The government as user and client of AI
The use of artificial intelligence within the Dutch government has shifted over the past few years from incidental experiments to a structural component of digital public administration. Where algorithms were previously used mainly in well-defined administrative processes, government organisations are now increasingly making use of advanced data analytics, machine learning and language models. This development goes to the heart of public service delivery: from granting subsidies and detecting fraud to optimising the physical living environment and supporting policy staff on complex dossiers.
As both a major consumer and purchaser of technology, the government occupies a special position. On the one hand, automation offers opportunities to increase the efficiency of public services, reduce workloads for civil servants and make better use of large volumes of societal data. On the other hand, public institutions face strict legal and societal requirements regarding proper administration, equal treatment, explainability and legal protection. After all, a decision supported by an algorithm must always be defensible to the citizen and the administrative court.
In this overview, we analyse how the Dutch government deals with AI in practice. We examine how applications are procured and developed, the frameworks and safeguards that apply to decision-making, the use of algorithm registers, the differences between large implementing organisations and municipalities, and the bottlenecks around knowledge building and staffing. For a specific overview of legal enforcement and the role of supervisory authorities, we refer to our article on AI oversight in the Netherlands.
Where and how is AI used within the government?
The range of AI applications within the public sector is diverse and growing steadily. To get a good picture of practice, it is useful to distinguish between internal process support, customer contact, and substantive judgement or enforcement.
In internal process support, organisations focus on streamlining everyday tasks. Think of automatically summarising extensive policy documents, making archives searchable, or processing incoming mail and application flows. With the rise of generative language models, various ministries and implementing agencies are experimenting with internal assistants that help civil servants draft responses or make sense of complex regulations. Although the risk profile of this type of use is considered relatively low, human oversight of the output remains necessary to prevent errors.
In customer contact and service delivery, automated systems are used to handle questions from citizens and businesses more quickly. Chatbots and virtual assistants on government websites guide visitors through complex application procedures or answer frequently asked questions about taxes, permits or benefits. As soon as a question deviates from the standard, the conversation is handed over to a human employee.
The most sensitive category concerns the use of algorithms in risk analyses, enforcement and automated decision-making. Implementing organisations and investigative services analyse large datasets to detect patterns of misuse or anomalies. Patterns are also analysed in spatial planning and environmental management, for example to monitor water quality, traffic flows or illegal construction via satellite and camera imagery. Because of the potential impact on individual civil rights, it is precisely these applications that are subject to the strictest quality requirements and transparency obligations.
Key principle for public AI: A government decision may never be the result solely of an automated process when it has legal consequences for a citizen. There must always be meaningful human involvement (the 'human-in-the-loop' principle) and clear explainability of the criteria used.
Procurement, development and decision-making
The journey of an AI application from idea to implementation within the government is subject to strict governance structures and procurement procedures. When acquiring AI functionality, government bodies face the choice between in-house development or purchasing commercial software and services from external market parties.
When the government develops custom solutions itself, control over the source code, data processing and retention periods is optimal. However, many agencies do not have enough specialised developers or data scientists to build complex models entirely in-house. This creates a strong dependence on external suppliers, IT service providers and consultancies. When purchasing off-the-shelf AI products, government organisations run into the dilemma that commercial models often function as 'black boxes': the exact operation and the training data used are not always fully transparent due to the supplier's trade secrets.
To ensure that procurement and tendering meet public values, organisations increasingly use standardised procurement conditions and assessment frameworks. For organisations seeking further strategic guidance on complex digital transformations and responsible implementations, the LLMnet Consultancy network offers in-depth insights and tailored advice.
During the decision-making process, government initiators must complete a number of mandatory assessments. An important part of this is conducting a Fundamental Rights and Algorithms Impact Assessment (IAMA) or a data protection impact assessment (DPIA). This identifies and mitigates potential risks in the areas of discrimination, privacy violations and arbitrariness in advance.
Safeguards, standards and the algorithm register
Transparency is a fundamental condition for maintaining public trust in a digital government. Citizens must be able to check what data is processed about them and the logic by which decisions are made. To guarantee this transparency, the national government Algorithm Register was established.
In this public register, government bodies publish information about the algorithms they use or plan to use. A registration typically includes details about:
- The purpose of the algorithm and the societal challenge for which it is used;
- The categories of data used and the origin of the datasets;
- The technical operation and the chosen type of model (from simple decision trees to advanced neural networks);
- The human safeguards applied and the way the output is checked;
- The results of the risk analyses and impact assessments carried out.
Although populating the algorithm register is an important step forward, practice remains stubbornly difficult. The quality and completeness of registrations varies by organisation. Moreover, explaining a complex algorithm in an understandable way to a lay audience is a challenge in itself. Governments are constantly looking for the right balance between legal-technical accuracy and accessible communication.
In addition to publicity, substantive safeguards apply in the area of data quality. Training on historical government data can cause the model to unintentionally adopt societal bias or outdated policy choices. Continuously testing for bias and ensuring representative training data are therefore standard parts of the management cycle.
Differences between implementing organisations and municipalities
The development and deployment of AI expertise is not evenly distributed across public administration. There is a clear difference in scale, resources and maturity between the large national implementing organisations and the various municipalities in the country.
National implementing organisations process enormous volumes of uniform applications and transactions every day. They have their own data labs, a substantial IT budget and specialised teams dedicated exclusively to responsible AI and data ethics. This makes them better able to develop their own frameworks, run pilots and meet complex obligations such as preparing comprehensive impact assessments.
Municipalities, by contrast, operate much closer to citizens and have an extremely diverse portfolio of responsibilities — from the social domain and debt counselling to spatial planning and waste management. The available capacity and knowledge varies greatly from one municipality to another:
- Large municipalities: Often have their own chief data officer, an ethics committee and specialised data science teams. They sometimes develop their own neighbourhood and management models.
- Medium-sized and small municipalities: Depend largely on external IT suppliers and regional partnerships for their software. For them, independently assessing the ethical and technical aspects of purchased AI software is a major challenge.
To prevent fragmentation and reinventing the wheel, national standards, umbrella organisations and joint procurement templates are increasingly being used. This allows smaller levels of government to benefit from proven methods for responsible AI as well.
Staff, knowledge building and organisational challenges
In addition to technological and legal issues, the human factor is decisive for the successful and responsible application of AI in government. The labour market for AI specialists, data engineers and ethicists is extremely competitive. Government bodies face direct competition from the private sector and the tech industry, where primary terms of employment are often more generous. You can read how this broader set of forces affects the labour market in our article on AI and the changing world of work.
To still attract and retain enough talent, the government focuses on societal relevance, training opportunities and flexible staffing arrangements. Knowledge building is not limited to the IT department, by the way. There is a broad need for greater 'AI literacy' among policymakers, legal professionals, buyers and case managers. They must be able to ask critical questions about the outcomes of an algorithm and judge when an outcome is illogical or unjust.
In summary, a number of persistent bottlenecks and dilemmas can be identified in everyday practice:
- Legacy systems and data quality: Many government processes still run on outdated IT systems. Connecting modern AI models to these legacy systems is technically complex and carries risks for data quality and privacy.
- Dependence on suppliers (vendor lock-in): Purchasing proprietary cloud and AI services can lead to strong dependence on a small number of large technology companies, putting pressure on the strategic autonomy of the government.
- Changing regulations and administrative burden: Implementing mandatory documentation, registration and assessment frameworks requires a great deal of time and capacity. The challenge is to avoid bureaucratic burden without compromising on thoroughness.
- Explainability versus complexity: The more advanced the AI model, the harder it is to demonstrate which specific variable was decisive in an individual case. This sometimes conflicts with the administrative law requirement to provide reasons.
As an AI user and purchaser, the government thus operates at a continuous intersection of innovation and safeguarding. In the years ahead, the focus will not only be on discovering new technological possibilities, but above all on embedding public values, transparency and a human dimension into every step of the digital process.
