breadth of tasks it can imitate. It can write a standard email, summarise a text, turn a free-form response into a table, draft code, prepare an outline or translate a document. Those are still not whole occupations. They are parts of a chain of work, with different inputs, different costs of error and different needs for human context.
The same job contains steps with very different potential for automation.
Extract the amount, date and identifier from a digital document.
Enter the values into predefined fields and the correct format.
Find the purchase order, contract and expected price; flag the difference.
Assess an exception, request a correction, explain the choice and bear its consequences.
“Replacing” can also mean different things. Taking over data entry changes the content of a job; leaving two of three posts unfilled changes the number of jobs. Yet a lower price can attract more customers and increase employment. A single technical capability permits several economic outcomes.
A debate that looks for a list of “occupations destined to disappear” therefore starts at too high a level. The answer is one floor down: in the tasks that make up an occupation, the time each consumes and what an organisation does with the gain created by automation.
An occupation is too coarse a unit for prediction
In its 2025 update, the International Labour Organization assessed 29,753 tasks mapped to 436 detailed ISCO‑08 occupational groups. Its conclusion was not that a quarter of people would lose their jobs. Roughly one in four workers worldwide is in an occupation with some degree of exposure to generative AI, but the authors regard transformation of work as a more likely outcome than the complete elimination of a job. Exposure remains highest in clerical and administrative occupations.[1]
Exposure is a possibility, not a dismissal notice. It means that today’s systems could theoretically perform a significant share of the tasks. It does not say that a company will buy the tool, connect it to databases, resolve data-protection issues, accept the risk of errors and redesign its organisational process. Still less does it say whether higher performance will lead to fewer people.
The European Commission’s Joint Research Centre captures the problem with a different measure. In an analysis published in 2026, it linked 352 AI research benchmarks to fourteen cognitive abilities, 108 work tasks and 127 occupational groups, tracking the evolution of exposure from 2008 to 2024. Information- and knowledge-intensive occupations emerge as more exposed than elementary occupations.[3]
This corrects a second popular misconception: the issue is not confined to simple office work. AI also reaches programmers, teachers, engineers, analysts and lawyers because it works with text, search, information sorting and abstract patterns. Yet high exposure can mean either the possibility of substitution or the possibility of powerful augmentation. A system may take over some routine assessment from a radiologist while increasing the number of cases they can handle safely. A lawyer may spend less time on the initial search and more on checking relevance and shaping litigation strategy.
The right question is therefore not, “Is my occupation safe?” It is: “Which parts of my output can be separated, specified precisely, produced automatically and verified cheaply?” The greater the share of the working day covered by those four conditions, the stronger the downward pressure on the price of that work.
Five signs of the task whose price will fall first
The fastest task to automate is one with a digital input. An email, document, database row, audio transcript or image can be handed to a system without physical movement in the world. If the procedure is also repeatable, the model receives a similar instruction again and again. A standard output — a classification, short summary, completed form or piece of code — reduces uncertainty about what should be produced.
The fourth sign is an abundance of examples. The organisation has an archive of earlier answers, resolved cases or approved documents. The fifth is cheap verification. If an error can be caught by a rule, test, total or quick human review, generation can be run at scale. When all five properties coincide, the economic pressure is strong even if AI is imperfect. It need only be good enough and cheaper than the current process.
By contrast, tasks tend to be more resilient when they involve incomplete context, a physical environment, variable exceptions, conflicting objectives and a high cost of error. Not because AI could never perform them, but because integration, verification and responsibility eat into the savings. A person who must enter someone else’s home, calm a client, assess safety and improvise when something unexpected happens is not performing one clean operation.
The test does not judge a person’s worth. It assesses the technical and economic structure of an activity.
Text, images, audio or structured data can be processed without physical intervention.
The procedure can be converted into stable rules, an instruction or a few examples.
The saving on one case is small, but it is multiplied across thousands of repetitions.
There is an expected answer, form, classification or measurable format.
Correctness can be checked by a test, database, total or brief review by a worker.
If an error requires expert judgement and an accountable person, automation does not end when an output is generated.
Routine does not mean unimportant. Data checks protect companies from fraud and customers from error. Data entry can be a gateway to expertise. The issue is not the moral quality of a task, but the ease with which its output can be separated from the person, measured and purchased more cheaply.
Potential for automation does not mean immediate automation, either. Legacy systems, poor data or regulation may hold a simple process in place for years; a digital platform, by contrast, can spread change within months.
A “capable person” is neither a credential nor a pose
The word capable can easily slide into a flattering label for people who already have power, education and a well-paid position. In this article it means something narrower and observable: a person can turn an ambiguous objective into a verifiable instruction, recognise a faulty input, choose the right degree of scrutiny, explain a trade-off and accept responsibility for the result.
That capability can appear in an accountant just as readily as in a plumber, nurse, lawyer or warehouse manager. It is not the same as typing speed, the volume of knowledge held in memory or the ability to write an impressive prompt. A prompt is cheap. Domain context, knowledge of exceptions and other people’s trust take longer to build.
A capable worker can notice when the assignment itself leads to the wrong result. They know when a figure adds up but measures the wrong thing. They recognise that an automatically drafted contract does not address the real allocation of risk. They can say “I don’t know”, request more information and stop a process whose error would cost more than the time saved.
This definition is not a celebration of the lone genius. Many people perform a narrow routine because that is how their company designed the job: it did not give them the information, authority, time or education required to make decisions. A person can be capable and still spend all day copying data because departing from the procedure would be punished. Automation then does not reveal their low worth. It reveals the narrowness of the role the organisation created.
AI does not diminish the worth of a person whose role consisted of routine clicking. It lowers the price of the click that the company separated from judgement.
Resilience therefore does not emerge from individual learning alone. It also requires access to a broader part of the process. An employee who sees only one field will struggle to learn how to decide the whole case. A company that wants to use human judgement after automation must create the conditions for people to exercise it before the change.
An uncomfortable paradox: AI sometimes helps workers with lower baseline performance the most
The simple story says that the best workers will receive AI and widen their lead. A field study, however, reveals the opposite mechanism. The analysis covered 3,006,395 chats handled by 5,172 customer-support agents from September 2019 to June 2021; a pilot randomised rollout involving about 50 people took place in August 2020, followed by the broader rollout from October 2020 to May 2021. The main staggered-rollout difference‑in‑differences estimate was a 15.2 per cent increase in issues resolved per hour.[4]
The average conceals the most important difference. Workers in the lowest performance quintile improved by about 36 per cent, while the study found no statistically significant change in productivity for the top-performing group. Less experienced people and workers with lower initial performance became faster and improved quality; gains among the most experienced were small, and some quality measures declined slightly. The study concerned one company, one occupation and an assistant trained for a specific process. It is not a universal forecast.
The mechanism is nevertheless important. The system was able to spread response patterns previously used by the best workers. Some tacit experience became guidance available to a newcomer. That may be socially good news: a person learns faster, the customer receives a more consistent service and the language barrier loses some of its force.
At the same time, an economic question arises. If a newcomer with an assistant can handle almost as much as a more experienced colleague, the company may need fewer entry-level staff. AI can thus help an individual newcomer while narrowing the doorway to the market for newcomers as a whole. There is no contradiction. The productivity of one worker and the number of jobs are two different variables.
A result from one specific customer-support setting cannot automatically be transferred to other occupations.
Estimated increase in issues resolved per hour among the lowest performance quintile after the assistant was introduced.
Advantage: faster learning and access to practices that were previously hard to acquire.The study found no significant productivity growth among the strongest workers; some quality measures declined slightly.
Risk: a recommendation may distract an expert or tempt them towards an easier but worse answer.In its Skills Outlook 2025, the OECD connects this paradox with developments in online labour markets. Within companies, AI may improve newcomers’ performance, while online consulting and freelance marketplaces see declining demand for short-term assignments and novice work and a growing relative premium on people who can solve more complex tasks beyond the frontier of current models.[2] In other words, a tool can democratise performance while lowering the price of entry-level work.
The jagged frontier: an expert must know when not to trust AI
AI capabilities do not rise in a straight line from simple to complex. A system may summarise a long document brilliantly and fail on a short question that requires connecting a hidden fact with a rule. It may write a persuasive proposal and overlook the one condition that reverses the entire recommendation. Two tasks that seem similarly difficult to a person may sit on opposite sides of the technological frontier.
A field experiment with 758 Boston Consulting Group consultants called this unevenness the “jagged technological frontier”. The consultants were split into non-overlapping arms: 385 completed eighteen realistic tasks inside GPT‑4’s capabilities during a five-week window spanning May and June 2023, while 373 completed the single outside-frontier task. In the first arm, people using AI completed 12.2 per cent more tasks, worked 25.1 per cent faster on average and submitted higher-quality solutions. In the second arm, they reached the correct result on the one complex task 19 percentage points less often than the control group.[5]
The study’s limitations are part of its message. The negative result rests on a single type of task, involved elite consultants and used a model from 2023. It cannot support a claim that AI generally makes expert work worse. It does safely show that a professional needs a distinct skill: recognising whether a particular step falls within a tool’s reliable capabilities and setting the level of review according to the cost of a possible error.
A similarly uncomfortable result came from a 2025 randomised experiment by METR with sixteen experienced developers working on 246 real tasks in their own large open-source repositories. With the AI tools available at the time, the work took them 19 per cent longer, even though they expected to become faster. The small, specific sample does not permit generalisation to programming as a whole. It does show, however, that a subjective sense of speed is not a measure of productivity.[6]
A capable person is therefore not the one who uses AI most often. It is the one who chooses the tool for the task, not the task for the tool. For an easily verifiable draft, they may generate quickly and inspect a sample. For a court filing, safety decision or intervention in an account, they must verify the sources, the process and the final consequence. Speed without calibrated trust may only accelerate the production of error.
Higher productivity has not yet decided how many people remain
Imagine that a team saves one fifth of its time through AI. The technical experiment ends there. The economic story is only beginning. Management can cut the number of jobs by one fifth and keep output unchanged. It can retain the team and increase production. It can remove a waiting time that customers previously regarded as normal. Or it can lower the price, attract more customers and eventually hire additional people for activities that were not automated.
A study of roughly 58 million profile positions and 14 million job postings finds both directions, but its sample is not the whole market: it covers US, English-language positions linked to publicly listed Compustat firms. Postings span 2010–2023 and profile positions effectively begin in 2011. Greater average exposure of tasks to AI reduced demand for the work concerned. When exposure was concentrated in a few steps, however, workers shifted their effort elsewhere, and more productive firms expanded employment in other occupations as well. The total effect on the number of jobs was therefore smaller than the direct substitution of tasks.[7]
This is not proof that employment always “balances itself out”. The outcome depends on demand, competition, prices, the speed of adoption and whether the technology creates new human tasks. The economic framework developed by Acemoglu and Restrepo distinguishes displacement of labour, a productivity effect and the creation of new activities. Automation reduces demand for people in the task it takes over; cheaper production can increase demand for labour elsewhere; new tasks can bring people back into the process.[16]
The model does not determine the outcome on its own. The business and organisational response does.
Fewer jobs
The company maintains output and, after some departures, either leaves positions vacant or removes them outright.
More output
The same team serves more clients; employment is unchanged, but output per person rises.
The queue and backlog disappear
The saved time is redirected to review, better service availability and cases that used to be postponed.
Lower price, larger market
A cheaper service attracts more customers and creates work in sales, integration, review or care.
A Danish study linked a survey of about 25,000 workers at 7,000 workplaces in eleven exposed occupations with administrative data. Roughly two years after ChatGPT’s launch — not after two years of individual adoption — it found no statistically significant average effect on earnings or hours worked and ruled out average effects larger than two per cent. At the same time, tasks related to AI review and integration increased.[8]
This is Denmark, an early period and a particular group of chatbots. The result is not proof of permanent safety. It is evidence that technological capability does not translate into the labour market immediately or through a single channel. An organisation stands between the model and a redundancy decision.
No economy-wide wave has been established yet. But the first cracks are uneven
The most honest answer today is this: aggregate data do not yet show mass displacement of employment caused by generative AI, but some narrowly defined markets and recruitment pathways are seeing changes consistent with automation. Both halves of the sentence are necessary. The first prevents panic; the second prevents comfortable denial.
The study cleaned a dataset of 1,388,711 advertised projects on one global freelance platform; its main estimates used 1,218,463 fixed-price projects in eight clusters from July 2021 to July 2023. During the eight months after ChatGPT’s launch, writing and programming projects fell by 21 per cent relative to manually intensive projects. After image generators were introduced, visual-creation projects fell by 17 per cent in relative terms. The remaining projects were more complex and better paid; the average number of bids per open project rose by 8.57 per cent.[9]
This environment is exceptionally conducive to rapid impact: digital input and output, divisible assignments, easy price comparison and a weak relationship with the client. The result cannot be transferred to a hospital, construction site or public authority; it shows the risk where a standard output is sold on its own.
The latest US signal comes from a Stanford Digital Economy Lab working paper based on ADP payroll data covering millions of US workers through June 2026. The 12 August revision reports that employment among people aged 22 to 25 in highly exposed occupations lagged by 19 per cent behind a hypothetical trajectory in which it kept pace with their peers in less-exposed occupations. The gap mainly reflected weaker hiring, not increased dismissals; there was no comparable gap among more experienced workers.[10]
The authors also stress that this is a descriptive, not a causal, result. The difference narrows after controlling for education, part of the trend began before generative AI, and the ADP sample shows a stronger effect than some nationwide surveys. The study therefore does not say that AI “caused a nineteen-per-cent loss of jobs”. It says that a signal concentrated in time and by occupation has emerged and needs to be monitored.
The Stanford AI Index 2026 summarises the same picture: an economy-wide employment collapse has not been demonstrated, but the first signals are concentrated in weaker hiring of young people in exposed fields. Among 1,753 respondents at organisations that regularly used AI, surveyed from 25 June to 29 July 2025, 32 per cent expected their total workforce to shrink by at least three per cent in the following year, while 43 per cent expected little or no change. These are respondents’ expectations, not realised layoffs or a representative forecast for all companies.[11]
The first wave may therefore take the form of a missing junior vacancy, an unrefilled position, a smaller external contract or more cases per worker. Redundancy statistics may fail to capture it for a long time.
If we remove the rung, where will future experts come from?
Simple tasks have a dual nature. For a company, they are a cost. For a beginner, they are training. A junior lawyer learns the structure of a decision through research. A programmer comes to understand a system’s architecture by fixing defects. An accountant develops a feel for exceptions by handling standard documents. A doctor does not become experienced by studying a textbook alone; they need a sequence of ordinary situations before they can safely handle a rare one.
If AI takes over drafts and straightforward cases, the company saves money but removes part of the environment in which a newcomer learned to recognise an error. A junior then receives only exceptions, without experience of ordinary cases, or reviews outputs whose reasoning they did not construct.
The answer is not a return to copying, but a new form of training: first solve a sample independently, then compare it with AI and explain the difference; rotate between creation, review and client contact; evaluate the ability to detect a faulty input, not only speed.
The aim is not to preserve pointless clicking, but to make experience-building deliberate.
Risk: AI produces a finished output and the junior merely approves it.
Training response: first solve a sample independently, then compare the results and explain the difference.
Risk: the worker receives only difficult cases without having built the foundations.
Training response: managed escalation, a mentor and a record of why the standard procedure failed.
Risk: senior staff become scarce because the company has spent several years failing to develop successors.
Training response: gradually assume responsibility for the whole case, maintain client contact and audit decisions.
The OECD notes that a shift in demand for entry-level work and a rise in the value of more complex tasks can occur at the same time.[2] If the market stops buying the first stages of expertise, the individual advice to “be better” is not enough. Every senior was once a junior. A training pathway is infrastructure for an occupation, not a beginner’s private hobby.
The proposition about capability becomes an organisational obligation here. A person can learn, but a company must preserve room to practise. Schools can strengthen evidence literacy, and the state can improve access to education; neither can decide on an employer’s behalf whether a productivity gain will be turned into training.
Even the best employee does not decide the fate of their job alone
Capability increases the number of options. An expert can assume oversight, design a new process, work with clients or move time towards tasks AI cannot perform. That is a real advantage. It is not an insurance policy.
A company may deploy automation across the board without being able to distinguish good work from bad. It can merge teams to meet a spreadsheet target. It can transfer an activity to a supplier whose price has fallen because of AI. An entire industry may face lower demand even when individual workers remain capable. And a productivity gain may accrue to owners, customers or a platform without appearing in a worker’s wage.
That is why it is dangerous to turn an article about AI into a moral judgement of people who are dismissed. Losing a job is not automatic evidence that a person “could do nothing extra”. It results from the interaction of capability, task composition, the business model, bargaining power and organisational decisions. Capability influences the probability, not the entire mechanism.
The cross-sectional JRC AIM‑WORK survey ran from October 2024 to January 2025 and collected responses from 70,316 people aged 16 to 65 in all 27 EU Member States. In the paid-employment module, about 29.9 per cent of working respondents reported having used an AI tool at work at least once in the previous twelve months. Self-reports do not establish causality; at the same time, they also reveal digital monitoring and algorithmic management, raising the question of autonomy and responsibility without corresponding authority.[14]
A more productive worker may be freer because routine disappears. They may also be worked more intensively because the target rises and every minute saved is immediately filled with another case. The same software can augment judgement or automate supervision. The difference lies in job design and in who is allowed to speak about that design.
Capability is an advantage within a work process. It is not a vote on company strategy.
Czechia: rapid adoption, unequal opportunities and no causal account yet
AI use is no longer a fringe experiment in Czechia. According to a Czech Statistical Office survey, 17.6 per cent of enterprises with ten or more employees used at least one of the measured AI technologies in 2025. The figure had been 11.3 per cent one year earlier and only 5.9 per cent in 2023. The indicator, however, covers a broad range of technologies — from text analysis and content generation to process automation — and says nothing about the intensity of use or the number of jobs eliminated.[12]
The difference by company size is pronounced. AI was used by 13.4 per cent of small enterprises with 10 to 49 employees, 28.8 per cent of medium-sized enterprises and 54.1 per cent of large enterprises with at least 250 employees. In the same year, the European average for enterprises with ten or more workers was 20 per cent; Czechia’s figure in Eurostat’s harmonised data was 17.6 per cent.[13]
A large company has more data, specialists, legal support and use cases from which to earn a return on investment. A small company can buy an inexpensive tool, but it has more difficulty financing secure integration and training. The same occupation may therefore have a different future in different organisations.
Czech companies also automate for different reasons: some because of costs, others because shifts remain unfilled or volumes are growing. AI can therefore eliminate a job or fill a capacity gap; a national total does not reveal what is happening within occupations.
In a survey of 5,342 employees aged 16 to 64 in eleven EU countries, conducted from February to May 2024, Cedefop reports that 28 per cent said either they or their colleagues used AI at work. It also found that 61 per cent expected to need new knowledge and skills because of AI within five years. This is neither an EU27 survey nor a forecast of Czech layoffs. The inequality of access to learning matters: someone who receives the tool, time and the whole process starts from a different position than a person whose role remains narrow until it is removed.[15]
Czech adoption therefore cannot honestly support either the claim that AI is taking jobs or the claim that it is not. We know that use is growing rapidly and unevenly. A genuine account would require linked workplace-level data on the tool, tasks, hiring, wages, hours and quality, and those data are missing.
Faster output is not the same as better work
Productivity is easy to quantify only where the result is easy to measure. Chats, documents or tickets are straightforward to count. In diagnosis, strategy, teaching or legal advice, speed is only one component; quality, consequences and the cost of review are others.
AI may cut the preparation of a first draft from one hour to ten minutes but require another thirty minutes of verification. It may increase the number of texts while adding small inaccuracies for which nobody assumes responsibility. Code may be produced faster while hidden maintenance costs rise. If we count outputs alone, we move the cost of error outside the chart.
The statement that a “human will remain in the loop” guarantees nothing by itself. A reviewer needs expertise, time, resources and the right to reject the output. If they must approve hundreds of proposals in a minute and dissent is punished, human oversight is merely decoration.
A capable person detects the error the metric cannot see. But the organisation must also recognise as productive the work that prevents an error and averts harm.
Different studies answer different questions. Their figures cannot be added together into a single forecast.
What kind of claim are we seeking?
What do we actually observe?
Where does an exaggerated conclusion begin?
ILO and JRC task indices.
Both administrative and knowledge work contain steps that can be affected.
Exposure is neither adoption nor an economic decision.
Field experiments in QJE and Organization Science.
The result depends on the task, the worker and the capability frontier.
One company or experiment does not represent an entire economy.
The freelance market, Danish data and US payroll data.
Weaker junior hiring and fewer projects in some digital categories.
No aggregate wave has yet been established, and some findings are observational.
- A long-term fall in employment that hit judgement-based roles and routine roles equally strongly, regardless of their task composition.
- Results showing that domain review and the ability to reject AI do not improve performance once quality is taken into account.
- Sustained growth in simple entry-level jobs even though AI can cheaply produce and verify their complete output.
Capability is not only the ability to create an output, but also the ability to choose the measure of its quality. Someone who merely accelerates a standard step competes with the price of a machine. Someone who determines the purpose of the step, the acceptable error and responsibility for the consequences moves their work towards decisions that, for now, are harder to separate from accountability.
Nor is this sphere a permanent human monopoly. The frontier moves. Protection therefore does not lie in a fixed list of skills learned once and for all, but in the ability to break a problem down afresh, change the procedure and admit what the tool does better — and what must not be entrusted to it without oversight.
This is not about fear. It is about a place in the chain of responsibility
The morning began with an invoice and six fields. After automation, data entry for the amount, date and identifier disappears; discrepancies and the decision whether to let the document proceed remain. If a company has narrowed the role to standard data entry, the price of that output falls sharply. A person who understands the process can take over exceptions, oversight and system improvement.
That is the core of the optimistic part of the story: a person who can frame a problem correctly, recognise an error, work with people and answer for the result has more ways to remain useful. AI can remove routine and extend their reach. Capability ceases to be the speed at which a first draft is produced and increasingly becomes the quality of the decisions around it.
The pessimistic part is equally important. A company need not turn the time saved into better work. It may eliminate the position before offering the person a route to a broader role. It may remove the junior rungs, raise the pace and leave responsibility with the employee. Even a capable person can still be harmed by a decision they did not make.
It is therefore inaccurate to say that a capable person has nothing to fear. This is more precise: capability expands the range of options and therefore resilience, not immunity. Resilience is strengthened further by an organisation that needs judgement, develops it and gives people the authority to exercise it.
Artificial intelligence does not start with the job title written into an employment contract. It starts with a step that can be extracted, described in one sentence and checked cheaply. A person’s resilience then depends in part on whether their work ended with that step — or only began beyond it.
What this article claims — and what it does not
The article combines task indices of technical exposure, productivity experiments, administrative data and worker surveys. These sources are not interchangeable. An experiment can establish a change in performance on a particular task, but not the number of future jobs. Observational data can capture concurrent developments, but without further identification they do not prove that AI caused them.
- “Capability” here is a set of observable work competencies, not a moral judgement of a person.
- The illustrative opening scenario is not a real case or a claim about a particular company.
- US, Danish and platform findings are not presented as a forecast for Czechia.
- Czech data measure the use of technologies in enterprises, not their causal impact on employment.
- The status of sources and links was checked as at 24 August 2026; rapidly changing working papers require a fresh check before publication.