Poland is one of the economies most exposed to the next wave of artificial-intelligence automation. Under the “Special Agent” scenario, 17.5% of the task content performed across the Polish workforce is technically susceptible to automation. The most important conclusion, however, is not that AI will eliminate one job in six. The more plausible outcome is a deep redesign of occupations, career ladders, corporate structures and the distribution of income between labour and capital.
The Coface and Observatoire des Emplois Menacés et Émergents (OEM) study covers 923 occupations and 32 countries. It measures the technical exposure of tasks, not a mechanical forecast of layoffs. A profession may remain necessary even when a large share of its present activities can be executed by AI systems. The economic impact will therefore depend on whether companies use productivity gains to expand output, improve quality and create new products—or primarily to reduce staffing needs.
17.5% of tasks does not mean 17.5% of jobs will disappear
The headline result is easy to misread. Employment is measured in people and positions, while the Coface-OEM model measures the content of work. A job is a bundle of activities with different levels of automation potential. A financial analyst may gather data, clean spreadsheets, prepare commentary, check documentation, discuss assumptions with management, assess exceptions and accept responsibility for a recommendation. An AI system may take over the first four functions long before it can replace the entire role.
The first effect of automation is therefore likely to be a change in the composition of work. Employees will spend less time collecting, reformatting and summarising information and more time validating outputs, interpreting unusual cases, communicating with clients and making decisions for which the organisation needs a clearly accountable person. Only later does the change in task composition translate into a change in headcount.
Employment adjustment can take several forms at the same time: fewer vacancies, slower replacement of departing staff, consolidation of roles, larger client portfolios per employee, flatter management structures and the transfer of selected activities to central technology teams. In many organisations, the earliest visible sign will not be a wave of dismissals. It will be a gradual decline in entry-level recruitment and a rising output expectation for every employee who remains.
The 17.5% figure is a map of technological potential. The portion that ultimately changes employment will depend on implementation costs, data quality, regulation, legal liability, error rates, cybersecurity, customer acceptance and the ability of companies to redesign end-to-end processes.
Why this automation wave is different from factory robotisation
Previous technological waves mainly targeted repetitive manual work and routine cognitive activities governed by explicit rules. Industrial robots transformed production lines, while enterprise software and robotic process automation reduced the need for manual data entry, standard accounting operations and repetitive administrative processing. The strongest pressure was usually concentrated on middle-skill employment.
Agentic AI moves the frontier towards cognitive and non-routine work. Its most favourable environment is a process in which both the input and the output are digital: an email, document, table, database, code repository, legal file, report, presentation or recorded conversation. Where information is already machine-readable and the expected result can be standardised, an AI agent can retrieve inputs, apply tools, make intermediate decisions, draft an output and route the case to the next stage.
The report describes this through a simple analytical distinction. Work is least exposed when it operates from matter to matter—building, repairing, installing, moving or handling objects in an unstable physical environment. Exposure is highest when work operates from data to data—drafting, classifying, summarising, translating, comparing and restructuring information. Tasks involving people occupy the middle ground: language can be automated, but trust, care, authority, responsibility and adaptation to genuinely unpredictable behaviour remain more difficult to delegate.
The difference between a copilot and an agent
A conventional generative-AI tool responds to an individual prompt. The employee still determines the sequence, gathers the data, evaluates the output and moves the process forward. A “Special Agent” system can instead coordinate a sequence of actions: collect documents, compare them with a policy, identify missing information, query a database, draft a response, create a record and escalate only exceptions. The decisive change is not better text generation. It is the movement from assistance with isolated tasks to partial execution of whole workflows.
| Technology phase | Operational meaning | Main labour-market consequence |
|---|---|---|
| Copilot | Employees prompt, verify and manually integrate outputs. | Faster individual tasks, but processes remain largely unchanged. |
| Special Agent | AI uses tools and coordinates multi-step workflows; people supervise validation points. | End-to-end processes can be redesigned and teams can become smaller. |
| Conductor | Persistent episodic memory supports long-running planning and coordination. | Management, orchestration and process-control roles become more exposed. |
| Prometheus Bound | Agentic systems and robotics mature without a breakthrough in causal reasoning. | Automation spreads towards installation, repair and embodied tasks. |
| Superhuman | AI exceeds human performance across cognitive domains. | A speculative outer bound in which nearly all occupations are structurally affected. |
The most exposed workers are not necessarily those in low-skill jobs
Engineering and computational occupations have the highest exposure, with 29% of task content at risk. Legal and financial professions and creative and content roles each reach 27%. Management and administrative work stands at 24%, commercial occupations at 19%, care and education at 14%, skilled trades and industrial production at 8%, and in-person manual services at 7%.
This pattern reverses the conventional assumption that higher education automatically provides protection from automation. Formal knowledge is no longer a sufficient barrier when models can access vast bodies of technical material, analyse documents and produce specialist outputs at a scale unavailable to an individual worker. The protective factor is not specialisation alone, but the combination of expertise with accountability, ambiguous judgement, client trust, organisational context and the ability to resolve exceptions that are poorly represented in historical data.
Creative work provides a useful example. The demand for content will not disappear, but the cost of producing a first draft of a text, image, presentation, analysis or design is falling sharply. Economic value moves away from basic execution and towards concept development, selection, positioning, fact checking, audience knowledge and responsibility for the final effect. Standardised output will face strong price pressure. Work in which reputation, strategic coherence and the cost of error matter will retain more human value.
Vertical complexity no longer guarantees safety
Deep domain knowledge can increasingly be reproduced by models trained on large technical corpora. Expertise protects a worker only when it is tied to context, responsibility and non-standard decisions.
Horizontal complexity remains a barrier
Jobs that confront continuously changing, unique situations are harder to automate because historical cases provide a weaker guide to the next real-world problem.
120 occupations are heading for deep transformation
The authors use a 30% threshold to identify occupations likely to undergo significant structural change. This is not a forecast that the position will be eliminated. It means that the core task mix, skills profile, team design and economics of employment may change materially. Under the Special Agent scenario, 120 of the 923 occupations studied—13% of the total—cross this threshold.
The largest numbers are found in computer and mathematical occupations and in architecture and engineering, with 19 occupations in each group. Business and finance follow with 18, and office and administration with 17. Life and social sciences account for 13, while education and library occupations account for 12. The exposure is therefore much broader than a small group of frequently discussed professions such as copywriters, translators or graphic designers.
The number of affected occupations does not capture the full depth of the shock. Sales contains fewer occupations above the threshold, but some of those roles have extremely high average exposure. In technology and engineering, a larger number of occupations will be affected, but many are more likely to be redesigned than removed. A developer using agents may produce much more code, while architecture, testing, security, integration and understanding the commercial objective become more valuable.
At the other end of the distribution, 136 occupations have exposure below 5%. Production, construction, extraction, installation and repair account for 84 of them. These roles require physical manipulation, spatial navigation, real-time sensory feedback and adaptation to conditions that cannot be fully described in advance. Without a major advance in robotics, language-based AI cannot substitute for this embodied complexity.
Poland ranks close to the most exposed advanced economies
In the detailed country comparison, the United Kingdom records 19.5% and the Netherlands 19.0%. Denmark reaches 18.2%, Canada 17.6%, and Poland, Germany and Australia each reach 17.5%. Sweden stands at 17.3%, the United States at 16.8%, France and Finland at 16.3%, and Japan at 14.6%.
Poland is therefore not positioned as a low-cost industrial economy largely insulated from cognitive automation. Its employment structure combines a substantial industrial and construction base with a large urban service economy built around finance, law, administration, IT, consulting and business-process delivery. The first group reduces the national average; the second pushes it upwards.
Poland’s exposure is close to Western European levels and among the highest in Central and Eastern Europe. The result reflects the scale of legal, financial, administrative and management work, while manufacturing and construction still provide a partial buffer.
Poland’s industrial buffer is useful—but it also reveals a growth dilemma
A high share of construction and manufacturing employment lowers the average exposure of the Polish economy because many tasks still require a person to act in the physical world. This creates short-term resilience. It does not automatically produce a strong long-term growth model.
Traditional sectors can protect employment from language-based automation while simultaneously offering less room for rapid productivity growth if investment, robotics, energy systems and management practices do not improve. Poland therefore faces a two-sided challenge. The knowledge-intensive service economy is exposed to AI substitution, while parts of the physical economy that remain more protected may struggle to generate faster value-added growth.
The favourable outcome would combine both structures: AI-supported engineering, industrial analytics, predictive maintenance, energy optimisation, logistics, embedded systems and digital twins. These fields use AI, but their value does not come from text generation alone. It comes from integrating digital intelligence with machines, infrastructure and specialist industrial knowledge—areas in which Poland can build more defensible capabilities.
The business-services sector is at the centre of the Polish risk
Poland’s business-services sector grew by combining a large talent pool, competitive labour costs, foreign-language skills, process discipline and access to the European market. Many centres have moved from transaction processing towards analytics, finance, compliance, IT, procurement and global process ownership. That upgrading is economically valuable, but it also places the sector in the group most exposed to agentic automation.
Processes performed in shared-service and outsourcing centres often have precisely the characteristics AI agents favour: high volumes, digital inputs, standard documentation, measurable service levels and a large historical record. Invoice handling, reconciliation, reporting, document review, employee support, procurement administration, customer communication and first-line compliance can be broken into repeatable workflows and partially automated.
The likely effect is not the disappearance of the sector, but a change in its economic model. A centre that competes primarily through lower labour cost faces strong pressure. A centre that owns a global process, manages exceptions, integrates systems, controls risk and develops automation can gain importance. Poland’s strategic objective should therefore be to move faster from “service delivery location” to “process owner and technology integrator”.
Most vulnerable model
High-volume execution of standard digital tasks with limited process ownership and weak client differentiation.
Transitional model
Human teams supervise AI workflows, handle exceptions and provide quality assurance across several markets.
More defensible model
Polish teams own the process, data, architecture, controls and business outcomes rather than only labour-intensive execution.
The first labour-market shock may appear in junior recruitment
The earliest evidence described in the report is concentrated among younger workers and entry-level positions in highly exposed occupations. This is economically logical. Junior employees often perform the tasks that are easiest to codify: preparing first drafts, collecting information, updating systems, conducting basic analysis, reviewing standard documents and producing routine reports.
A company does not need to dismiss an experienced specialist to reduce labour demand. It can preserve senior roles while hiring fewer graduates beneath them. The immediate financial benefit may look attractive, but it creates a long-term organisational problem. Senior competence is normally produced through years of progressively more difficult assignments. If AI performs the elementary work, companies must design a new way for inexperienced workers to acquire judgement.
This creates a “career-ladder problem”. Traditional professional development assumed that an employee learned through repeated exposure to basic cases before handling exceptions and taking responsibility. When basic cases are automated, young employees can lose the training ground that previously created future experts. Organisations may then save on headcount today but face shortages of experienced decision-makers several years later.
Automation could shift income from labour towards capital
The occupations most exposed in the Special Agent phase are central to wage formation, tax revenue and value creation in developed economies. They include large, well-paid groups in finance, engineering, law, business services and administration. This makes the macroeconomic effect different from a narrow shock to a declining industry.
If employment losses or slower recruitment are not offset by the creation of new work, labour income comes under pressure. Productivity gains initially accrue to the companies deploying AI. A further share may flow upstream to the firms controlling models, cloud infrastructure, chips and data-centre capacity. For countries that import most of this infrastructure, part of the economic surplus can leave the domestic economy.
Poland could therefore experience a double leakage. The domestic wage and payroll tax base may weaken if well-paid cognitive employment grows more slowly, while part of the capital income generated by automation is captured abroad. The fiscal effect would not be immediate, but it matters because public finances are heavily connected to employment income, consumption financed by wages and social contributions.
The distribution of gains will matter as much as their size
A productivity increase is not automatically a social gain for every group. If an organisation produces the same output with fewer employees, profits rise but wage income may fall. If it uses lower production costs to expand, export, reduce prices and create new products, employment can remain stable or even increase. The final outcome depends on demand, market structure and investment decisions.
Highly concentrated markets may convert AI productivity into margins rather than lower prices. Competitive markets are more likely to pass part of the benefit to customers. Labour institutions, taxation and corporate governance will influence how much of the surplus is reinvested, distributed to workers or paid to owners of technology and capital.
Education will be tested by the declining value of routine expertise
The report does not imply that university education becomes irrelevant. It suggests that a diploma alone will provide less protection when the knowledge encoded in textbooks, regulations and professional documents can be accessed instantly by AI. The premium shifts towards judgement, adaptability, problem definition, communication, accountability and the ability to combine domain knowledge with AI tools.
Education systems face a difficult design choice. Teaching only the use of current applications will produce skills that become obsolete quickly. Excluding AI from the classroom would also be unrealistic because graduates will work in AI-enabled organisations. The more durable approach is to teach students how to verify, question and improve machine output while maintaining independent analytical competence.
For universities and employers, assessment must move away from easily generated deliverables towards oral defence, project work, evidence of reasoning, data provenance and decision-making under uncertainty. Professional training will also need to accelerate. Workers in finance, law, media, administration and IT will require continuous adaptation rather than a single period of education at the beginning of a career.
AI creates new operational and geopolitical dependencies
Automation does not eliminate risk; it changes its location. A labour-intensive process is distributed across many employees and offices. An AI-intensive process becomes more dependent on a smaller number of models, cloud platforms, semiconductor suppliers, data centres and integration layers. Productivity can rise while the number of single points of failure increases.
For Polish companies, this means exposure to cloud outages, cyberattacks, model changes, pricing decisions, export controls, energy constraints and regulatory fragmentation. A process that appears cheaper under normal conditions may be less resilient during a geopolitical or technical disruption. The relevant comparison is therefore not “AI cost versus salary” but the full cost of ownership, continuity and control.
Strategic autonomy does not require every company or country to build a frontier model. It does require clarity about which data, workflows and capabilities are critical; the ability to switch providers; strong security; contractual control; and enough domestic competence to audit, integrate and maintain systems. Without that layer, AI adoption can increase productivity while reducing decision-making sovereignty.
Agentic AI still faces major economic and technical limits
The exposure estimates should not be read as a deployment timetable. Agentic systems remain difficult to audit because their decisions are not generated from transparent, fixed rules. Errors can cascade across a chain of agents, tool calls can fail, context can be lost, and hallucinated information can be propagated into subsequent steps. The more complex the workflow, the more difficult real-time monitoring becomes.
Costs can also rise rapidly. A multi-agent architecture may require many model calls, persistent storage, integration, security, human review, testing and maintenance. An activity may be technically automatable but economically unattractive once inference, supervision and failure costs are included. Regulated sectors face an additional barrier: responsibility cannot be delegated to a model, even when a model performs most of the processing.
This gap between technical capability and economic deployment is important for Poland. Companies that treat agentic AI as a plug-in will often fail to capture value. The main investment is not the model subscription. It is process mapping, data preparation, integration, control design, employee training and organisational change.
Metropolitan labour markets may face greater disruption
The report’s local analysis is based on France, where large urban centres show higher exposure because they concentrate knowledge-intensive employment. The same mechanism can be cautiously applied to Poland, although it is an inference rather than a separate OEM estimate for Polish regions.
Warsaw, Kraków, Wrocław, the Tri-City and Poznań contain large shares of finance, IT, consulting, business services, media, corporate administration and management functions. These cities are therefore likely to experience a faster change in the composition of office work than regions with a greater role for construction, tourism, local production or in-person services.
This does not mean that metropolitan areas will lose their advantage. They have universities, capital, clients, infrastructure and entrepreneurial density. They can create new companies and higher-value jobs. But they may also face weaker demand for traditional office space, stronger competition for advanced roles and a larger reskilling challenge. Regions centred on physical production may be more stable in task terms while capturing less of the productivity upside.
Later scenarios show how the exposure can spread
The Special Agent scenario is the most relevant for the coming years because it does not require a fundamental breakthrough in the underlying language models. It assumes that current engineering problems around tool use, coordination, memory and reliability are gradually resolved. The report also examines more speculative stages.
In the Conductor phase, AI gains episodic memory and can maintain a persistent representation of a situation over time. The average automation score rises to 26.3%, and 375 of 923 occupations—40.6%—cross the 30% threshold. Coordination and orchestration roles become more exposed because systems can maintain plans, allocate tasks and manage dependencies that previously required human managers.
In the Prometheus Bound phase, agentic AI and robotics mature without achieving full causal reasoning. The mean score rises to 29.1%, while 412 occupations—44.6%—cross the threshold. The expansion is driven less by deeper automation of already exposed cognitive roles and more by the spread into installation, repair, trades and physically situated work.
The Superhuman phase is a speculative outer bound rather than a forecast. It assumes systems comparable to or better than humans across cognitive domains. The mean score reaches 49.5%, and 93.7% of occupations have at least 30% of their tasks exposed. In this scenario, cognitive complexity ceases to provide durable protection; physical, relational and embodied work remains the last area of resistance.
| Scenario | Average automation score | Occupations above 30% exposure | Primary expansion mechanism |
|---|---|---|---|
| Special Agent | Near-term agentic phase | 120 of 923 (13%) | Digital information workflows |
| Conductor | 26.3% | 375 of 923 (40.6%) | Persistent memory, planning and coordination |
| Prometheus Bound | 29.1% | 412 of 923 (44.6%) | Mature robotics and embodied automation |
| Superhuman | 49.5% | 93.7% of occupations | Removal of the cognitive ceiling |
What Polish companies should do now
- Map tasks rather than job titles. A department-level headcount view hides which activities are standardised, digital, high-volume and easy to validate.
- Automate complete processes selectively. Isolated tools create local productivity gains; major value appears only when data, approvals and responsibility are redesigned end to end.
- Define human control points. Organisations need explicit thresholds for escalation, audit trails, ownership of decisions and procedures for system failure.
- Protect data and process knowledge. The durable asset is not access to a general model but proprietary context, customer relationships, policies and knowledge of exceptions.
- Redesign career development. When entry-level tasks are automated, junior employees need structured simulations, supervised decisions and earlier exposure to complex cases.
- Measure the distribution of benefits. Productivity should be assessed together with employment, wages, service quality, risk, resilience and the company’s ability to create future expertise.
Which companies are likely to gain the most
The strongest advantage will go to companies with scale, high-quality data, trusted brands, access to capital and the capability to integrate technology into core operations. AI reduces the cost of executing many tasks, but it does not remove the importance of distribution, customer access, reputation and responsibility. It may therefore reinforce market leaders rather than equalise competition.
A large enterprise can spread the cost of building, monitoring and securing an agentic system across millions of transactions. It also has more data for evaluation. Smaller firms can use ready-made tools and may adapt processes faster, but they have less capacity to build audit systems and absorb costly failures. Their opportunity lies in entering specialised niches and combining AI with expertise or customer relationships that larger competitors cannot easily replicate.
The defensible assets will be those that are hard to copy: proprietary data, customer trust, regulatory knowledge, distribution channels, the ability to accept liability and the integration of digital output with physical operations. The simple ability to generate text, code, images or analysis will become increasingly commoditised.
Poland can benefit—but not through cost cutting alone
AI can help Poland close part of its productivity gap, compensate for demographic pressure and expand the scale of service exports. In a labour market facing slower growth in workforce supply, automation of selected tasks can relieve shortages. The condition is that employees move towards higher-value activities rather than being removed from the production system without replacement demand.
A strategy based only on wage savings can improve margins in the short term but is unlikely to create a durable advantage. Competitors will have access to similar models. Sustainable gains will come from combining AI with product redesign, faster innovation, better customer service, new pricing models and ownership of the underlying process.
Poland has a particular opportunity in areas that combine technology with industry: production automation, predictive maintenance, energy, logistics, cybersecurity, embedded systems, industrial data analysis and digital twins. These fields require more than a language interface. They reward integration with physical assets and engineering expertise.
The most likely outcome: less routine work, smaller teams and greater responsibility
Over the next several years, a gradual reorganisation is more likely than a single, economy-wide unemployment shock. Companies will hire fewer people to prepare first drafts of documents, code and analysis. They will place more value on employees who validate results, resolve exceptions, communicate with clients and accept responsibility. One person will be able to supervise output previously produced by several colleagues, but will also control a wider set of risks.
Not all productivity gains will translate into lower employment. Some firms will use lower production costs to serve new clients, improve products and enter markets that were previously uneconomic. The employment effect will depend on demand elasticity. Where cheaper services generate a large increase in demand, employment may remain stable despite automation. Where demand is limited, higher productivity will primarily reduce staffing needs.
The largest uncertainty concerns new tasks. Earlier technologies created occupations that did not previously exist. AI may do the same, but it differs in one important respect: newly created digital tasks can themselves become automatable quickly. The emergence of new specialisations therefore does not guarantee that all lost labour demand will be replaced.
Conclusion: Poland’s exposure is a strategic test, not a countdown to mass unemployment
The 17.5% result places Poland near the more exposed advanced economies. It reflects the combination of a large legal, financial, administrative and management sector with a substantial industrial and construction base. The first structure raises exposure; the second reduces it.
The greatest risk is not that machines suddenly replace one Polish worker in six. It is a faster contraction of entry-level opportunities, pressure on well-paid cognitive roles, a shift of income from labour to capital, dependence on foreign AI infrastructure and the weakening of traditional career ladders.
The greatest opportunity is also broader than cost reduction. AI can increase the productivity of Polish companies, expand service exports, reduce labour shortages and support new products. Success depends on moving from the role of a user of foreign tools towards the role of process owner, integrator and producer of sector-specific solutions.
The outcome will be determined not by access to AI alone, but by the organisation of its benefits. Poland gains most when productivity growth is connected to investment, new products, workforce development and the retention of a larger share of value within the domestic economy. Without these conditions, automation may improve the financial performance of individual firms while weakening the wage, tax and development base of the wider economy.
Sources: Coface and Observatoire des Emplois Menacés et Émergents (OEM), The Next Automation Frontier: A Scenario Map of AI Labour Exposure, April 2026; Coface press release on AI automation exposure in Poland, 17 July 2026.
The indicators measure technical exposure of task content to automation, weighted by employment structure. They are not direct forecasts of redundancies. The discussion of Polish metropolitan areas is a structural inference based on the mechanism identified in the report, not a separate OEM estimate for individual Polish regions.
