Work and AI: what the labor market figures really show

Around the breakthrough of generative AI and advanced language models, leading macroeconomic reports predicted an unprecedented shockwave in the labour market. According to early model calculations, hundreds of thousands of jobs in administrative support, legal analysis, software development, graphic design and customer contact would disappear or shrink substantially within a few years. When we look at the statistics actually recorded by official bodies such as Statistics Netherlands (CBS), the employee insurance agency UWV and international institutions such as the OECD and Eurostat over the period 2024 to mid-2026, however, a fundamentally different and far more layered picture emerges.

Instead of a sudden rise in unemployment, the hard labour market figures show a subtler pattern of task shifts, changing job requirements and a selective slowdown in the intake of entry-level roles. The background article on which tasks are changing fastest through AI focuses on the task-based breakdown per occupational group; in this dossier we analyse realised macroeconomic occupancy rates, vacancy volume data, wage developments and the interplay between automation and the continuing tightness of the Dutch labour market.

The task-based approach versus job losses

A persistent misconception in the public debate is equating an automated sub-task with the disappearance of a full job. For labour market evaluations, economists generally use the task-based framework, in which an occupation is broken down into dozens of separate actions and interactions. Empirical analyses show that generative language models and multimodal systems can rarely take over a full job from beginning to end independently. Instead they automate specific sub-tasks, such as drafting initial text, summarising bulky documents or generating standard code blocks.

When an employee completes 25 percent of their routine actions faster thanks to tooling, this rarely results in practice in the immediate dismissal of a quarter of the department. Within the Dutch economy, which has struggled for years with structural tightness in sectors such as healthcare, education, technical installation and public administration, the time freed up is absorbed almost immediately by backlogs, quality improvement or more complex customer interactions. At macro level, automation on the task side therefore leads mainly to task reallocation and an intensification of human validation, rather than to measurable net shrinkage in the total number of people in work.

Every round of automation also brings new accompanying tasks. Think of formulating effective system instructions, checking model output for hallucinations, setting up data quality controls and safeguarding privacy and compliance standards. The time saved on executive actions is therefore partly offset by an increase in supervisory and integrating work within the same job.

Vacancy trends and skills demand in the Netherlands (2024–2026)

When we analyse the official vacancy databases and labour market analyses from UWV and CBS over the past two years, what stands out is that the total volume of open vacancies for knowledge workers has not collapsed. There is, however, a clear shift in the competences demanded. Where vacancy texts previously asked for basic skills in data processing, standard reporting or coding in specific programming languages, in 2026 employers increasingly require experience with AI-driven development tools, structured data extraction and the critical review of synthetic content.

At the same time, labour market researchers observe an asymmetric dynamic by experience level. For senior roles, where strategic insight, domain-specific context, stakeholder management and ultimate responsibility are central, demand remains exceptionally high. Employers are prepared to offer substantial terms for professionals who can deploy AI instruments to deliver complex projects faster. For generic entry-level roles (such as junior copywriters, entry-level data analysts and junior front-end developers), however, the vacancy figures show a flattening or slight decline.

Companies fill simple creation and translation tasks partly with automated workflows, which puts pressure on the traditional training ground for newcomers. Where a junior employee previously started by writing basic reports or building standard forms, these tasks are now generated directly by software layers. This creates a recruitment paradox: demand for experienced people rises, while the intake channels in which juniors can gain that experience are gradually narrowing.

Productivity gains in practice: what measurements among knowledge workers show

Several controlled field studies among software developers, customer service staff and financial analysts show considerable productivity gains from using generative assistants. Developers deliver repetitive code faster and helpdesk teams handle standard questions in less time. This gain does not, however, translate linearly into fewer working hours needed per organisation. Here the familiar economic phenomenon known as the Jevons paradox occurs: as a process becomes more efficient and cheaper, total consumption of and demand for the end product rises.

In software development, for instance, faster code production leads organisations to build more features, iterate more often and maintain larger software architectures. The complexity of the overall system increases, so demand for experienced engineers who track down bugs and take architectural decisions stays the same or even rises. On automated routing of workloads through software layers, the guide to the power of an LLM API aggregator explains how companies optimise costs and model performance when dozens of internal applications call on computing power simultaneously.

Productivity measurements also show that the return on AI systems depends heavily on organisational structure. Companies that only give staff access to a generic web interface often see marginal efficiency gains at best, because personnel lose a lot of time to manual copying and pasting. Organisations that instead build targeted API integrations within their existing ERP or CRM systems and train staff in validation protocols achieve measurable lead-time reductions of 20 to 35 percent on administrative processes.

Sectoral differences: who feels the impact first?

The impact of AI does not unfold evenly across all sectors of the Dutch economy. Where administrative, creative and legal occupations experience direct effects in their daily work processes, physical, operational and relational occupations remain largely untouched by the current generation of software models. Within financial services and the legal domain we see compliance checks, document analysis and credit assessments leaning ever more heavily on automated extraction.

This reduces manual searching and sorting, but simultaneously increases the need for legal experts who oversee interpretation, statutory liability and ethical frameworks. In accountancy we see bookings and invoice processing running almost entirely autonomously through pattern recognition, so that the accountant transforms from a processor of figures into a strategic business adviser.

In the creative sector (including graphic design, translation agencies and commercial copywriters), freelancers and agencies report clear pressure on hourly rates for routine production work. Simple localisation and basic stock material are increasingly generated by software. By contrast, services focused on brand positioning, complex final editing, fact-checking and strategy are flourishing. The split between low-value volume output and high-value bespoke work is more pronounced here than in any other industry.

In healthcare and education the situation is fundamentally different. Although language models are deployed to relieve administrative burdens (such as patient records and lesson preparation), physical and interpersonal interaction remains the heart of the profession. Because of the enormous ageing of the population and persistent staff shortages, time savings in these sectors do not lead to shrinkage but to a much-needed easing of workload and better staffing at the bedside or in the classroom.

The rise of autonomous workflows and the impact on junior positions

An important technological evolution since 2025 is the transition from static chatbots to agent systems that can carry out multi-step actions independently. These systems not only generate text, but call external APIs, consult databases, validate intermediate results and execute automated workflows without continuous human intervention. For a deeper reading of systems that carry out multi-step plans independently, the analysis of autonomous agentic AI systems offers insight into how decision loops shift from human input to model-driven orchestration.

This development particularly affects traditional stepping-stone roles within organisations. Historically, new graduates learned the trade by performing repetitive background tasks: scanning files, drafting standard letters or writing simple regression tests. When an agentic workflow handles these tasks reliably, friction arises in internal talent development: how does a junior build the practical knowledge needed to grow into a senior role if the fundamental practice material has been automated?

Companies that fail to recognise this problem in time risk a serious shortage of experienced managers and subject-matter experts over a five to ten year period. Progressive organisations are therefore restructuring their onboarding programmes: juniors are paired directly with experienced seniors to audit model output together, analyse design errors and discuss ethical dilemmas, rather than acting as manual executors.

Comparative analysis: expectation versus realised figures

To make the distance between early forecasts and current reality clear, the table below compares the expected consequences with the actual observations from Dutch and European labour market statistics over the period 2024 to mid-2026.

Legislation and regulation around AI in the workplace

Alongside technological and economic factors, legal preconditions play a decisive role in the pace at which organisations may deploy systems. With the phased entry into force of the European AI Act, strict requirements have been set on the use of algorithms in staff selection, evaluation and workplace monitoring. AI systems deployed to rank applicants, assess promotion prospects or steer task allocation fall into the high-risk category.

This brings heavy legal obligations regarding data quality, transparency, human oversight and the prevention of bias. Employers may not leave decisions with legal consequences solely to automated models. The overview of the AI Act timeline in 2026 describes which obligations for HR and selection systems now apply operationally. This regulation prevents organisations from deploying black-box models uncontrolled for staffing decisions, keeping the human dimension in personnel management legally anchored.

Works councils and trade unions also play an active role through the Works Councils Act (WOR). Introducing or substantially changing staff monitoring systems or algorithmic assessment tools requires explicit consent from the works council in the Netherlands. In practice this leads to a considered, phased rollout in which pilots are first thoroughly tested for privacy and employee wellbeing before being introduced organisation-wide.

The cost side of implementation: why adoption slows

An often overlooked factor in labour market forecasts is the total cost of ownership of advanced AI systems. Replacing an employee with a software model is less trivial in business terms than theoretical simulations suggest. Besides the direct token costs for API calls, organisations have to invest in infrastructure, enterprise licences, data cleaning, compliance audits and ongoing maintenance.

When a company integrates an AI agent into critical operational processes, a continuous need for monitoring and validation arises. Models can degrade when input formats change (data drift), generate unexpected exceptions or carry out erroneous actions in connected systems. The cost of specialised engineers to maintain and monitor these systems is substantially higher than the operational saving on simple administrative hours. This economic balance restrains overly reckless automation and forces companies into a pragmatic cost-benefit analysis per business process.

Methodological limitations and measurement errors in labour market research

When interpreting labour market data around emerging technologies, methodological caution is called for. Traditional surveys by statistical offices often lag several quarters behind technical reality. Moreover, standard occupational classifications (such as the international ISCO standard or the Dutch ROA-CBS occupational classification) measure only job titles and not the underlying composition of the working week.

When an employee keeps the same job title but spends half their day validating AI output instead of entering data manually, this substantial job transformation remains invisible in conventional macro statistics. To understand how these figures come about and where reporting differences between institutions originate, the dossier on Dutch AI in figures describes the measurement methodology of public bodies such as CBS, DNB and CPB. Combining traditional register data with real-time vacancy scraping and surveys among entrepreneurs offers the most reliable insight into how actual adaptation in the workplace is proceeding.

Surveys among employees also often show a "social desirability bias", or conversely underreporting out of fear of a change in role. Employees who use shadow AI (such as personal language model subscriptions for work-related tasks) rarely report this officially to their employer or in surveys. This creates a discrepancy between actual operational adoption in the workplace and the formal IT adoption figures that appear in management reports.

Conclusion and strategic implications for the Dutch labour market

The empirical labour market figures for the period up to mid-2026 refute both the apocalyptic scenarios of acute mass unemployment and the exaggerated promises of effortless productivity explosions. The Dutch labour market is characterised by a high degree of resilience and adaptability, driven by persistent demographic tightness, a strong services sector and a robust legal framework through European legislation.

The essential transformation is taking place at the level of skills, task allocation and quality standards. Where routine cognitive actions are automating rapidly, the economic value of critical evaluation, domain knowledge, systems architecture and interpersonal communication rises correspondingly. The greatest policy and management challenge for the coming years lies not in redistributing scarce work, but in redesigning training and career paths. Only when new talent gets the chance to build fundamental skills in an AI-supported working environment will the knowledge-intensive economy remain agile and sustainable in the long term.