What becomes enforceable on 2 August 2026 under the EU AI Act high-risk obligations, where most European financial services firms still have gaps, and what realistic triage looks like in six weeks.
On 2 August 2026 — six weeks from now — the substantive provisions of Chapter III of the EU AI Act become enforceable for high-risk AI systems. This is the most consequential application date in the Act's staged rollout for European financial services. From that day, any AI system in scope of Annex III that is in use across the institution falls under the operational, documentation, and oversight obligations of Articles 8 through 49.
Across the engagements we have reviewed, most institutions are not where they need to be. Not because the work cannot be done, but because the work is unfamiliar — the EU AI Act sits next to existing model risk and ICT frameworks rather than replacing them, and institutions that have not yet mapped their AI estate against Annex III are discovering, late, that systems they thought were out of scope are not. Six weeks is enough time to triage; it is not enough time to build a compliance posture from scratch.
This article covers what specifically becomes enforceable on 2 August 2026, where most financial institutions still have gaps, what realistic triage looks like in six weeks, and what "not ready" actually looks like when a national regulator picks the file up later this year.
The EU AI Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024 and applies in stages. The prohibited practices (Article 5) became applicable on 2 February 2025. The general-purpose AI obligations (Articles 51-55) became applicable on 2 August 2025. From 2 August 2026, the obligations on high-risk AI systems listed in Annex III apply in full.
For financial services specifically, Annex III point 5(b) explicitly covers AI used to evaluate creditworthiness or to establish credit scores, with the carve-out for AI used for the purpose of detecting financial fraud. Annex III point 5(c) covers AI used for risk assessment and pricing in relation to natural persons in life and health insurance. Other Annex III categories (biometrics, employment, essential services) may also apply depending on the institution's AI estate.
Where a system is classified as high-risk, the provider's obligations under Chapter III, Section 2 apply: risk management (Article 9), data and data governance (Article 10), technical documentation (Article 11 and Annex IV), record-keeping (Article 12), transparency to deployers (Article 13), human oversight (Article 14), and accuracy, robustness, and cybersecurity (Article 15). Conformity assessment (Article 43) and registration in the EU database (Article 49) also apply.
For institutions that use third-party AI systems classified as high-risk, the deployer obligations under Article 26 apply: using the system in accordance with its instructions, ensuring input data is relevant, monitoring operation, retaining logs, informing affected natural persons, and conducting fundamental rights impact assessments where applicable under Article 27.
The point most often missed: a financial institution can be in scope as a deployer even when its AI capability comes entirely from a vendor. Whether the institution built the system or bought it, the deployer obligations land on the institution from 2 August.
Five gaps come up consistently across the institutions we have worked with on EU AI Act readiness.
No complete inventory of AI systems mapped against Annex III. Article 28 of DORA produces an ICT third-party register; the EU AI Act expects a different list — every AI system in use, classified against Annex III risk categories, with rationale documented. Most institutions can produce one or the other; few can produce both, and the two registers do not align cleanly because their classification axes are different.
Classification rationale not documented. Where institutions have classified some AI systems as high-risk and others as not, the rationale is frequently absent or thin. The regulator's reading of an undocumented classification is conservative — the assumption is the institution was avoiding the high-risk classification rather than reasoning into it. A classification with a defensible written rationale, including how the Annex III definitions were applied, is a different artefact from one with the box ticked.
Annex IV technical documentation incomplete or in draft. Article 11 and Annex IV set out the technical documentation required for high-risk systems. The list is comprehensive: intended purpose, system architecture, data used for training, performance metrics, risk-management measures, post-market monitoring plan. Most institutions have fragments of this in different places — a model card here, a risk assessment there, a deployment runbook somewhere else — but not a single consolidated Annex IV pack per high-risk system.
Human oversight policy without operational evidence. Article 14 requires high-risk systems to be designed so they can be effectively overseen by natural persons. Most institutions have a policy statement asserting human oversight exists. Few can show, on demand, what the oversight reviewer actually sees, how often they intervene, how long the queue is, and how disagreements are resolved. Policy without operational evidence is the gap regulators surface most quickly. As covered in the regulator-review piece, this is where review findings most often land.
Logs incomplete or not retrievable on demand. Article 12 requires automatic event recording over the lifetime of the high-risk system. Most institutions log model inputs and outputs in some form, but few have logs at the granularity required, retained for the period required, and retrievable per individual case on demand. The audit-trail layer is the deployment work that is most commonly under-built — and the easiest for a regulator to test.
Six weeks is not enough time to design a comprehensive EU AI Act compliance programme. It is enough time to triage the highest-risk gaps and produce defensible documentation for systems that are clearly in scope. The institutions that handle the next six weeks well tend to follow a sequence of four priorities.
Priority 1 — complete the AI system inventory. A list of every AI system in production or active development, with what it does, which business function owns it, which model it uses, and a preliminary Annex III classification. This is the document that frames everything else. Without it, no other work can be prioritised. Most institutions can complete this in a fortnight if they devote the resource.
Priority 2 — finalise classification rationale for in-scope systems. For each system flagged as potentially high-risk in the inventory, document the reasoning that places it in or out of the Annex III categories. Pay particular attention to creditworthiness systems (5(b)) and insurance pricing systems (5(c)) — these are the most commonly contested classifications. Where the classification is borderline, the rationale matters more than the conclusion.
Priority 3 — consolidate Annex IV documentation for high-risk systems. Take the fragments — model documentation, deployment notes, risk assessments — and consolidate them into a single Annex IV-shaped pack per high-risk system. Identify which Annex IV elements are missing entirely and prioritise filling those gaps. A consolidated pack with known gaps is in a better regulatory position than a complete pack that nobody can locate.
Priority 4 — document human oversight and logging in their current state. For each high-risk system, describe what oversight actually looks like operationally — who reviews what, how often, with what authority. Where there are gaps between policy and practice, document them honestly. The same applies to logging: describe what is being recorded today, what is retrievable, and what would not be reconstructable on demand. Honest documentation of an imperfect state is more defensible than aspirational documentation that does not match operations.
What is not realistic in six weeks: building a new model risk framework specifically for AI, completing conformity assessment for any system that requires it, redesigning the deployment architecture, or hiring an EU AI Act compliance team. Those are 12-24 month programmes. The six weeks before 2 August are about being able to defend the position the institution is currently in, not transforming it.
National regulators across BaFin, FINMA, the FCA (where it applies in the UK context, with separate EU AI Act considerations), and the relevant EU member-state authorities will not all initiate AI-focused reviews on 3 August. The realistic timeline for the first wave of post-deadline reviews is the autumn of 2026 through the first half of 2027, as part of broader supervisory cycles. But every interaction with a regulator from 2 August onward — DORA reviews, prudential reviews, Consumer Duty thematic reviews where the FCA applies, MaRisk reviews under BaFin — will increasingly include AI-related questions, and the answers will be measured against EU AI Act expectations.
The institution that is "not ready" presents recognisable signs. The inventory does not exist or is materially out of date. The classification is undocumented or asserted without rationale. The Annex IV technical documentation is described as "in progress" with no timeline for completion. Human oversight is documented as a policy but the operational evidence is absent. Logs are described as "available" but cannot be produced for a specific historical case on demand. As covered in the DORA register piece, these documentation gaps appear consistently across the regulator-facing surface.
The institution that is "ready enough" — not perfect, not transformed, but defensibly working through the obligations — looks different. An inventory exists and is current. Classifications are documented with reasoning. Annex IV packs are partial but identifiable per system, with known gaps and a documented plan to close them. Human oversight has operational evidence as well as policy. Logs are demonstrably retrievable on a sample case-by-case basis.
The gap between "not ready" and "ready enough" is mostly documentation discipline, not technology. The institutions that close it in the six weeks ahead spend the autumn defending a position. The institutions that arrive at 2 August without doing the work spend the autumn under retrospective scrutiny.
The EU AI Act high-risk obligations become enforceable on 2 August 2026. For European financial services, the systems most clearly in scope are those covered by Annex III point 5 — creditworthiness and credit scoring, life and health insurance pricing — and the deployer obligations under Article 26 apply to institutions using third-party high-risk AI even where they did not build it.
Most institutions are not fully ready, primarily because the work is unfamiliar rather than because it is intractable. The gaps cluster in five places: incomplete AI system inventory, undocumented classification rationale, fragmented Annex IV technical documentation, human oversight policy without operational evidence, and incomplete or non-retrievable logs.
Six weeks is enough time for triage, not transformation. The institutions that use the time well complete the inventory, document the classification reasoning, consolidate Annex IV documentation per high-risk system, and describe human oversight and logging in their current operational state — honestly, gaps included. The institutions that try to build a comprehensive compliance posture from scratch in six weeks tend not to finish either.
The realistic regulator response is not an immediate audit on 3 August. It is a series of supervisory interactions across the autumn of 2026 and 2027 in which AI-related questions are increasingly part of the conversation. The institution that has done the documentation work defends its position; the institution that has not finds itself working backwards under scrutiny.
If your institution is working through the EU AI Act high-risk gap analysis in the next six weeks and wants to pressure-test the inventory, the classification rationale, and the Annex IV documentation before the deadline, we can help you work through it.
Related reading:
Stop renting generic models. Start building specialized AI that runs on your infrastructure, knows your business, and stays under your control.