debt-collectiongdprautomated-decisionsai-actcredit-management

AI in debt collection: what the Dutch DPA now expects

November 18, 20269 min readPixel Management

This article is also available in Dutch

On 29 October 2026 the Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, published a study of algorithms and AI at 29 debt collection agencies. Its finding: the sector can't reliably tell when a system is making the decision. That matters just as much to the credit management team of an energy supplier, telecom operator or online shop collecting its own arrears.

AI in debt collection in the Netherlands is allowed, but a system that decides about a debtor without meaningful human involvement falls under the Article 22 GDPR ban, even when the decision favours the debtor, such as an automatically approved payment plan. The AP considers it "plausible" that such decisions already happen without agencies noticing. If you score the likelihood of collecting a debt, the AI Act's high-risk rules will probably apply as well from 2 December 2027.

What did the Dutch DPA find in debt collection?

The AP's announcement of 29 October 2026 summarises a research report (in Dutch). In October 2025 the AP sent a questionnaire to 29 agencies on the Dutch collection register. 22 sent usable answers. Ten of them used at least one algorithmic system, so 10 out of 22, or 45 percent. Three of those ten also used actual AI systems; the rest ran on rule-based if-then logic.

Small numbers, but a deliberate sample. The AP only picked registered agencies and calls the result "the most favourable scenario". Agencies ignoring the registration duty weren't in it.

The report lists six uses the AP found in practice:

  • choosing the contact channel and frequency (post, email, phone, text; every 30, 60 or 90 days)
  • answering questions through chatbots, voice systems and language models reading email, including fully automated payment arrangements
  • estimating the likelihood of collection from inputs such as credit score, location and payment history
  • enriching debtor files with data from registers and the internet, social media activity included
  • detecting vulnerable debtors
  • analysing phone calls for sentiment and compliance with internal policy

AP board member Katja Mur says in the announcement there must always be a collections employee "of flesh and blood" who can be held to account. The AP holds a round table with the sector on 2 December 2026; registration closed on 11 November.

Which AI uses in collections are fine, need a DPIA, or cross Article 22?

The uses don't carry equal weight. This sorting is our reading of the AP report, the GDPR and the AI Act, not legal advice on your process, but it's the order in which we'd audit a collections workflow.

UsePositionWhy
Summarising a call for the agentFine, with safeguardsNo decision; watch for sensitive data in free text
Chatbot or voice bot answering questionsFine, with AI disclosureArticle 50 AI Act: the debtor must know it's AI
Choosing channel and frequency automaticallyDPIA, test against Article 22Driven by payment behaviour; the AP says it's unclear agencies checked whether this is a decision
Collection likelihood score for prioritisingDPIA, profiling, possibly high-riskA score can itself be a decision (CJEU, SCHUFA)
Auto-approving a payment planCrosses Article 22A favourable decision is still covered
Auto-rejecting a payment planCrosses Article 22Legal effect with no human
Enriching files from social mediaStop, or justify in writingThe AP calls this profiling with a basic lawfulness question
Vulnerability detectionDPIA, legal basis firstAlmost certainly special category data such as health
Sentiment analysis of callsDPIASystematic evaluation of personal aspects

Two rows surprise most credit managers. First, a positive outcome doesn't save you. One agency told the AP it only decides automatically when a payment plan request is granted, with rejections always going to a person. The report's answer is that the intended outcome "makes no difference to the ban". Second, a human step that only exists when the debtor asks for it isn't human intervention under the GDPR.

When is a collections decision automated under Article 22 GDPR?

Article 22 GDPR gives a debtor the right not to be subject to a decision based solely on automated processing, including profiling, that produces legal effects or similarly significantly affects them. There are three exceptions: the decision is necessary for a contract, the law authorises it, or the person gives explicit consent. Under the first and the third you still have to offer human intervention, a chance to give their view, and a way to contest.

The Court of Justice made clear in its SCHUFA judgment of 7 December 2023 (C-634/21) that a probability score on someone's ability to pay is itself an automated decision when the party using it "draws strongly" on it. The AP applies that to collections: an intermediate step like calculating a risk score can count as the decision. An agent working down the system's list from top to bottom doesn't turn a collection score into a human decision.

So the AP's report says to test every automated process against Article 22, check whether an exception applies, and stop processing that amounts to profiling the debtor. Where a person is added to stay outside Article 22, that involvement has to be meaningful: enough time, good instructions, and a system that doesn't steer them. The AP set this out in its guidance on meaningful human intervention of 23 July 2025 (in Dutch). Making such a decision explainable to a debtor is covered in AI explainability in decision-making.

Time saved

Save 6 hours per week on manually tracing where a system actually decides, per collections process

Is a collection score high-risk under the AI Act?

A collection score is probably high-risk once it's an AI system assessing a private individual's creditworthiness or willingness to pay. Annex III, point 5(b), of the AI Act covers AI systems "intended to be used to evaluate the creditworthiness of natural persons or establish their credit score", with fraud detection excluded. Read literally, that text is about lending, not collecting.

The AP reads it more broadly. In a box in its report it says the category also covers systems that assess how willing a person is to meet payment obligations, and that collection firms developing such systems themselves must meet the high-risk requirements "from 2 December 2027". That date holds: the Digital Omnibus, Regulation (EU) 2026/1744, published on 24 July 2026, moved the Annex III high-risk rules from 2 August 2026 to 2 December 2027. We cover the background in our piece on the AI Act Digital Omnibus.

Where does the line sit? We work with three boundaries:

  1. A rule-based decision tree ("text after 30 days") is usually not an AI system, so only the GDPR applies.
  2. A trained model ranking debtors by likelihood of payment, fed with credit scores and payment history, assesses willingness to pay. Treat it as high-risk until a regulator says otherwise.
  3. If you buy such a model, you're the deployer. For Annex III point 5(b), Article 27 then requires a fundamental rights impact assessment before you put it into use, on top of your DPIA.

A chatbot isn't covered by point 5(b). Article 50 applies to it instead, since 2 August 2026: the debtor has to be told they're dealing with an AI system. More in AI transparency requirements for businesses.

Does the Dutch collections act cover your in-house team?

The Wet kwaliteit incassodienstverlening (Wki, the Dutch Debt Collection Quality Act), in force since 1 April 2024, only covers out-of-court collection "for a third party or after transfer of the claim", from a private individual living in the Netherlands (Article 2). Anyone doing that as a business has to be on the register kept by the Minister for Legal Protection (Articles 3 and 4). Bailiffs and lawyers are exempt from registration. The Justice and Security Inspectorate supervises.

A health insurer, housing association or online shop collecting its own claims, not transferred to anyone, sits outside the Wki. That doesn't mean fewer rules. Article 22 GDPR, the DPIA duty and the AI Act don't care who collects the debt, only what the system does. In-house credit management teams carry the most risk here: the Wki framework doesn't apply, the software comes from the ERP vendor, and they read the AP's study as something for collection agencies.

The company type doesn't matter for a DPIA either. Article 35(3)(a) GDPR requires one for a systematic and extensive evaluation of personal aspects based on automated processing, including profiling, on which decisions with significant effects are based. A collection model is the textbook case. Our step-by-step DPIA for AI projects shows how to build one.

Vendors adding AI overnight

The AP notes that collection software vendors actively sell AI, "sometimes by (suddenly) adding an AI feature to software". Some agencies had no data processing agreement, sometimes because they believed no personal data went into the system. The report shows how that goes wrong: some agencies didn't count the debt itself, or a model's output, as personal data, although both are about an identifiable debtor.

For a buyer, that means three questions at every update of the collections package. Which new feature decides or ranks? What data goes in and what comes out? Who carries the provider obligations under the AI Act? Those belong in the contract, see buying AI software.

What to settle before 2 December 2027

The GDPR obligations apply today; only the high-risk requirements wait until December 2027. This is the order we'd take it in:

  1. Map every automated step in the collections flow, from first reminder to handover to the bailiff, and note for each whether a system decides, ranks or only informs.
  2. Test every deciding step against Article 22 and its three exceptions. An auto-approved payment plan belongs on that list.
  3. Stop enriching files from social media unless you can defend the legal basis and necessity on paper.
  4. Run a DPIA for collection models, vulnerability detection and call analysis.
  5. Send the privacy information actively with the first contact, with a separate, readable passage on algorithms. The report found that half of the privacy notices didn't match what agencies told the AP in the survey.
  6. Decide for each purchased model whether it falls under Annex III point 5(b), and schedule the fundamental rights impact assessment.

How this fits the wider GDPR and AI Act picture is in our guide to AI legislation in the Netherlands and the EU AI Act. If you want a collections process audited before you add AI to it, or an existing model tested, we do that as part of AI consulting.

Service

Learn more about AI consulting?

View service →

Curious how much time you could save?

Request a free efficiency audit. We'll analyze your processes and show you where the gains are, no strings attached.