Article 22 of the GDPR gives people the right not to be subject to decisions based solely on automated processing that produce legal or similarly significant effects, and two Court of Justice rulings have made that right much wider and much more practical. In SCHUFA (C-634/21, 7 December 2023) the Court held that a credit score can itself be an automated decision when a lender draws strongly on it. In Dun & Bradstreet Austria (C-203/22, 27 February 2025) it held that people are entitled to an intelligible explanation of the procedure and principles actually applied, and that trade secrets do not justify a blanket refusal.
For teams deploying scoring models, fraud engines and AI agents that act on people, these rulings move Article 22 from a niche provision to a design constraint. This guide covers the text, the two judgments, and what they mean for systems where software, not a person, makes the call. For the broader interaction between AI and data protection law, see the AI and GDPR Guide.
This page is general information, not legal advice. Status as of 17 Nov 2026. Based on the GDPR text on EUR-Lex and the judgments in Case C-634/21 and Case C-203/22. Take advice on your own processing.
Key takeaways
- Article 22(1) applies where there is a decision, based solely on automated processing including profiling, that produces legal effects concerning the person or similarly significantly affects them.
- The Court treats Article 22(1) as a prohibition in principle: automated decisions are allowed only under one of the Article 22(2) exceptions and with the safeguards in Article 22(3) and (4) (C-634/21, paras 52 to 55 and 64).
- SCHUFA: producing a probability score is itself an automated decision where a third party draws strongly on it to establish, implement or terminate a contract. Treating the score as a mere preparatory step would leave a gap in protection.
- Dun & Bradstreet: Article 15(1)(h) gives a genuine right to an explanation of the procedure and principles actually applied, in a concise, transparent, intelligible and easily accessible form. Handing over a complex formula or the full algorithm is not enough.
- Where the controller claims trade secrets or third-party data, it must give the information to the supervisory authority or court, which balances the interests case by case. National rules that exclude access as a rule are precluded.
- For AI agents, "a human was technically in the loop" is not a defence if the human rubber-stamps the output. Design for meaningful review, explanation and contest from the start.
Article 22 at a glance
| Provision | What it says |
|---|---|
| Article 22(1) | Right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects or similarly significantly affects the person. |
| Article 22(2) | Exceptions: (a) necessary for entering into or performing a contract; (b) authorised by Union or Member State law with suitable safeguards; (c) based on explicit consent. |
| Article 22(3) | For (a) and (c), the controller must implement suitable measures, at least the right to obtain human intervention, to express a point of view and to contest the decision. |
| Article 22(4) | Decisions must not be based on special category data unless Article 9(2)(a) or (g) applies and suitable safeguards are in place. |
| Articles 13(2)(f), 14(2)(g), 15(1)(h) | Transparency and access: the existence of automated decision-making, meaningful information about the logic involved, and the significance and envisaged consequences. |
| Article 21 | Right to object to processing based on public interest or legitimate interests, including profiling. |
| Article 35(3)(a) | Data protection impact assessment required for systematic and extensive evaluation based on automated processing, including profiling, on which decisions with legal or similarly significant effects are based. |
Scope: what counts as an automated decision
The Court in SCHUFA (paras 43 to 50) broke Article 22(1) into three cumulative conditions and read each broadly.
There must be a decision
The GDPR does not define "decision". The Court held that the concept is broad enough to cover a number of acts that may affect the person in many ways, and that recital 71 expressly mentions the automatic refusal of an online credit application. A creditworthiness score fell within it (paras 44 to 46).
Based solely on automated processing, including profiling
SCHUFA's activity met the Article 4(4) definition of profiling, so the second condition was satisfied (para 47). The key practical question is the word "solely". A process where a human formally signs off but in practice follows the machine is at risk of being treated as solely automated. The Court's focus in SCHUFA on whether the score plays a determining role points the same way.
Legal or similarly significant effects
Where a lender almost always refuses credit to a person with an insufficient score, the score significantly affects the person (paras 48 and 49). Credit, employment, insurance, housing, access to public services and account closures are typical examples. Recital 71 also mentions e-recruiting practices without human intervention.
The SCHUFA ruling (C-634/21)
The case was a reference from the Administrative Court of Wiesbaden in a dispute between a data subject (OQ) and the Land Hessen, with SCHUFA Holding AG, a German credit information agency, intervening. The dispute concerned a credit score supplied by SCHUFA to a lender and the data subject's request for information about it.
The operative part holds that Article 22(1) must be interpreted as meaning that the automated establishment, by a credit information agency, of a probability value based on personal data relating to a person and concerning their ability to meet payment commitments in the future constitutes automated individual decision-making, where a third party to which that value is transmitted draws strongly on it to establish, implement or terminate a contractual relationship with that person.
The reasoning matters as much as the outcome. If only the lender's final decision counted, the agency that actually computes the score would sit outside Article 22, and the person could not get the Article 15(1)(h) information from the agency, while the lender would generally not hold it. The Court called that a risk of circumventing Article 22 and a lacuna in legal protection (paras 61 to 63).
On Article 22(2)(b), the Court held that Member State law authorising automated decisions must lay down suitable safeguards, including human intervention, the right to express a point of view and the right to contest (paras 65 and 66). Such processing must also comply with Articles 5 and 6, and Member States cannot add conditions to the Article 6 legal bases or predetermine the result of a legitimate interests balance (paras 67 to 70). Whether the German provision (section 31 of the Federal Data Protection Act) met those tests was left to the national court.
The Dun & Bradstreet Austria ruling (C-203/22)
The case concerned a credit assessment carried out by Dun & Bradstreet Austria. The data subject (CK) sought meaningful information about the logic involved, and the Austrian authorities and courts referred questions on how far Article 15(1)(h) reaches and how it interacts with trade secrets.
What "meaningful information about the logic involved" means
The Court held that the data subject is entitled to an explanation of the procedure and principles actually applied in order to use, by automated means, their personal data to obtain a specific result, such as a credit profile, in a concise, transparent, intelligible and easily accessible form (operative part, point 1, and para 58). Recital 71 supports a genuine right to an explanation (para 57).
- Merely communicating a complex mathematical formula, such as an algorithm, or a detailed description of every step, does not satisfy the requirement, because neither is concise and intelligible (para 59).
- The controller does not have to disclose the algorithm itself, but complexity does not excuse it from explaining (paras 60 and 61).
- It may be sufficient to tell the person to what extent a variation in the personal data taken into account would have led to a different result, which is in effect a counterfactual explanation (para 62).
- If the actual result differs from what the explanation would predict, the controller must explain the procedure and principles behind the actual result (paras 64 and 65).
Trade secrets and third-party data
Where the controller considers that the information includes third-party personal data or trade secrets within the meaning of Directive (EU) 2016/943, it must provide that allegedly protected information to the competent supervisory authority or court, which balances the rights and interests at issue and decides the extent of access (operative part, point 2, and paras 73 and 74). Article 15(1)(h) precludes a national provision, such as section 4(6) of the Austrian Data Protection Act, that excludes access as a rule where it would compromise a business or trade secret (para 75).
Key obligations for controllers
- Identify Article 22 processing. Map every process where an automated output determines or strongly drives a decision with legal or similarly significant effects, including scores you supply to others.
- Find an exception. Rely on contract necessity, explicit consent or a specific Union or Member State law. Legitimate interests alone is not an Article 22(2) exception.
- Implement safeguards. Human intervention, the right to express a point of view and the right to contest (Article 22(3)), plus the Article 22(4) restriction on special category data.
- Be transparent up front. Disclose the existence of automated decision-making, meaningful information about the logic and the envisaged consequences in privacy notices (Articles 13(2)(f) and 14(2)(g)).
- Answer access requests properly. Provide the explanation required by C-203/22, not source code or a formula, and route trade secret claims to the authority or court rather than refusing.
- Assess. Carry out a data protection impact assessment where Article 35(3)(a) applies.
Enforcement and penalties
Article 22 is enforced by national supervisory authorities, and individuals can bring complaints and court claims. Infringements can lead to administrative fines under Article 83; check Article 83(5) on EUR-Lex for the tier that covers data subject rights. The two rulings are preliminary references, so they bind national courts and authorities on interpretation, and the individual cases were returned to the referring courts to apply. For enforcement trends, see GDPR Fines and Enforcement Statistics.
What it means for security, identity and AI agent teams
Article 22 is usually discussed for credit and hiring, but security and identity systems make automated decisions about people every day.
- Fraud and risk scoring. Automatic account blocks, transaction refusals or onboarding rejections driven by a risk engine can have legal or similarly significant effects. SCHUFA means a vendor supplying the score may itself be making the decision if the customer draws strongly on it.
- Identity verification. Automated rejection at identity proofing, for example a failed document or liveness check that blocks access to a financial or public service, needs a human review route and a way to contest.
- Insider risk and employee monitoring. Automated access revocation or disciplinary flags based on behaviour analytics should be reviewed by a person with authority to overrule before effects land on the employee.
- AI agents. An agent that approves refunds, closes accounts, ranks applicants or sets prices is executing decisions. If no person reviews the specific outcome, it is likely solely automated. Use approval gates for decisions with significant effects, see the AI Agent Action Approvals and Spending Limits Guide.
Two engineering consequences follow from Dun & Bradstreet. First, you need to be able to explain the specific result for a specific person, which means logging the inputs, model or rule version, and output for each decision, tied to the identity of the system or agent that made it. Second, counterfactual explanations ("if X had been different, the result would have changed") are explicitly endorsed as a possible form of explanation, so building them into the decision service is a sound investment. Where the system is also a high-risk AI system, human oversight duties overlap with the EU AI Act Article 14 Guide.
Compliance steps
- Build an inventory of automated decision points, including scores and recommendations you send to, or receive from, third parties.
- For each, record the Article 22(2) exception, the legal basis under Article 6, and whether special category data is involved.
- Define meaningful human review: who reviews, what information they see, what authority they have to overrule, and how often they disagree with the system.
- Create a contest channel and service level for responses.
- Instrument decisions: per decision logging of inputs, version, output, reviewing human and acting system identity, retained for the access and complaint window.
- Prepare explanation templates describing the procedure and principles applied, plus counterfactual statements where feasible.
- Agree internally how trade secret claims will be handled, with material ready to give to the supervisory authority or court.
- Complete or update the DPIA and review it when models or rules change.
Common failure patterns
- Labelling a process "human in the loop" when reviewers approve nearly every automated outcome without reading the case.
- Score vendors assuming Article 22 belongs only to their customers, which SCHUFA rules out where the customer draws strongly on the score.
- Answering access requests with a model card, source code or formula instead of an intelligible explanation.
- Refusing explanations outright on trade secret grounds rather than using the authority or court route.
- Relying on legitimate interests as if it were an Article 22(2) exception.
- No per decision logs, so the controller cannot reconstruct why a specific person was refused.
- AI agents given authority to take significant actions without an approval gate or contest route.
Practitioner checklist
- All automated decision points with legal or similarly significant effects are inventoried, including third-party scores.
- Each has a documented Article 22(2) exception and Article 22(3) safeguards.
- Human reviewers have authority, information and time to change outcomes, and override rates are monitored.
- Privacy notices describe automated decision-making, the logic and consequences.
- Access request responses explain the procedure and principles actually applied, with counterfactuals where feasible.
- Decision logs record inputs, versions, outputs and the identity of the system, agent or person acting.
- Trade secret claims follow a documented route to the supervisory authority or court.
- A DPIA covers each Article 35(3)(a) process and is reviewed on model change.
Standards and references
- EUR-Lex: Regulation (EU) 2016/679, General Data Protection Regulation
- EUR-Lex: Judgment of 7 December 2023, OQ v Land Hessen (SCHUFA), C-634/21
- EUR-Lex: Judgment of 27 February 2025, CK v Magistrat der Stadt Wien (Dun & Bradstreet Austria), C-203/22
- EUR-Lex: Regulation (EU) 2024/1689 (Artificial Intelligence Act)
- NIST: SP 800-53 Rev. 5 (AU-3 Content of Audit Records)
- ISO: ISO/IEC 27001:2022 Information security management systems
Related NHI Mgmt Group resources: AI and GDPR Guide · EU AI Act vs GDPR · EU AI Act Article 14 Guide · GDPR Article 32 Security of Processing Guide