This is a hypothetical, illustrative example. The invented facts are: a lender uses an assistant that scores loan applications, explains the score factors, and recommends approval, referral or decline, with a human underwriter making the final decision. On those facts, the screening in this guide finds that the assistant is an AI system in scope, that the software vendor is a provider and the lender is a deployer, that no Article 5 prohibition is shown, and that the Annex III essential services use case in Article 6(2) must be examined, with the Article 6(3) filter checked and the profiling condition applied. The conclusion is provisional, tied to the invented facts, and must be confirmed with a competent authority or legal adviser for any real system. This example does not determine any real credit scoring system.
What are the invented facts?
The invented facts are stated here so that the screening can be tested against them. A lender, Deployer L, uses an assistant, Assistant M, supplied by a vendor, Provider N, to score loan applications for personal loans. Assistant M reads the application, the credit history and the affordability information, produces a score, lists the main factors that influenced the score, and recommends approval, referral to a human underwriter, or decline. A human underwriter reviews the referred and declined cases, and makes the final decision on every application. Assistant M does not decide the loan on its own, and the underwriter can override the recommendation. No biometric identification, emotion inference or social scoring is used. The assistant is placed on the Union market by Provider N and put into service by Deployer L in the Union, and is used for credit decisions in the Union. Those facts are invented and simple, deliberately, so that the screening steps are visible. A real credit scoring system will have more complex facts, including the data sources, the model validation, the fairness testing, the explainability design, the human oversight, the record keeping, and the sector regulation, each of which can change the screening.
How is scope and role screened?
Begin with what Assistant M actually does, and who answers for it. The assistant reads an application, credit history and affordability information, produces a score with the main factors behind it, and recommends approval, referral or decline, while a human underwriter reviews referred and declined cases and makes the final decision on every application, with power to override. That picture gives the two roles at once: Provider N develops the assistant and places it on the market under its own name, and Deployer L, the lender, uses it under its own authority for a professional activity. The assistant meets the Article 3(1) definition of an AI system on these facts, subject to the Commission guidelines on the definition, and it is placed on the Union market and used in the Union, so Article 2 is satisfied and no exclusion applies. The lender's regulated status does not change the AI Act roles, but sector, prudential and consumer law apply alongside the Act and must be checked separately. Deployer L does not take on provider duties on these facts, since it neither substantially modifies the assistant nor puts its name on it, but that position must be rechecked if the lender customises the assistant or changes its purpose.
For the broader regulatory framework, see questions and misconceptions overview.
How is the prohibition screen done?
Test each prohibition against the assistant's actual behaviour, and write down why each one does or does not bite. The scoring and factor explanations are not a manipulative practice on these facts, but the explanation design, the applicant information and the marketing must still be checked for deceptive presentation. Affordability and personal data must be checked for vulnerability exploitation, including financial difficulty. Social scoring is the closest call: Assistant M scores creditworthiness for a specific loan, not general social behaviour across contexts, so the prohibition is not shown, but the distinction needs recording with reasons covering the specific purpose, the data and the context. No emotion inference, biometric categorisation or remote biometric identification is used, and the new December 2026 prohibitions are not shown, since the assistant generates no intimate or sexual content. Every conclusion here is provisional and tied to the invented facts, and any change to the data, the purpose or the presentation reopens the screen.
How is the Annex route screened?
Annex III, point 5 covers AI systems intended to evaluate the creditworthiness of natural persons or to establish their credit score, except systems used for detecting financial fraud. Assistant M evaluates creditworthiness and establishes a credit score for personal loans, so it falls within that use case in outline. Article 6(3) is then tested and fails on these facts: the assistant does more than a narrow procedural task, more than improving a prior human result, more than detecting decision-making patterns without replacing human assessment, and more than a preparatory task, because it scores, explains and recommends a decision that shapes the underwriting outcome. Profiling closes the route in any event, since scoring applicants is profiling, and the filter is unavailable where a system profiles natural persons. The provisional conclusion, tied to the invented facts, is that the assistant is a high-risk candidate under Article 6(2), with the full classification record, including the Article 6(3) reasoning, to be completed and confirmed before any reliance is placed on it.
What sector and rights questions remain open?
The open questions for Assistant M include the data governance, the bias and fairness testing across protected groups, the accuracy and robustness validation, the explainability of the score factors to underwriters and to applicants, the human oversight design and the underwriter training and override record, the data protection impact assessment, the sector regulation for credit, the consumer information and the right to an explanation under other law, the fundamental rights impact assessment position for Deployer L, the post market monitoring plan, and the application date for the high risk duties, which for Annex III systems is 2 December 2027. Each open question needs an owner, a date and evidence. The example also leaves open whether the lender is a public body or a private entity providing a public service for Article 27 purposes, which is unlikely on the invented facts for a private lender, but which must be checked where a public service is involved, and whether the general purpose model component, if any, has systemic risk. A credit scoring file that does not show fairness testing, validation and human oversight evidence will be difficult to defend, even where the classification is clear. The screening record should list the open questions with the action taken, and should be updated as the evidence is produced.
Continue with misconception: all AI is high risk.
What does this example not decide?
This walkthrough settles nothing about a real credit scoring system. It does not fix the Article 27 position for any lender, and it gives no financial regulation, data protection or consumer law opinion for any product. It is a hypothetical exercise, built from invented facts, with a provisional conclusion tied to those facts and a confirm-with-authority exit. A provider or deployer with a real credit system should run the screening order on its own facts, check the consolidated Regulation, the Commission guidelines and the sector regulation, and confirm the classification and the duties with a competent authority or legal adviser. The method is transferable; the answer for a real system is not.
Frequently asked questions
Is every credit scoring system high risk?
Annex III point 5 covers systems intended to evaluate creditworthiness or establish a credit score, except fraud detection systems. A system that only performs a narrow procedural task without profiling can fall within the Article 6(3) filter, but that must be checked. Most scoring, ranking and recommendation systems for credit will need a careful Annex III analysis.
Does a human underwriter remove the high risk classification?
No, not by itself. Human oversight is a requirement for high risk systems, not a route out of classification. A system that scores and recommends a decision can be high risk even where a human makes the final decision, because the system shapes the decision and the human can be influenced by the score.
What is the difference between credit scoring and fraud detection?
Annex III point 5 expressly excepts systems used for detecting financial fraud from the creditworthiness use case. A fraud detection system must still be screened for prohibition, transparency and other law, and must be checked against the other Annex III categories, but it is not classified under the creditworthiness point on that ground alone.
What fairness evidence is needed?
The high risk requirements include data governance, accuracy, robustness, human oversight and record keeping, with bias testing and fairness evidence as part of the data and risk management file. The exact evidence depends on the system, the data and the sector regulation, and must be worked out with the provider, the deployer and, where needed, a legal adviser. This example does not list the evidence for any real system.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk timeline ↗