EU AI Act High-Risk Rules for Banks and Insurers in 2027

The EU AI Act's high-risk obligations for banks and insurers no longer start on 2 August 2026. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and moves the deadline for Annex III systems, which include credit scoring and life and health insurance pricing, to 2 December 2027. That is sixteen extra months, and they are now law, not a proposal. This post explains exactly which of your systems are caught, which duties apply to you as a deployer, and how to sequence the work so December 2027 is a formality rather than a scramble.

What the Digital Omnibus actually changed

The AI Act, Regulation (EU) 2024/1689, originally applied its Chapter III high-risk rules from 2 August 2026. The problem, stated bluntly in the Omnibus recitals, was that the harmonised standards, common specifications, guidance, and national competent authorities needed to make those rules work were not ready. Rather than let the deadline collapse into divergent national improvisation, the co-legislators moved it.

Under the amended Article 113, the application dates are now:

• Annex III high-risk systems (creditworthiness and credit scoring of natural persons, risk assessment and pricing in life and health insurance, employment screening, and the rest): 2 December 2027.
• Annex I high-risk systems (AI embedded as a safety component in regulated products): 2 August 2028.
• Prohibited practices: in force since 2 February 2025, with new prohibitions added by the Omnibus applying from 2 December 2026.
• Transparency and general provisions: still applying from 2026 as before. The Omnibus moved the high-risk regime, not the rest of the Act.

Two details matter for planning. First, the Omnibus is published and in force; this is settled law, so you can plan against these dates. Second, systems already on the market or in service before the application date are grandfathered unless they undergo significant design changes, under the amended Article 111. A scoring model you materially retrain or re-architect after December 2027 is no longer legacy. Plan as if every system you actively maintain is in scope.

You can read the amending regulation yourself: Regulation (EU) 2026/1744 on EUR-Lex, recital 40 and the amended Article 113 are the load-bearing provisions.

Which bank and insurer systems are high-risk

The classification lives in Annex III of the AI Act, point 5, and the wording is narrower than most commentary suggests:

• Point 5(b): AI systems used to evaluate the creditworthiness of natural persons or establish their credit score, with an explicit exception for systems used to detect financial fraud.
• Point 5(c): AI systems used for risk assessment and pricing in relation to natural persons in life and health insurance.

Read those carefully and three boundaries appear.

Natural persons only. A model that prices commercial property or corporate liability for a legal entity is not caught by 5(c). A credit model scoring sole traders and consumers is caught by 5(b); a model scoring only corporates sits outside it. Banks and insurers writing mixed books must classify system by system, not programme by programme.

Fraud detection is excluded, AML is unsettled. Pure fraud-detection scoring is expressly carved out of 5(b). Anti-money-laundering transaction monitoring has no equally clear classification either way; treat it as an open question to revisit when guidance lands, not as settled in either direction.

Insurance telematics is the paradigm case. Pay-as-you-drive motor pricing that builds an individual premium from a named driver's behaviour is exactly what 5(c)-style individualisation targets, as the sector analysis at regulation-ai.eu spells out. The same logic catches any personal-lines model that moves from pool-level actuarial pricing to individual AI-driven scores. And note that 5(c) names life and health explicitly; motor telematics is argued in via the general structure of Annex III point 5, which is why a classification memo per system, reviewed by counsel, beats a one-line assumption.

One more system banks forget: AI used in recruitment screening is separately high-risk under Annex III point 4, so the HR tool your retail bank bought is in scope too, independent of anything in the credit book.

Deployer or provider: the question that sets your workload

Most institutions will be deployers: they use a system built by a vendor under their own authority. Deployers carry the duties in Article 26. But two situations flip you into the far heavier provider role under Articles 9 to 15:

• You build the system in-house, or have it built and place it on the market under your own name.
• You use a vendor system outside its instructions for use in a way that substantially modifies it. The regulation-ai.eu analysis is direct on this: deploy a vendor actuarial model outside its validated population or use case and you can inherit full provider obligations.

Provider status means owning the whole lifecycle: risk management system, data governance with active bias detection, Annex IV technical documentation, conformity assessment, registration, post-market monitoring. Deployer status means the six operational duties in the next section. The commercial consequence is simple: your vendor contracts now decide your regulatory workload. Demand the Declaration of Conformity, CE marking, Annex IV documentation, instructions for use with validated-use parameters, and post-market monitoring data as contractual deliverables. Our AI vendor due diligence guide for banks and insurers has a working checklist for exactly this conversation.

The six deployer obligations that actually bite

The banking.vision analysis of high-risk AI in credit institutions distils the deployer duties well. Here they are with the operational content filled in.

1. Use per intended purpose. A model validated for consumer credit cannot be quietly repurposed to small-business lending. Purpose drift is the cheapest compliance failure to commit and one of the easiest for a supervisor to find.

2. Qualified human oversight. A case handler must be able to understand the recommendation well enough to correct it, not rubber-stamp it. This is a training and interface-design requirement, not a policy sentence: the handler needs reason codes, confidence information, and genuine authority to override.

3. Logging and retention. The Act sets a six-month minimum log retention, but sector rules run longer: five to ten years across commercial, tax, securities, and AML record-keeping regimes. Build retention to the longest applicable clock. If your logging layer is not already designed, our post on LLM monitoring and audit logging for banks and insurers covers the architecture.

4. Serious incident reporting. A scoring model that systematically misrejects a protected group is a reportable serious incident under Article 73, to both the provider and the competent authority, without delay. You need an internal definition, a runbook, and someone on call for it.

5. Fundamental rights impact assessment. Article 27 names deployers of Annex III 5(b) and 5(c) systems explicitly: before first deployment you must document your processes, the period and frequency of use, the groups of persons affected, the specific risks of harm to them, your human oversight measures, and what happens when a risk materialises, including complaint handling. The result is notified to the market surveillance authority on the Commission's template. This is the single most bank-and-insurer-specific duty in the Act, and it cannot be outsourced to your vendor because it assesses your deployment context.

6. Individual explanations. Under Article 86, a person refused credit or priced by an AI system can demand a clear, meaningful explanation of the decision's main factors. A black-box model with no reason-code layer is a direct compliance risk here, whatever its AUC says.

Get the fines right, because much of the ranking content does not. Under Article 99: breaching the Article 5 prohibitions costs up to EUR 35 million or 7 percent of worldwide turnover; breaching the high-risk operator obligations costs up to EUR 15 million or 3 percent; supplying misleading information to authorities costs up to EUR 7.5 million or 1 percent. Some explainers still cite EUR 30 million or 6 percent, a figure from an earlier draft, and others imply the 35 million tier applies to credit scoring breaches. It does not. The 15 million or 3 percent tier is the one that hangs over Annex III deployers and providers.

What the extra sixteen months are actually for

A realistic sequence for a mid-size institution with a credit scoring model and a handful of other AI systems, starting now:

Q4 2026: inventory and classify. List every system that infers predictions or decisions from data, including vendor tools and the HR screening tool. For each, record: natural persons affected yes or no, Annex III point, deployer or provider role, intended purpose as documented by the vendor. Expect a retail bank to find that a minority of its systems are high-risk, but that those systems sit on its most regulated journeys.

Q1 2027: gap analysis and vendor paper. For each high-risk system, map current state against the six deployer duties. In parallel, send every relevant vendor the contractual request for conformity documentation, instructions for use, and monitoring data. Vendors will be slow; start early.

Q2 2027: oversight and explanations. Train case handlers, define override authority, and stand up the reason-code and explanation layer for customer-facing decisions. This is the longest-lead internal work because it touches frontline process, not just documentation.

Q3 2027: FRIA and incident runbook. Complete the Article 27 assessment for each 5(b) and 5(c) system using the Commission template, run the serious-incident reporting drill once, and reconcile log retention with the longest sector clock.

Q4 2027: notification and dry run. Notify FRIA results to the market surveillance authority, finalise the registration picture with your providers, and run one end-to-end review as if a supervisor asked for the file. 2 December 2027 should be a date you notice in passing.

Insurers have one extra stream. EIOPA's Opinion on AI governance of August 2025 deliberately excludes AI Act high-risk systems from its scope and instead tells national supervisors how Solvency II and the IDD already apply to all other insurance AI: data governance, record-keeping, fairness, explainability, human oversight. So an insurer's compliance map has two layers, the AI Act for high-risk pricing systems and EIOPA-aligned sectoral governance for everything else, and supervisors will expect both.

Common mistakes to avoid

• Planning against 2 August 2026 anyway. Some teams treat the old date as the safe date. It is not wrong to finish early, but misreading the Omnibus cuts the other way too: teams that half-heard the news sometimes assume everything moved, including the transparency obligations that still apply from 2026.
• Believing "AI Act ready" on a vendor slide. There is no vendor certification that discharges your Article 26 and 27 duties as deployer. The FRIA in particular is yours by law.
• Classifying by system name instead of by use. The same underwriting workbench is high-risk in personal lines and out of scope in corporate lines. Classification follows the population affected, not the product SKU.
• Letting purpose drift turn you into a provider. Repurposing a vendor model outside its instructions for use is a technical decision with provider-grade regulatory consequences. Route it through risk sign-off.
• Quoting the wrong fine tier in the board pack. EUR 15 million or 3 percent is the operative tier for high-risk breaches. Boards remember wrong numbers.

When you can worry less

This regime may genuinely not bite you. If you lend only to corporates, write only commercial lines, and use AI purely for fraud detection and internal productivity, Annex III points 5(b) and 5(c) do not capture those systems. Non-EU institutions with no EU customers are outside the Act's territorial scope, though a Kenyan or East African bank scoring EU-resident customers, or an insurer pricing EU policyholders, is inside it. And even where the Act does not apply, the discipline it demands, inventory, classification, documented oversight, explanations, is the same discipline your own supervisors and the EU's sectoral regulators increasingly expect anyway, which is why the cheapest strategy is usually to build the file once and reuse it.

Frequently asked questions

Did the EU AI Act deadline really move to December 2027? Yes. Regulation (EU) 2026/1744 was published in the Official Journal on 24 July 2026 and is in force. Annex III high-risk obligations, including credit scoring and life and health insurance pricing, apply from 2 December 2027; Annex I embedded systems from 2 August 2028.

Is our credit scoring model high-risk even if a human makes the final decision? Yes. Annex III point 5(b) classifies the system by what it does, evaluating creditworthiness of natural persons, not by who signs the decision. The human-in-the-loop is an obligation (Article 14 oversight), not an exemption.

Does the AI Act apply to anti-money-laundering systems? Fraud-detection systems are expressly excluded from the credit-scoring category, but AML transaction monitoring has no equally clear classification either way. The prudent approach is to document your classification reasoning and watch for Commission guidance.

What is a FRIA and does it apply to us? A fundamental rights impact assessment under Article 27. It applies explicitly to deployers of Annex III 5(b) and 5(c) systems, which is exactly banks scoring credit and insurers pricing life and health risk. You complete it before first deployment and notify the results to the market surveillance authority.

What are the fines for getting this wrong? Up to EUR 15 million or 3 percent of worldwide turnover for breaching the high-risk operator obligations; up to EUR 35 million or 7 percent only for the prohibited practices; up to EUR 7.5 million or 1 percent for misleading the authorities.

We already comply with GDPR and Solvency II or Basel model risk rules. Is that enough? No. The AI Act stacks on top of sectoral law rather than replacing it. EIOPA's 2025 Opinion is explicit that existing insurance legislation continues to apply to AI systems alongside the Act. Model risk documentation helps, and our model risk documentation guide shows the overlap, but it does not substitute for the Act's specific duties. GDPR in particular still governs the personal data flowing into these systems, which is why the pipeline discipline in our PII redaction guide remains relevant for any model that touches customer data.

Where to go from here

The sixteen months to December 2027 are enough time to do this properly: inventory, classify, fix vendor contracts, build oversight and explanations, complete the FRIA. They are not enough time to do it starting in autumn 2027. We implement enterprise AI governance for banks and insurers, including the classification memos, evaluation and monitoring layers, and governance documentation this regime demands, and our Claude for Financial Services work covers the agent side of the same file. If you want a second pair of eyes on your inventory, talk to an engineer.