EU AI Act | Financial Services

EU AI Act High-Risk Requirements for Creditworthiness Assessment at Banks

AI systems used to evaluate the creditworthiness or credit score of natural persons are high-risk under the EU AI Act, with narrow exceptions such as fraud detection and prudential capital calculation. As of October 2026, the Annex III high-risk rules apply from 2 December 2027. Banks acting as deployers must use these systems according to their instructions, monitor them and act on risks. Runtime policy enforcement can evidence what a system was permitted to do, what it did and when a human was involved. Model validation, dataset governance and conformity assessment remain separate responsibilities.

  1. Inventory every AI use in credit decisioning and classify each as in or out of scope, recording the reasoning for any fraud-detection, prudential-capital or derogation exclusion.
  2. Identify the provider of each in-scope system and obtain its instructions of use, since deployer use and monitoring duties are defined by them.
  3. Assign a control owner and an evidence source to each deployer duty: use, monitoring, risk response and input data.
  4. Test whether you can show, for one credit interaction, the permitted actions, the actual action and any human review.
  5. Confirm with your national competent authority which existing financial-services-law controls may be relied on.
  6. Check the fundamental rights impact assessment expectation against the operative text for your systems.

Translating operational duties into runtime controls

Event recording

Runtime control: every interaction passes through policy enforcement. Evidence: a timestamped record of the request, the policy version applied and the action taken.

Human oversight

Runtime control: defined conditions escalate for human approval before the interaction proceeds. Evidence: the interaction, the policy that triggered escalation, the reviewer and the outcome.

Use per instructions of use

Runtime control: policy limits permitted tools, data and actions to what the instructions allow. Evidence: policy rules traceable to the provider's instructions.

Deployer monitoring

Runtime control: allow, redact, block and escalate counts by policy. Evidence: periodic views that feed risk review. The bank must still act on what they show.

Input data handling

Runtime control: redact or block fields outside the intended use before they reach the system. Evidence: records of redaction and blocking. Representativeness of data remains a governance task.

Accuracy, robustness, conformity

Runtime control: none. Model validation, testing and the provider's conformity assessment sit outside the execution path.

Which credit-related AI uses are high-risk

The Act classifies AI systems used to evaluate the credit score or creditworthiness of natural persons as high-risk. Its stated reasoning is that these systems determine access to financial resources and essential services such as housing, electricity and telecommunication services. Annex III lists systems intended to evaluate the creditworthiness of natural persons among the high-risk areas.

Two carve-outs matter for banks. The recitals state that AI systems provided for by Union law to detect fraud in the offering of financial services, and those used for prudential purposes to calculate the capital requirements of credit institutions and insurance undertakings, should not be considered high-risk. Separately, an Annex III system is not high-risk where it does not pose a significant risk of harm to health, safety or fundamental rights, including by not materially influencing the outcome of decision making.

In practice, classification should be done per use, not per vendor or per model. A scoring model, an assistant that summarises an applicant file for an underwriter and a fraud screen can sit in the same workflow and fall on different sides of the line. Record the reasoning for each exclusion you rely on. A system whose output feeds a credit decision is hard to place under the derogation, so that conclusion needs to be tested rather than assumed. The prior question is whether the technique is an AI system at all.

Provider and deployer obligations for banks

The high-risk requirements cover risk management, data quality and relevance, documentation and traceability, record-keeping, transparency and information to deployers, human oversight, accuracy, robustness and cybersecurity. Most of these are design-time obligations that fall on the provider. Before placing a high-risk system on the EU market or putting it into service, the provider must complete a conformity assessment, repeated if the system or its purpose is substantially modified. Providers also remain responsible across the lifecycle for serious-incident reporting, corrective actions and cooperation with market surveillance authorities.

A bank that buys a credit scoring system is typically the deployer. Deployers must use the system in accordance with its instructions of use and take all technical means to that end. They must monitor operation based on those instructions and act on identified risks or serious incidents. Where the deployer supplies input data, that data must be relevant and sufficiently representative for the intended purpose. Recital 96 adds that banking and insurance deployers of certain annexed high-risk systems should carry out a fundamental rights impact assessment before use.

Role allocation needs explicit attention where a bank builds a system or substantially modifies a purchased one, because that is where provider-side duties, including repeated conformity assessment, can arrive inside the bank.

Fitting the Act alongside prudential supervision and the timeline

Banks already operate under prudential supervision and model risk management, so the Act does not arrive in a vacuum. An ECB opinion published in the Official Journal on 15 April 2026 records the ECB's understanding that credit institutions acting as providers or deployers of high-risk systems are deemed to comply with certain AI Act provisions when they comply with similar obligations under Union financial services law. It also records that conformity assessments, breach identification and AI Act enforcement are tasks of market surveillance authorities, not part of the ECB's prudential mandate. Expect two supervisory conversations, and confirm with your national competent authority which existing controls can be relied on.

The same opinion suggested clarifying that generalised linear models such as logistic regression used for credit scoring should be excluded from the high-risk definition. That is a comment on a proposal, not an exclusion in force, and it should not be used to take systems out of scope.

On timing, as of October 2026, Annex III high-risk rules apply from 2 December 2027, and rules for high-risk AI embedded in regulated products apply from 2 August 2028. Amending Regulation (EU) 2026/1744 changed the application dates for high-risk requirements, so earlier timelines are superseded. Enforcement powers of the AI Office and national competent authorities apply from 2 August 2026.

Aligning the human oversight record with the existing model risk management framework avoids a parallel process. The record is best owned by the business and risk function, not only by IT.

Where runtime enforcement helps and where it does not

The weakness in many programmes is reconstruction. Teams can describe the intended controls, but for a single credit interaction they cannot show what the system was permitted to do, what it did and whether a person was involved without assembling fragments from several logs. Enforcement in the execution path addresses that gap because the evidence is created by the control itself, at the moment the action is allowed, redacted, blocked or escalated.

Trussed enforces policy in the execution path on individual interactions and tool calls, with actions that include allow, redact, block and escalate for human approval. It produces audit-ready evidence from that enforcement, mapped to the EU AI Act among other frameworks, and deploys self-managed in the customer cloud with configurable retention and data residency. Insurance, also a regulated financial sector, is where Trussed is deepest.

The boundary is as important as the capability. Runtime enforcement evidences operation. Model validation, dataset governance, risk management testing and the provider's conformity assessment remain separate responsibilities, and no runtime record substitutes for them. Treat enforcement as one evidence layer inside a wider model risk management programme.

From obligation to evidence

Label: Classify

Description: Decide which credit-related AI uses evaluate natural persons and which fall under an exception.

Label: Assign

Description: Separate provider duties from deployer duties for each system.

Label: Enforce

Description: Apply policy to individual interactions: allow, redact, block, escalate for human approval.

Label: Evidence

Description: Keep a per-interaction record of the policy applied, the action taken and any human involvement.

Test your evidence against a single credit interaction

See how policy enforced in the execution path produces per-interaction records for human oversight and event recording, and where other controls must carry the load.

See How Evidence Is Produced