Annex III of the EU AI Act lists creditworthiness assessment and credit scoring of natural persons as high-risk. An AI agent that judges a person's ability to repay, or produces a score a lender relies on, is inside it. Providers then owe risk management, data governance and logging; deploying lenders owe oversight, monitoring and a fundamental rights impact assessment.
What does Annex III say about credit, in plain words?
The EU AI Act (Regulation (EU) 2024/1689) sorts AI systems by risk. Most uses carry few duties. Annex III names specific uses that are high-risk, and one of them is AI used to evaluate the creditworthiness of natural persons or establish their credit score. Risk assessment and pricing for life and health insurance sits in the same annex, for similar reasons, as we explain in our piece on Annex III life and health insurance pricing.
The reason is access. A credit decision decides whether someone can rent a home, buy a car or cover an emergency. The Act treats a system that shapes that decision as one that must be built and run under strict controls. If you need the basics first, start with what credit scoring is.
Two limits matter. The listing covers natural persons, so assessing a company's credit is outside it, although a sole trader's or a guarantor's assessment may not be. And the Act carves out systems used only to detect financial fraud.
Who does the high-risk listing apply to?
The Act splits duties between providers and deployers. The provider builds the system or puts it on the market under its name. The deployer uses it. A lender that buys a credit agent is usually a deployer; a lender that builds its own, or substantially modifies a bought one, can become a provider too.
Providers carry the heavier load: a risk management system (Article 9), data governance (Article 10), automatic logs (Article 12), information to deployers (Article 13), design for human oversight (Article 14) and accuracy, robustness and cybersecurity (Article 15). Deployers must use the system as instructed, assign human oversight to competent people, monitor it, keep its logs and inform people affected (Article 26). Deployers using AI for credit scoring of individuals must also carry out a fundamental rights impact assessment (Article 27).
The Act entered into force on 1 August 2024 and applies in phases. The dates for high-risk duties have been subject to proposed changes, so plan on substance rather than a date. Our overview of the EU AI Act for fintechs deploying agents covers the wider set of duties.
Why is an agent in creditworthiness assessment high-risk even if it does not decide?
This is where most teams get it wrong. They reason that a person makes the final call, so the agent is only an assistant and not high-risk.
The Act allows a provider to argue that a listed system is not high-risk when it performs only a narrow procedural task, or prepares work without materially influencing the outcome. That argument fails for any system that profiles natural persons, and an agent that reads a person's statements, income and spending to assess whether they can repay is profiling them. Its summary and recommendation also tend to shape the decision, which is the opposite of a narrow procedural task.
So consider an agent that retrieves bank statements, computes affordability, flags risk factors and drafts a recommendation. It is part of creditworthiness assessment even if an underwriter clicks "approve". Our short answer on whether a credit scoring agent is high-risk walks through the edge cases.
What may fall outside: an agent that only checks a document is legible, routes files to queues, or books appointments. The test is whether its output feeds the judgement about a person's ability to pay.
How do you meet the duties when the agent is high-risk?
The table maps the main duties to the controls an agent needs and the evidence each leaves.
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Risk management across the lifecycle | Named risks for each capability: wrong affordability, proxy discrimination, overreach into decisions | Risk register entries tied to the agent and its capabilities |
| Data governance | Defined sources for the agent, checked for relevance, gaps and bias | Source list with provenance and review dates |
| Automatic logs | Every retrieval, summary, recommendation and hand-over recorded | Append-only event log per application |
| Human oversight | The agent recommends at depth 2; a trained underwriter decides and can override | Decision records showing who decided and any override |
| Accuracy and robustness | Replay against real past applications before release and after each change | Test results with pass floors and sign-off |
| Fundamental rights impact assessment | Assessment of who is affected and how, before first use | Signed assessment, reviewed on material change |
| Information to people affected | Applicants told an AI system is used in assessing them | Notice wording and delivery records |
What will a supervisor ask?
In banking, financial supervisors are expected to take a lead role in supervising AI used by the institutions they already oversee, so expect these questions from familiar faces:
- Which of your agents take part in assessing individuals' creditworthiness, and are you their provider or deployer?
- Show us the fundamental rights impact assessment for this agent.
- Who exercises human oversight, what training do they have, and how often do they override?
- What do the logs show for this declined application?
- How do you know the agent's affordability figures are accurate, and when were they last tested?
- What changed in the agent since it went live, and what was re-tested?
What does Annex III not cover?
The listing does not, by itself, decide what reasons you give a declined applicant; that sits in consumer credit law and, where solely automated decisions occur, GDPR Article 22. It does not cover business lending to companies. It does not replace existing banking rules on model risk, credit risk or conduct, which our guide to what AI agents in banking may and may not do covers. And it does not make every agent in a lender high-risk: a service agent answering balance questions is in a different category, with a transparency duty to tell people they are talking to an AI system. For writing decline reasons well, see our playbook on how to draft credit decline reasons with an AI agent.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Pryme separates the capabilities inside a credit agent, so reading statements can sit at depth 1 while recommending a decision sits at depth 2 and deciding is refused. Every retrieval and recommendation is filed in an append-only record, which serves as the log trail the Act expects. Attest replays the agent against your own past applications before it goes live.
Questions people ask
Is a lender that buys a credit agent a provider or a deployer?
Usually a deployer. It can become a provider if it puts the system on the market under its own name or substantially modifies it, so changes to a bought agent need a deliberate check.
Does a human final decision take the agent out of the high-risk category?
Not if the agent profiles the applicant and its output feeds the decision. Human review is one of the required controls, not an exemption from them.
Does Annex III apply to small business lending?
The listing covers natural persons. Lending to companies falls outside it, but assessments of sole traders or personal guarantors may fall inside, so map each product.