EU AI Act for Financial Institutions: The 2027 Deadline and What to Do Now

If you run credit scoring, creditworthiness assessment or AI-assisted lending decisions in the EU, your systems are high-risk under Annex III and that has not changed. What did change is the clock: the Digital Omnibus moved the high-risk compliance date from August 2026 to 2 December 2027. Read carefully, because the deferral is narrower than the relief it appears to offer — several obligations are already in force, and the work that takes longest was never the part that was postponed.

Published: March 2026 · updated August 2026 · Author: Dominic Fui Dodzi-Nusenu · Reading time: ~8 minutes

Disclosure: This article was drafted using generative AI assistance and subsequently edited, verified, and expanded by human legal and tech policy specialists. The organizational case study presented below is a composite hypothetical scenario designed to illustrate compliance risks.

Why financial services sits at the centre of this

The AI Act is horizontal — it regulates by use case, not by industry. Financial services nonetheless attracts more of it than almost any other sector, because several of the uses named in Annex III are things banks and insurers do as core business rather than as experiments.

Note the last column. Buying a model does not move the obligation elsewhere. A deployer of a high-risk system carries duties of its own under Article 26 — human oversight, input data relevance, log retention, and monitoring — and those cannot be contracted away to the vendor. This is the single most common misreading we encounter.

What actually changed in 2026, and what did not

The Digital Omnibus deferral was widely reported as « the AI Act being delayed ». That framing is wrong in a way that costs money. Only the high-risk tier moved.

So an institution that reads « delayed » and stands down has stood down from nothing that was actually deferred, while remaining exposed on four sets of obligations that are enforceable now. The AI literacy duty in particular is frequently overlooked: it applies to every organisation using AI, at every risk tier, with no threshold.

Case scenario: how a mid-sized lender ends up in scope

A German consumer-credit provider with 180 staff licenses a scoring model from a vendor, retrains it quarterly on its own repayment data, and runs it behind a manual review desk. Management’s view is that the vendor is the provider and the desk satisfies human oversight. Three things are wrong with that.

1. Retraining on own data can make you the provider. Substantially modifying a high-risk system, or placing it on the market under your own name, transfers provider obligations to you under Article 25 — including Annex IV technical documentation for the modified system.
2. A review desk is not automatically Article 14 oversight. Oversight must be exercised by people with the competence, authority and information to override the system. A desk that sees a score and a recommendation, with no view of the drivers and no realistic mandate to reject, is a rubber stamp with a headcount.
3. Quarterly retraining without drift records is the gap that closes last. Article 72 post-market monitoring expects evidence over time. Starting to collect it in late 2027 leaves you with a few months of history for a system that has run for years.

This scenario is a composite, but each of its three failure modes is one we see repeatedly — and the third is the one that cannot be fixed by working harder closer to the deadline.

The three gaps that actually delay financial institutions

An AI inventory that stops at the models the CIO knows about

Almost every institution we work with underestimates its own AI footprint, usually by a wide margin. The gap is rarely the flagship scoring model; it is the vendor product with a machine-learning feature nobody classified, the analytics tool that added a generative assistant in a release note, and the department that solved its own problem with a spreadsheet and an API key. You cannot classify what you have not found, and classification is the first gate.

Developer documentation mistaken for statutory evidence

Model cards in a wiki, notebooks in a repository and a data dictionary in Confluence are good engineering practice and they are not Article 11 technical documentation. Annex IV specifies content, and supervisory evidence has to be version-controlled, attributable, and reconstructable as it stood on a given date. The test is simple: if a supervisor asks what your model looked like in Q3 last year, can you produce it without reconstructing it from memory?

Treating the deadline as the project plan

Articles 9, 15 and 72 all describe continuous processes, not deliverables. A risk management system that was stood up the month before the deadline satisfies the letter of nothing — it has no iterations, no identified-and-mitigated risks, and no monitoring history to show. This is why the deferral to December 2027 is less generous than it sounds.

A sequenced 90-day start

You do not need a multi-year programme to reach a defensible position. You need to do the sequence-dependent work first.

1. Days 1–20 — inventory. Enumerate every AI system in use, including embedded vendor features and anything staff adopted directly. Record purpose, data, owner and vendor.
2. Days 21–40 — classify. Map each system to prohibited, high-risk (Annex III), transparency-only (Article 50) or minimal. Record the reasoning and the provision, not just the answer — the reasoning is what a supervisor tests.
3. Days 41–60 — close the live duties. Article 5 prohibitions, Article 4 AI literacy and Article 50 transparency are enforceable now. Finish these before touching 2027 work.
4. Days 61–75 — start the clock on evidence. Turn on logging, drift monitoring and bias measurement for high-risk systems. Every week you delay is a week of history you will not have.
5. Days 76–90 — gap-assess against Annex IV. Now you know what you have and what it must become, size the documentation work honestly.

Related reading

• German AI Regulation: Navigating BaFin, DORA and the KI-MIG
• EU AI Act Compliance Checklist for High-Risk AI Systems
• Understanding AI Bias in Financial Services
• Shadow AI: The Hidden Risk in Your Organization

Frequently asked questions

Has the EU AI Act been delayed?

Only the high-risk tier. Standalone Annex III systems now apply from 2 December 2027 and Annex I embedded systems from 2 August 2028. The prohibited-practice ban and AI literacy duty (2 February 2025), GPAI obligations (2 August 2025) and Article 50 transparency (2 August 2026) are all in force and unaffected.

We license our scoring model from a vendor. Are we still liable?

Yes. Deployers of high-risk systems carry independent obligations under Article 26 covering human oversight, input data relevance, log retention and monitoring. In addition, if you retrain the model on your own data or put it on the market under your own name, Article 25 can make you the provider, which brings the full Annex IV documentation duty with it.

What are the maximum penalties?

Up to 35 million euros or 7 percent of total worldwide annual turnover, whichever is higher, for prohibited practices. Up to 15 million euros or 3 percent for most other breaches, including high-risk obligations. Supplying incorrect or misleading information to authorities carries its own tier.

Does BaFin supervision replace the AI Act?

No — they stack. BaFin’s MaRisk and BAIT expectations continue to apply to your institution, and Germany’s KI-MIG — ce que cela signifieKI-MIGKI-Marktüberwachungs- und Innovationsförderungsgesetz — loi allemande de mise en œuvre de l'EU AI ActLa loi allemande qui transpose l'EU AI Act en droit national, en vigueur depuis le 29 juillet 2026. Elle désigne la Bundesnetzagentur comme autorité centrale de surveillance du marché de l'IA en dehors des secteurs réglementés, et y établit un centre de coordination, le KoKIVO. La BaFin conserve le mandat sectoriel pour les systèmes d'IA utilisés en lien direct avec une activité financière réglementée, ce qui explique qu'une banque et une société de logement relèvent d'autorités différentes pour un même type de modèle.S'applique àFournisseurs et déployeurs de systèmes d'IA supervisés en Allemagne. L'autorité compétente dépend du secteur.Lire le texte source (s'ouvre dans un nouvel onglet)Information générale sur l'instrument cité, ne constituant pas un conseil juridique.Fermer designates market surveillance for the AI Act itself. Institutions should expect to evidence the same systems against both frameworks rather than choosing one.

Is an internal model validation report enough for Article 11?

Usually not on its own. Validation reports address model quality; Annex IV specifies a broader content set including intended purpose, system architecture, data governance, risk management, and post-market monitoring arrangements. Validation output is an input to that documentation, not a substitute for it.

Where should we start if we have done nothing?

Inventory, then classification. Everything else depends on knowing which systems you have and which tier each falls into, and both are cheap relative to the documentation and monitoring work they scope.