DBA and false self-employment: the line between a contract for services and employment

The market for freelance machine learning specialists, prompt engineers, and AI architects has matured, but with it the legal complexity surrounding hiring has also grown. Anyone working as a self-employed professional on AI systems in the Netherlands operates within the enforcement framework of the Wet Deregulering Beoordeling Arbeidsrelaties (Wet DBA). The line between a legitimate contract for services and a disguised employment relationship (false self-employment) is not about intentions on paper, but about how the collaboration actually works in practice.

For both clients and self-employed professionals, a clear understanding of the assessment criteria is necessary to avoid back taxes, fines, and employment law claims. In this guide, we analyze the legal touchstones using concrete situations in software and model development. We look at the influence of authority relationships, organizational integration, entrepreneurial risk, and specific complications surrounding hardware, data, and intellectual property.

The legal playing field: from DBA to the Deliveroo ruling

Under Dutch law, the qualification of an employment relationship rests primarily on article 7:610 of the Dutch Civil Code (Burgerlijk Wetboek). An employment contract exists as soon as three cumulative elements are present: work, pay, and a relationship of authority over a certain period. Where the focus traditionally lay mainly on formal contract clauses, the Dutch Supreme Court (Hoge Raad) definitively established with the Deliveroo ruling (2023) that a holistic weighing is required. All circumstances of the case must be weighed in relation to one another.

This means that a clause such as "the parties expressly intend not to enter into an employment contract" is legally worthless as soon as the actual course of events points to employment. During checks, the Belastingdienst assesses whether the zzp'er can work free of instructions, bears economic risk, and can arrange for a substitute. Anyone considering the switch from a permanent job to self-employment must first master the basic principles of market positioning and acquisition; see the guide on freelancing as an AI specialist for setting up a sustainable practice and finding assignments outside of secondment.

The core of the assessment breaks down into three main pillars that together determine whether an assignment can be carried out outside of employment:

• Authority and direction: Does the client have the authority to give operational instructions about how the AI model is trained or evaluated?
• Organizational integration: Does the task form a structural part of the company's core activities, and does the specialist take part in regular performance-review cycles?
• Entrepreneurship and risk: Does the specialist bear financial risk in case of non-performance, invest in their own tooling, and work for multiple clients at the same time?

Relationship of authority versus specialist autonomy

The distinction between subject-matter specifications and operational authority is often subtle in the AI sector. A client may set functional requirements for the end result, such as a maximum inference latency of 200 milliseconds or a minimum F1 score of 0.92 on a specific validation set. This falls under the client's authority to give instructions about the desired result (art. 7:402 BW).

It goes wrong as soon as the client interferes with the way the work is done: mandating fixed working hours, requiring attendance at general staff meetings, or forcibly imposing a specific development environment without an objective technical necessity for it. As soon as an engineering manager assigns tasks daily via stand-ups as if the freelancer were a regular junior developer, the presumption of a relationship of authority arises.

Organizational integration in agile teams

Many tech companies work with multidisciplinary scrum teams in which permanent employees and hired professionals mix seamlessly. This practice in particular poses a significant risk under DBA enforcement. When an AI engineer works through the same Jira backlog as salaried developers, takes part in the on-call rotation, and participates in performance reviews, the tax authorities consider this person organizationally integrated.

To prevent integration, the assignment must be formulated as a well-defined project with a clear scope. Instead of "supporting the data team for six months", the assignment description would, for example, read: "design, implementation, and documentation of a RAG pipeline for the customer service department, including integration tests". The freelancer reports on progress to a project owner and not to a line manager.

Anyone considering combining the Dutch market with foreign clients will find, in the overview of international remote AI jobs , practical guidance for setting up cross-border contracts without local tax entanglement.

Materials, hardware, and data access: the paradox in AI

In traditional IT assignments, the use of one's own hardware (the "Bring Your Own Device" rule) is considered a strong indication of self-employment. In machine learning and LLM engineering, however, a paradox arises here due to security requirements, privacy legislation (AVG/GDPR), and enormous compute costs.

A freelancer can hardly bring their own cluster of eight Nvidia H100 GPUs to a client. Nor may a bank or healthcare institution allow sensitive training data to be downloaded to an external specialist's private laptop. Access via virtual desktops (VDI), shielded cloud environments (such as AWS GovCloud or Azure Private Link), and hardware provided by the client is often a hard compliance requirement in this sector.

The Belastingdienst acknowledges that providing equipment and environments does not automatically lead to an employment relationship if this follows from strict security or confidentiality requirements. However, it must then be established, both contractually and in practice, that the access serves exclusively for carrying out the project and is not accompanied by further-reaching control over the working person.

Compensation and rate as an indicator

Although a high hourly rate does not rule out false self-employment, the financial profile weighs heavily in the overall assessment. In case law on false self-employment, a low rate is often seen as an indication of economic dependency and employment. In the AI sector, hourly rates for experienced engineers and strategists are typically well above the market averages for general software development.

A self-employed professional must demonstrably charge rates that provide sufficient margin for entrepreneurial risk, disability insurance, pension accrual, and periodic investments in their own hardware and training. To get a clear picture of the financial and legal differences between permanent and interim work, the comparison on Self-employment (zzp) or salaried employment in AI offers insight into net income, liabilities, and secondary benefits.

Contractual structure: from hour factory to result-based agreements

The traditional model of billing by the hour based on open-ended contracts increases the risk of reclassification. A safe collaboration requires a well-considered contract for services (model agreement) with explicit obligations of result, clear deliverables, and clear Service Level Agreements.

When an AI model is put into production, agreements about uptime, recovery times, and model drift must be carefully documented without the freelancer having to be on standby 24/7 like a pseudo-employee; for this, consult the guide on AI contracts and SLA provisions to legally separate operational guarantees from personal effort in a watertight way.

Below is an example of how a task definition in a Statement of Work (SOW) should be formulated to safeguard the entrepreneurial character of the assignment:

// Slecht: Wijst op gezag en operationele inbedding
Opdrachtnemer verleent gedurende 6 maanden data science ondersteuning 
aan het Marketing Analytics team voor 36 uur per week, onder leiding 
van de Lead Data Scientist. Werkzaamheden worden uitgevoerd op kantoor.

// Goed: Resultaatgericht, autonoom en buiten gezagsverhouding
Opdrachtnemer realiseert vóór 15 november een getraind classificatiemodel 
voor fraudedetectie, inclusief Dockerized inferentie-API en validatierapport. 
Opdrachtnemer bepaalt zelfstandig werktijden en werklocatie, mits voldaan 
wordt aan de vastgelegde acceptatiecriteria (F1-score >= 0.90).

The consequences of reclassification for both parties

The risks of false self-employment are asymmetric but significantly affect both the hiring organization and the self-employed professional once the inspectorate proceeds with a correction:

For the client

• Back-tax assessments for payroll taxes: The Belastingdienst retroactively collects payroll tax and national insurance contributions, plus tax interest and any fines.
• Employment law claims: The falsely self-employed worker can retroactively claim employee rights, such as continued payment during illness, vacation days, severance pay (transitievergoeding), and pension accrual.
• Chain rule (ketenregeling): Successive freelance assignments can unintentionally turn into an employment contract for an indefinite period.

For the self-employed specialist

• Loss of entrepreneurial tax benefits: The self-employed tax deduction (zelfstandigenaftrek), starter's deduction, and SME profit exemption are cancelled for the years in question, resulting in substantial income tax back-payments.
• VAT corrections: VAT already invoiced can be corrected if the service turns out to actually be wages from employment.
• Loss of flexibility: The specialist loses their tax status as an IB entrepreneur (income tax business owner) for the income stream in question.

Practical checklist for AI practice

To ensure that a collaboration stays within the boundaries of the Wet DBA and the Deliveroo framework, the client and contractor must actively monitor the following points before and during the project:

1. Work with project phases and milestones: Link invoicing to measurable deliverables (e.g., delivery of the data pipeline, baseline evaluation, final model) rather than purely hours worked.
2. Safeguard the substitution clause: Include a genuine provision stating that the freelancer may be substituted by an equally qualified professional in case of illness or leave.
3. Avoid internal HR processes: No performance review meetings, no inclusion in internal phone lists or organizational charts as a team member, and no participation in non-work-related team outings.
4. Maintain multiple clients: Ensure that income is not structurally derived for more than 70% from a single party over a period of several years.
5. Use your own business branding: Invoice from your own business domain, maintain your own website and portfolio, and present yourself externally as an independent entity.

By systematically implementing these safeguards in contracts and day-to-day communication, organizations can continue to benefit from specialized external AI expertise without unintentionally taking on the role of employer.