Autonomous AI Agents in Regulated Financial Services — Governance Framework for Agentic Systems Under DORA and the EU AI Act

This specification provides the first published governance framework for autonomous AI agents deployed in regulated financial services environments subject to Regulation (EU) 2022/2554 (DORA) and Regulation (EU) 2024/1689 (EU AI Act). No existing regulatory instrument — DORA, the AI Act, SR 11-7, MiFID II, the ECB Guide to Internal Models, EBA guidelines, or EIOPA opinions — addresses the governance challenges created by AI systems that autonomously pursue goals, dynamically select tools, observe their environment, plan across multiple steps, and modify external state without per-action human approval.The specification defines three tiers of agent autonomy for financial services (Assistive, Supervised Autonomous, Fully Autonomous) and identifies the governance thresholds at which existing frameworks exhibit gaps (Tier 2) and break entirely (Tier 3). A systematic article-by-article analysis of DORA identifies thirteen provisions where the regulation's implicit application to AI fails for agentic systems, including the inverted threat model (the agent is the actor, not the target), incident onset ambiguity for autonomous actions, unbounded action spaces preventing exhaustive resilience testing, and dynamic tool connections creating Register of Information registration challenges.The specification identifies and analyses the structural paradox between the AI Act's Article 14 human oversight requirements and the autonomous operation that defines agentic systems, proposing an Agent Oversight Architecture with three intervention models (pre-commitment gates, statistical process control with automatic halt, post-action audit with reversal) calibrated to autonomy tier and action reversibility.An Agentic Risk Taxonomy (ART-1 through ART-6) defines six risk categories specific to autonomous agents in financial services: Autonomous Decision Drift (ART-1), Tool Chain Escalation (ART-2), Fiduciary Boundary Violation (ART-3), Autonomous Concentration (ART-4), Cross-System State Propagation (ART-5), and Accountability Void (ART-6). Three categories have no analogue in the compositional risk taxonomy — they are uniquely agentic risks arising from autonomous action, regulatory boundary crossing, and systemic emergence.Four architecture-specific compliance profiles address the dominant agentic deployment patterns in financial services: autonomous trade execution (MiFID II and DORA dual-compliance), AML investigation (fraud detection exemption boundary analysis), claims processing (Annex III life/health insurance pricing boundary), and customer onboarding (creditworthiness assessment boundary and GDPR Article 22). Each profile includes an architecture reference diagram, risk category assessment, regulatory classification analysis across multiple frameworks, DORA compliance requirements, and a safe halt specification.Operational methodologies are provided for three critical compliance functions: a Tool Inventory Registration Methodology mapping agent tool connections to the DORA Register of Information ITS template (Commission Implementing Regulation 2024/2956), an Agent Incident Classification Decision Logic resolving the onset ambiguity problem for autonomous agent actions under the DORA incident reporting framework (CDR 2024/1772), and a Portable Agent Specification enabling credible exit strategies under DORA Article 28(8) for agent-based systems.This document extends the Compositional Risk Taxonomy (CRT-1 through CRT-8) established in the CRSA-1 EU Edition and is part of the CRSA-1 EU Compliance Series within the Auburn Governance Stack.