The AI Act Deadline for Credit Scoring Already Passed. Is Your Agent Still Legal?
Most compliance calendars still mark the EU AI Act's high risk application date as a future milestone. It is not. Article 113 set the general application date for Annex III high risk systems at 2 August 2026. That date is now more than two months behind us. If your institution runs an AI system that evaluates creditworthiness or generates a credit score, and it has not gone through a conformity assessment, you are not approaching a deadline. You are already past it, and the exposure is live, not theoretical.
This matters because credit scoring is not a gray area under the Act. Annex III, point 5(b) places AI systems used to evaluate the creditworthiness of natural persons, or to establish their credit score, squarely in the high risk category. The only carve out in that same list is for AI used purely to detect financial fraud, which sits in a different clause entirely. If your model touches creditworthiness, affordability, or default probability, fraud detection language in your documentation will not move it out of Annex III.
The carve out most teams think they qualify for
Article 6, paragraphs 3 and 4, do offer a narrow task exemption. An AI system that performs a narrow procedural task, improves a previously completed human activity, or detects patterns without materially influencing a decision can fall outside Annex III even if it technically matches one of the use cases on the list. A lot of credit teams have read that paragraph and concluded their scoring assistant qualifies, because it only produces a recommendation and a human analyst signs off.
Here is the detail that gets missed. The same article states, without much room for interpretation, that the narrow task carve out does not apply if the system performs profiling of natural persons. Profiling here uses the GDPR definition: any automated processing of personal data to evaluate aspects relating to a person, including their economic situation, reliability, or behavior. Most credit scoring systems profile applicants by definition. That is the entire point of a scoring model. The moment your system builds a behavioral or financial profile to arrive at its score, the narrow task exemption is void, regardless of how much human sign off sits downstream of it. A recommendation engine built on profiling is a high risk system wearing an advisory costume.
This is not a drafting ambiguity institutions can argue around in an audit. It is a structural feature of how Article 6 was written. If your legal review of the carve out stopped at "it only makes recommendations" without separately testing whether the underlying model profiles applicants, that review was incomplete.
What a compliant system actually needs
For systems that are correctly, unavoidably high risk, the obligations are concrete. Article 14 requires effective human oversight, specifically the practical ability for a human to intervene or halt the system's operation, not a dashboard a compliance officer glances at once a quarter. A rubber stamp process where a human technically reviews every output but has no real ability or incentive to override the model does not satisfy Article 14's intent, and regulators examining automation bias in human in the loop setups have already flagged this exact failure pattern elsewhere in financial services oversight.
Beyond oversight, high risk credit scoring systems carry conformity assessment obligations, documentation requirements under Article 12 covering logging and traceability, and risk management obligations that need to be demonstrable, not just asserted in a policy document. None of this is new guidance. It has been the law since 2 August 2026, and for institutions that built their AI governance roadmap around "we have until next year," that roadmap is now wrong.
Where DORA adds a second layer
There is a compounding issue for any institution that bought or licensed its credit scoring model from a third party rather than building it in house, which describes most mid sized lenders and EMIs. DORA, in force since 17 January 2025, requires under Articles 28 through 30 a complete register of ICT third party contracts, a pre contract concentration risk assessment, and for functions classified as critical or important, a mandatory exit strategy with unrestricted audit rights written into the contract itself.
A credit scoring vendor relationship almost always qualifies as a critical or important function. If your MSA with that vendor predates DORA's binding date and has never been revisited, you likely have a compliance gap on the vendor contract side that sits underneath the AI Act gap on the model side. Two separate regulations, two separate audit trails, and both point at the same relationship.
This also cuts the other direction for any firm, including Novafin, that provides AI enabled scoring or decisioning capability to a client under contract. The provider becomes an ICT third party under the client's DORA obligations, which means the provider's own contracts, exit clauses, and audit rights language need to hold up to the same Article 30 minimum content requirements the client is being assessed against. If you supply decisioning models to banks or EMIs, this is worth checking in your own paper before a client's auditor finds it first.
What to check this week, not next quarter
Three questions separate institutions with real exposure from institutions with a defensible position. Does any production system evaluate creditworthiness, affordability, or credit score, even as one input among several? Does that system profile applicants as part of generating its output, which almost certainly voids any narrow task exemption your legal team relied on? And for every vendor supplying that capability, does the contract meet DORA Article 30's minimum content, including a genuine exit strategy and audit rights, not boilerplate language from before 2025?
None of these questions require a six month remediation program to answer. They require someone pulling the actual model documentation and the actual vendor contract and reading them against the actual regulatory text, this week. The institutions that will have a problem in an upcoming supervisory review are not the ones with imperfect systems. They are the ones who have not yet done this review at all, because the deadline still feels like it is in the future.