All posts
  • eu ai act
  • credit scoring
  • creditworthiness assessment
  • high-risk ai
  • banking
  • ai agents
  • digital omnibus
  • fria
  • consumer credit directive
  • regulatory compliance

EU AI Act Credit Scoring: What Banks Must Do Before December 2027

Creditworthiness AI is high-risk under the EU AI Act, with obligations from 2 December 2027. Here's what's in scope (including LLM agents), who does what, what the Omnibus changed, and a 10-point readiness checklist.

By Aodit

13 min read

Credit scoring is the use case in the EU AI Act that hits banks most directly. Annex III, point 5(b) names "AI systems intended to be used to evaluate the creditworthiness of natural persons or establish their credit score" as high-risk. Since the Digital Omnibus on AI (Regulation (EU) 2026/1744), the high-risk obligations for these systems apply from 2 December 2027.

That can sound like a long runway. It isn't, for three reasons:

  1. The second Consumer Credit Directive (CCD2) applies from 20 November 2026, three days from now. It already gives consumers rights over automated creditworthiness assessments.
  2. Lenders are moving AI into the lending journey, from LLM assistants that collect applicant data to agents that summarise affordability for underwriters. Many of those deployments have never been checked against Annex III.
  3. The AI Act requirements are evidence-heavy. Risk management, testing, logging and human oversight need months of operating history before a supervisor will find them credible.

This guide covers what is in scope, who carries which obligations, what the Omnibus changed, and a 10-point checklist risk and compliance teams can start on now.

Note: This article is general information, not legal advice. Where we interpret the regulation, we say so.

What Annex III point 5(b) covers

The text is short, so the details of the wording do a lot of work.

  • "Natural persons" only. Point 5(b) covers creditworthiness and credit scoring for individuals. Our reading: lending to companies (legal persons) falls outside 5(b) as written. Sole traders and other individuals borrowing for business purposes are still natural persons, so SME lending desks should not assume they are out of scope.
  • Fraud detection is excluded. The annex carves out "AI systems used for the purpose of detecting financial fraud". A model that scores fraud risk on a loan application is not a 5(b) system on that basis alone. A model that mixes fraud signals into a creditworthiness score is a harder case.
  • Life and health insurance is separate. Point 5(c) covers risk assessment and pricing for natural persons in life and health insurance. That matters for bancassurance groups.

The Article 6(3) filter rarely helps in credit

Article 6(3) lets a provider conclude that an Annex III system is not high-risk if it does not materially influence decision outcomes. Examples are a narrow procedural task, improving the result of a completed human activity, or a preparatory task.

There is a hard stop. An Annex III system "shall always be considered to be high-risk where the AI system performs profiling" of natural persons. Credit scoring is close to the textbook definition of profiling. In its commentary on the Commission's draft high-risk classification guidelines (published 19 May 2026), Osborne Clarke notes that systems in areas like credit assessment "usually perform profiling in some way". The filter is therefore unlikely to apply to most credit models.

If you do rely on Article 6(3), Article 6(4) requires you to document the assessment before placing the system on the market or putting it into service, and to register it. Commentary on the Omnibus indicates that registration remains, with a lighter filing.

Where AI agents and LLMs enter the picture

Most credit teams think of Annex III in terms of scorecards and gradient-boosted PD models. The newer exposure is generative and agentic AI placed around the decision. A few patterns we see, with our view of how each is likely to be treated:

DeploymentLikely treatment (our interpretation)
LLM chatbot answering general questions about loan productsNot 5(b) on its own. Article 50 transparency still applies (disclose AI interaction, applicable since 2 Aug 2026).
Agent that collects applicant data and pre-qualifies or soft-declines applicantsLikely 5(b): it evaluates creditworthiness and can materially influence the outcome.
LLM that summarises bank statements and flags affordability concerns for an underwriterGrey zone. "Preparatory task" could apply, but if the summary steers decisions and involves profiling, assume high-risk.
Agent that recommends limit increases or collections treatment from behaviour dataLikely profiling and creditworthiness evaluation, so assume high-risk.

Two practical points follow.

First, you may become the provider. Under Article 25, a deployer is treated as the provider of a high-risk system if it puts its name on the system, makes a substantial modification, or changes the intended purpose of a system (including a general-purpose AI system) so that it becomes high-risk. A bank that takes a general-purpose LLM and builds it into a pre-qualification agent is very likely the provider of that agent. The provider carries the full set of Articles 9–15 obligations, not the lighter deployer duties.

Second, agent behaviour is part of the risk. A scorecard gives the same output for the same input. An LLM agent may not. It can be talked out of policy, misstate eligibility criteria, or skip an escalation when a customer pushes back. Those behaviours bear directly on Article 15's accuracy and robustness requirements and on Article 14's human oversight requirements.

This is the gap AODIT was built to measure. AODIT stress-tests AI agents across six behavioural dimensions (Reliability, Integrity, Confidentiality, Judgment, Resistance and Resilience) using adversarial multi-turn conversations. You get an independent view of how a lending or servicing agent behaves under pressure before a supervisor asks.

Provider vs. deployer: who does what

The EBA's November 2025 mapping of the AI Act against banking law makes the same distinction. A bank that develops a credit AI system in-house is both provider and deployer. A bank that buys one is a deployer only.

Provider obligations (in-house models, or agents you built or repurposed)

  • Risk management system (Art. 9): a continuous, documented process. It includes testing against "prior defined metrics and probabilistic thresholds" before market placement.
  • Data and data governance (Art. 10): relevance, representativeness, error handling and bias examination of training, validation and test data.
  • Technical documentation (Art. 11) and record-keeping/logging (Art. 12).
  • Transparency and instructions for use (Art. 13) for deployers.
  • Human oversight by design (Art. 14).
  • Accuracy, robustness and cybersecurity (Art. 15), including resilience to attempts to manipulate the system.
  • Quality management system (Art. 17), conformity assessment (Art. 43), CE marking and EU database registration (Art. 49), plus post-market monitoring (Art. 72) and serious-incident reporting (Art. 73).

Deployer obligations (purchased or vendor systems)

  • Use the system according to its instructions, with appropriate technical and organisational measures (Art. 26(1)).
  • Assign human oversight to people with the necessary competence, training and authority (Art. 26(2)).
  • Ensure input data you control is relevant and sufficiently representative (Art. 26(4)).
  • Monitor operation, suspend use and notify the provider if a risk emerges, and report serious incidents (Art. 26(5)).
  • Keep automatically generated logs for at least six months (Art. 26(6)).
  • Inform applicants that they are subject to a high-risk AI system (Art. 26(11)).
  • Give clear and meaningful explanations of the AI system's role on request, where a decision significantly affects someone (Art. 86).

The FRIA obligation that is specific to credit

Article 27 normally requires a fundamental rights impact assessment (FRIA) only from public bodies and private entities providing public services. It explicitly extends to deployers of Annex III point 5(b) and (c) systems. Private lenders and life/health insurers are named directly.

The FRIA must describe:

  • the processes where the system is used, and the period and frequency of use;
  • the categories of people affected and the specific risks of harm to them;
  • the human oversight measures;
  • what happens when risks materialise, including governance and complaint mechanisms.

The results must be notified to the market surveillance authority. Where a GDPR DPIA already covers some elements, the FRIA complements it rather than duplicating it.

Financial-sector carve-outs: what is deemed covered, and what is not

The AI Act recognises that banks already operate under heavy governance rules. Several obligations are deemed fulfilled through compliance with EU financial services law:

  • Quality management (Art. 17(4)): deemed met by complying with internal governance rules under financial services law. The exceptions are risk management (Art. 9), post-market monitoring (Art. 72) and serious-incident reporting (Art. 73), which still need specific handling.
  • Deployer monitoring (Art. 26(5)) and log retention (Art. 26(6)) can be satisfied within existing internal governance and documentation regimes.
  • Supervision (Art. 74(6)): for AI systems directly connected to financial services, the market surveillance authority is in principle the national financial supervisor, unless a Member State designates otherwise.

The EBA concluded that it found "no significant contradictions" between the AI Act and EU banking law, and saw no immediate need for new EBA guidelines. That does not mean existing model risk frameworks cover everything. Article 9 risk management, Article 14 oversight design, Article 15 robustness against manipulation and the FRIA are where we see the largest gaps, particularly for LLM-based components that traditional model validation was not designed for.

What the Digital Omnibus changed (and what it didn't)

Regulation (EU) 2026/1744 was adopted on 8 July 2026 and entered into force on 27 July 2026. For credit AI, the key points are:

  • New date: Annex III high-risk obligations apply from 2 December 2027 (previously 2 August 2026).
  • Legacy systems: Article 111(2) keeps a grandfathering rule. High-risk systems already placed on the market or put into service before the application date are only caught if they later undergo significant changes in design. The Omnibus clarified that this turns on at least one unit of that type and model having been lawfully placed on the market before the date. Our view: a retrained or re-architected credit model, or an LLM agent swapped to a new underlying model, may well count as a significant design change. Don't build your plan on grandfathering.
  • Bias detection data: a new Article 4a provides an exceptional legal basis to process special categories of personal data where strictly necessary to detect and correct bias, subject to safeguards. It creates a permission, not an obligation.
  • AI literacy (Art. 4): softened from ensuring a sufficient level to taking measures to support it.
  • Unchanged: the Annex III 5(b) classification itself, the core Articles 9–15 requirements, and the FRIA obligation for credit deployers.

The other rules that already apply

Waiting for December 2027 ignores the rules that apply sooner.

  • CCD2 (Directive (EU) 2023/2225) applies from 20 November 2026. Under its creditworthiness provisions, consumers have the right to human intervention when an automated assessment leads to rejection, to a clear explanation of the assessment, and to a procedure to contest the decision. Special-category data such as health data may not be used. Check your national transposition for exact wording.
  • GDPR Article 22 restricts solely automated decisions with legal or similarly significant effects. In SCHUFA (C-634/21, December 2023), the CJEU held that generating a credit score can itself be an automated decision where third parties draw strongly on it.
  • EU AI Act Article 50 transparency duties, such as telling people they are interacting with an AI system, have applied since 2 August 2026.
  • DORA applies to the ICT stack behind any AI system, including LLM vendors. See our guide to DORA and AI agents.

For an AI agent in the lending journey, this means the human-intervention, explanation and transparency questions are live now, before Annex III obligations start.

A 10-point readiness checklist for December 2027

  1. Inventory every AI touchpoint in credit. Include origination, limit management, collections, pricing and any LLM or agent that interacts with applicants or underwriters. Note whether each one evaluates natural persons.
  2. Classify each system against Annex III 5(b)/(c) and record your Article 6(3) reasoning. Assume profiling means high-risk.
  3. Determine your role. Are you provider, deployer or both? Check Article 25 for any agent built on a general-purpose model.
  4. Map AI Act requirements to existing controls. Use EBA internal governance, loan origination and model risk frameworks. Flag the items that are not deemed covered: Art. 9, 72 and 73.
  5. Define testing metrics and thresholds before deployment (Art. 9). For LLM agents, cover factual accuracy on product terms, policy adherence under pressure and consistency across reworded inputs, not just aggregate accuracy.
  6. Test for manipulation and robustness (Art. 15). Applicants and fraudsters will try to argue, reframe or socially engineer an agent into a better outcome. See our prompt injection testing guide.
  7. Design and test human oversight (Art. 14 and 26(2)). Set escalation triggers, override authority and automation-bias controls. Make sure the CCD2 human-intervention route actually works. Our human oversight guide covers HITL vs HOTL design.
  8. Run bias and data-quality analysis on training and input data (Art. 10, 26(4)). Use the new Article 4a basis only where strictly necessary.
  9. Draft the FRIA (Art. 27) alongside your DPIA, and plan the notification to the authority.
  10. Build the evidence trail. That means logs, explanation templates (Art. 86 and CCD2), applicant notices (Art. 26(11)), post-market monitoring and incident reporting, retained in line with financial-sector documentation rules.

How AODIT supports this work, and where it doesn't

We want to be precise about fit.

AODIT helps with the behavioural layer of AI agents and LLM applications in lending and servicing journeys. Our testing protocol runs agents through structured adversarial conversations: baseline, reworded, complicated, grey-zone, hard push, manipulation, recovery and final judgment. It scores dimensions that map directly to the checklist above:

  • Reliability and Integrity: factual consistency, contradiction handling, uncertainty disclosure and resistance to false certainty. This is relevant to Art. 15 accuracy and to hallucinated eligibility rules (see our hallucination testing guide).
  • Judgment: proportional action, escalation to human oversight, policy-boundary adherence and blast-radius limitation. This is relevant to Art. 14 and CCD2 human intervention.
  • Resistance: jailbreak, prompt injection, social engineering, authority spoofing and adversarial reframing. This is relevant to Art. 15 robustness against manipulation.

Results come as a PDF-ready report that can sit in your Article 9 testing evidence and your FRIA. For sensitive data, AODIT can be deployed on-premise or air-gapped, so no applicant data leaves your environment.

What AODIT does not do: it does not perform conformity assessment, statistical fairness testing of scorecards, or legal classification of your systems. Its evaluations are advisory and do not constitute regulatory approval. Use it alongside your model validation, data-governance and legal teams, not in place of them.

You can see how an agent performs in a free two-minute live demo. Paste your agent's instructions and watch an 8-turn adversarial run scored turn by turn.

Key takeaways

  • Creditworthiness assessment and credit scoring of natural persons are high-risk under the EU AI Act, with obligations applying from 2 December 2027.
  • Profiling means the Article 6(3) exit is rarely available for credit systems.
  • Banks that build agents on general-purpose models are likely providers, with the full set of Articles 9–15 obligations.
  • Credit deployers must run a FRIA, an obligation that otherwise applies mainly to the public sector.
  • The financial-sector carve-outs are partial. Risk management, post-market monitoring and incident reporting still need specific work.
  • CCD2 rights on automated credit decisions apply from 20 November 2026. Human intervention and explanation need to work now.

FAQ

Is credit scoring high-risk under the EU AI Act?

Yes. Annex III point 5(b) classifies AI systems intended to evaluate the creditworthiness of natural persons or establish their credit score as high-risk, except systems used to detect financial fraud.

When do the high-risk obligations for credit scoring apply?

From 2 December 2027, following the Digital Omnibus on AI (Regulation (EU) 2026/1744). The original date was 2 August 2026.

Does the AI Act apply to business lending?

Point 5(b) refers to natural persons. In our reading, lending to legal entities is outside 5(b) as written, but lending to sole traders and other individuals is not.

Do banks need a fundamental rights impact assessment for credit AI?

Yes. Article 27 requires deployers of Annex III point 5(b) and 5(c) systems to carry out a FRIA before deployment and to notify the results to the market surveillance authority.

Is an LLM chatbot used in lending a high-risk AI system?

It depends on what it does. A chatbot answering general product questions is not, on its own, a creditworthiness system. An agent that pre-qualifies or declines applicants, or materially steers underwriting decisions, likely is. A bank that repurposes a general-purpose model for that role may become its provider under Article 25.

Who supervises AI Act compliance for banks?

For AI systems directly connected to financial services, Article 74(6) designates in principle the national authority responsible for the financial supervision of those institutions, unless a Member State designates another authority.

Prepare your evidence before your supervisor asks

December 2027 is one model-validation cycle away, and CCD2 is already here. The institutions that arrive ready will be the ones that can show, with evidence, how their credit AI behaves when applicants push, data is messy and humans need to step in.

Request an independent AODIT evaluation of the AI agents in your lending and servicing journeys. You can also start with the free live demo, compare plans, or read how AODIT approaches EU AI Act readiness.


Sources

  1. 11 min read

    Excessive Agency in Banking AI Agents: Permissions and Blast Radius

    Once an AI agent can issue refunds or move money, its permissions become your biggest risk. Here's what OWASP, the OCC, FSB and DORA say about excessive agency, and a 6-step framework to limit an agent's blast radius.

    • excessive agency
    • ai agent permissions
    • least privilege
    • blast radius

    Read

  2. 14 min read

    LLM Model Change Management for Banking AI Agents: A Retest Playbook

    Vendor model retirements, alias moves and prompt edits can quietly change how a banking AI agent behaves. Here's what the OCC, DORA, the EU AI Act and the PRA say about change, and a 7-step playbook to retest agents before the switch.

    • model change management
    • llm model updates
    • ai agents
    • banking

    Read

  3. 13 min read

    Human Oversight for AI Agents in Banking: HITL, HOTL and Escalation

    The FSB, FCA and OCC all said it in autumn 2026: human oversight of AI only counts if it's designed, resourced and tested. Here's what they expect, HITL vs HOTL, and a 6-step framework for escalation that holds under pressure.

    • human oversight
    • human in the loop
    • ai agents
    • banking

    Read