EU AI Act Annex III: which financial services AI is actually high-risk?

By Scanlex  ·  30 March 2026  ·  9 min read

The EU AI Act is now in force. Annex III sets out the categories of AI systems classified as high-risk, and several of them map directly to the AI tools that regulated financial institutions are already using or building. Understanding whether your AI systems fall within scope is the first compliance step, and for many firms it is not as straightforward as it seems.

This article explains which AI systems in financial services are explicitly listed in Annex III, how the provider vs deployer distinction affects your obligations, what the AML fraud detection carve-out actually says, and what timeline you are working against.

What Annex III actually says about financial services AI

Annex III lists eight categories of high-risk AI. Three of them directly affect regulated financial institutions.

Provider vs deployer: why the distinction matters

The EU AI Act divides obligations between two roles: providers (who develop and place AI systems on the market) and deployers (who use third-party AI systems in a professional context). The obligations are significantly different.

The provider/deployer boundary is not always obvious. Fine-tuning a third-party model on your own data, substantially configuring an AI tool, or combining multiple AI components into a system you place on the market under your name may make you a provider under Article 25, regardless of whether the underlying components were built by someone else. This determination requires legal analysis applied to the specific facts of your AI development and deployment.

The AML fraud detection carve-out: what it actually says

Recital 56 of the EU AI Act provides that AI systems used specifically for the purpose of detecting financial fraud and AML transaction monitoring are not classified as high-risk under Annex III point 5. This carve-out is widely referenced but frequently misapplied.

What the carve-out covers: AI transaction monitoring systems, AI-assisted sanctions screening, and AI fraud detection tools used purely in the context of AML/CFT compliance functions.

What the carve-out does not cover:

• AI systems that influence credit decisions, even if fraud risk is one of the inputs
• AI systems used for customer onboarding risk scoring where creditworthiness is assessed alongside fraud risk
• AI used for insurance underwriting where both fraud and risk pricing are outputs
• Any AI system where Annex III point 5(b) or 5(c) coverage applies to any function the system performs

The carve-out requires that the system is used specifically for AML/fraud detection, not that it incidentally touches AML or fraud. A system with multiple outputs, only one of which is a fraud flag, is not automatically carved out on the basis of that one output.

What high-risk classification means: the full obligation set

For a system classified as high-risk under Annex III, the following obligations apply before deployment (or before continued operation if the system was already deployed when obligations came into force):

• Technical documentation prepared to Annex IV standards: intended purpose, architecture, training data, performance metrics, limitations, human oversight measures
• Quality management system documented to Article 17: policies and procedures for ongoing compliance, risk management, data governance, post-market monitoring
• Risk management system under Article 9: identification, analysis and mitigation of risks to fundamental rights and safety throughout the AI lifecycle
• Conformity assessment completed before deployment: either internal (for most financial services AI) or third-party (for biometric and critical infrastructure AI)
• EU AI database registration under Article 71
• Fundamental Rights Impact Assessment (FRIA) under Article 27, for deployers of credit scoring and insurance pricing AI
• Human oversight measures implemented under Article 14: ability for a human to override, intervene, or withdraw the system from operation
• Post-market monitoring under Article 72: ongoing collection of data on system performance and adverse incidents

The enforcement timeline

August 2026 is not a distant deadline. Building a complete Annex III compliance programme, including technical documentation, QMS, FRIA, and conformity assessment preparation, typically takes 8 to 16 weeks once a decision to proceed is made. Firms that have not started the classification and scoping process have less runway than they may assume.

The GDPR and AML interaction: why integrated advice matters

For regulated financial institutions, the EU AI Act does not exist in isolation. Credit scoring AI interacts with GDPR Article 22 (automated individual decision-making). Insurance pricing AI interacts with GDPR legitimate interest assessments and DORA operational resilience requirements. AML transaction monitoring AI sits at the boundary of the Annex III carve-out and must be assessed against the specific functions the system performs.

Handling these frameworks through separate advisors, one for AI Act, one for GDPR, one for AML, creates gaps at every seam. A classification opinion that ignores the GDPR interaction is incomplete. An AI Act FRIA conducted without understanding the AML implications of the AI system being assessed will miss material obligations.

For regulated firms that also need AML compliance support alongside AI compliance advisory, see our KYC / ODD Team Outsourcing and AML Audit and Advisory services.