Rules
When a decision involving AI falls under the ban on automated decision-making, which exceptions apply, and how a human genuinely takes part.
When three things come together: a decision is taken about a person, that decision is based solely on automated processing, and it has legal effects or affects the person in a similarly significant way. The prohibition in the GDPR then applies. If a human sits in between who genuinely takes part in the decision, it is no longer an automated decision.
The rule is in Article 22 of the GDPR: a person has the right not to be subject to a decision based solely on automated processing, "including profiling", which produces legal effects concerning them or "similarly significantly affects" them. According to Article 4 of the same regulation, profiling is the automated evaluation of personal aspects, such as someone's economic situation, health or reliability.
The Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, gives examples of such consequences on its page on the automated decision: someone is not invited to a job interview, does not receive a benefit, or does not get a loan, insurance or a home.
This is an explanation, not legal advice. Whether a specific application falls under the rule is something to assess with your data protection officer or a lawyer.
For a long time the question was whether a credit score is itself a decision, or only the rejection by the bank. On 7 December 2023, in the case SCHUFA (C-634/21), the Court of Justice ruled that the automated establishment of such a probability value is already an automated decision if concluding, performing or terminating a contract depends "to a large extent" on that value.
For a firm, this means: anyone who adopts a score or a recommendation from an AI system without weighing it themselves may be taking an automated decision, even if the system belongs to someone else. On the same page, the AP warns that a company may not simply use a profile bought from a credit reference agency for an automated decision.
Article 22(2) has three exceptions. The decision is necessary for entering into or performing a contract with the data subject, it is authorised by a law that also provides suitable safeguards, or the data subject has given explicit consent.
In the case of a contract or consent, Article 22(3) requires you to offer at least three things: the right to human intervention, the right to express a point of view and the right to contest the decision. Article 22(4) adds that such a decision may in principle not be based on special category data, such as health data, unless explicit consent or a substantial public interest allows it and suitable measures are in place. More on that category is in Special category data: what is it, and may AI process it?.
The Dutch supplement is Article 40 of the Dutch GDPR Implementation Act. It allows an automated decision where it is necessary for a legal obligation or a task in the public interest, but only if it is not based on profiling. If the organisation is not a public body, human intervention, the right to give a point of view and the possibility to contest must be ensured there too.
This is where things often go wrong in practice. In their guidelines on automated decision-making and profiling, the European data protection authorities write that you cannot avoid the rule by "fabricating" human involvement. Anyone who routinely applies automatically generated profiles without any real influence on the outcome is still taking an automated decision. Oversight must be meaningful, carried out by someone with the authority and competence to change the decision, who considers all the relevant data.
On 23 July 2025 the AP published guidance on meaningful human intervention. According to the AP, the intervention may not have a merely symbolic function. How it is set up matters too: a staff member hindered by time pressure or an unclear system cannot really assess the outcome. If the intervention is meaningful, the decision does not fall under Article 22.
Four questions help you test this:
A DPIA is often mandatory here as well: Article 35(3) GDPR names a systematic and extensive evaluation of personal aspects based on automated processing, on which decisions with legal effects are based. How to approach it is in DPIA for AI tools: when is it mandatory and how do you approach it?.
If you take an automated decision, you say so in advance and, under the GDPR, provide "meaningful information about the logic involved", its significance and the envisaged consequences. On 27 February 2025 the Court of Justice worked this out in Dun & Bradstreet Austria (C-203/22): you explain concisely and intelligibly which procedure and principles were actually applied to arrive at a result, such as a credit profile.
How to put that in your privacy statement is in AI in your privacy statement: what do you tell clients and staff about AI use?.
The EU AI Act sits alongside the GDPR and looks at the system itself. Point 5 of Annex III includes AI that evaluates the creditworthiness of natural persons or establishes their credit score, with an exception for detecting financial fraud, and AI for risk assessment and pricing in life and health insurance. The annex contains a comparable point for recruiting and assessing staff; see AI in HR: applications, appraisals and the rules for high-risk AI.
According to Article 113, the obligations for such high-risk systems apply from 2 December 2027 (as at November 2026). Under Article 26, anyone using such a system must then assign human oversight to people with the necessary competence, training and authority, and inform the people being decided about that the system is in use.
Article 86 gives a person about whom a decision with serious consequences has been taken on the basis of such a system the right to a clear explanation of the system's role. That right applies only to the extent that it does not already follow from other Union legislation, such as the GDPR.
If you let an AI assistant summarise a client file, find a passage or write a draft letter, and a staff member then decides for themselves, there is usually no automated decision. That changes as soon as the assistant's output is in practice the decision: a summary that routinely advises "reject" and that nobody reads any more moves towards Article 22.
Where the line falls depends on what happens in practice, not on what the process is called on paper. The table puts five common situations side by side. It is a first assessment, not a verdict on your case.
| Situation | Automated decision? | What you arrange |
|---|---|---|
| A staff member has an AI assistant summarise a client file and then decides for themselves | Usually not, as long as the staff member genuinely weighs it up | The ordinary GDPR: lawful basis, data processing agreement, information to clients |
| A system rejects an online credit application without anyone looking at it | Yes | An exception from Article 22(2), a human on request, an explanation of the logic; from 2 December 2027 also the AI Act |
| A staff member always adopts a purchased credit score without checking it | Probably yes, given the SCHUFA case | Let the staff member genuinely weigh it up, or arrange an exception with safeguards |
| A filter rejects applicants before a recruiter sees them | Yes, for those rejected | An exception, safeguards and a DPIA; high risk under the AI Act |
| AI drafts a premium proposal for a life insurance policy that an underwriter assesses | Depends on how meaningful that assessment is | The four questions above; the system falls under Annex III |
If in doubt, record why you consider that a human genuinely decides. That helps with a question from a client and with a question from the supervisory authority.
Such an assistant still processes personal data, so the ordinary rules of the GDPR continue to apply. PrivaSearch searches and cites in your client files, and states on the page on the GDPR and the AI Act that it does not assess people and takes no decisions.
It can be. In the SCHUFA case, the Court of Justice ruled that the automated establishment of a score is already an automated decision if concluding, performing or terminating a contract depends to a large extent on that score. It then makes no difference that someone other than the agency calculating the score concludes the contract.
No, not if they merely rubber-stamp it. The European data protection authorities and the AP ask for intervention that is meaningful: by someone who may and can change the decision, who looks at the relevant data and has the time to do so. A routine confirmation leaves the decision automated.
Explicit consent is one of the three exceptions in the GDPR. Even then you must give the client the opportunity to have a human look at it, to express their point of view and to contest the decision. For health data and other special category data, additional conditions apply.
You provide meaningful information about the logic involved, its significance and the consequences. According to the Court of Justice, you describe intelligibly which procedure and principles were actually applied. Handing over a technical description or the source code is not enough for that.
The prohibitions and the rules on AI literacy already apply. The obligations for high-risk systems from Annex III, including AI for creditworthiness and for pricing life and health insurance, apply from 2 December 2027 (as at November 2026). The GDPR rules on automated decisions already apply now.
What confidentiality means when you put a client file in an AI chatbot, and what the NOvA, KNB, NBA and CCBE have published about this practice.
Privacy by design and by default for AI tools: what to settle when choosing a tool, which settings to switch off and what to ask of the supplier.
What the AIIA of the Dutch Ministry of Infrastructure is, who must use it, how it is structured and how it relates to the IAMA and the DPIA.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →