Cédric Jadoul, Partner, Intelligent Automation & Agentic AI, Deloitte Luxembourg 

Abderrahmane Saber, Partner, Banking Operational Efficiency, Deloitte Luxembourg

The inflection point

The banking industry is at a pivotal moment. For years, institutions layered digital tools and automation onto legacy processes to boost efficiency. While this approach delivered incremental value, it also introduced complexity and operational fragility.

Now a fundamental evolution is underway. Agentic AI that reason, plan, act, and adapt—is elevating automation from simple task execution to true operational autonomy. 

Unlike traditional Robotic Process Automation (RPA) or assisted AI, agentic systems dynamically orchestrate workflows, interpret unstructured data, and pursue defined goals with minimal human supervision. They do not wait for prompts: they autonomously monitor conditions, detect triggers, and initiate predefined actions.

This progression represents an important redesign of banking operations. When embedded into core processes, agentic AI unlocks a future where routine activities occur autonomously, reliably, and in real time. This is "ZeroOps": an operating model where manual intervention is the exception, not the norm.

For example, trade confirmations that once required manual reconciliation can be self-matched by agents that negotiate terms and flag only genuine exceptions. Regulatory reporting that previously took weeks of data aggregation can be turned into real-time submissions. Client onboarding, which used to rely on sequential handoffs, can instead flow as a single orchestrated process, drastically reducing lead times.

From task automation to autonomous operations

Banks have progressed through waves of automation: digitalizing client touchpoints, using RPA for repetitive back-office work, and applying assisted AI for decision support. But those technologies still rely on human orchestration.

Agentic AI adds a vital new dimension: goal-driven behavior. For example, a system is given the objective to "complete a cross-border wire transfer in compliance with anti-money laundering (AML) regulations". It will plan the steps, interact with data sources, and execute across domains without explicit scripting.

Three architectural patterns are emerging as dominant frameworks:

• Workflow patterns assign agents sequential roles. In securities settlement, one agent validates trade details, another confirms counterparty instructions, a third manages collateral, and a supervisor agent orchestrates the full cycle. Discrepancies trigger dynamic rerouting, only escalating to specialists if automated resolution fails.

• Swarm patterns deploy multiple agents to collaborate on complex problems. In AML monitoring, specialized agents simultaneously examine behavioral patterns, geographic anomalies, network relationships, and transaction structuring. The swarm converges on suspicious activity through collective reasoning, producing comprehensive and contextual investigations.

• Graph patterns organize agents around knowledge structures. In capital adequacy reporting, agents navigate interconnected regulatory taxonomies and trace calculation dependencies across risk-weighted assets. They systematically propagate updates when regulations change, eliminating the need for manual recalibration.

Where agentic AI and ZeroOps deliver value

Client lifecycle management and know your customer (KYC) 

Traditional onboarding is often a slow dance of coordination between relationship managers, compliance teams, and operations staff across multiple jurisdictions. Agentic systems transform this into continuous, parallel orchestration. 

When a wealth management client adds an investment entity, agents gather corporate documents, verify director identities, analyze beneficial ownership structures, and screen sanctions lists in real time—all while maintaining comprehensive audit trails. Relationship managers are notified when the process is complete, allowing them to focus on investment discussions rather than chasing paperwork. 

For ongoing monitoring, agents continuously monitor customer data for triggering events and initiate refresh cycles autonomously, escalating to human review only when professional judgment is genuinely required.

Credit operations and risk assessment

Agentic systems orchestrate the full credit review end-to-end. They extract data from financial statements across formats and languages, normalize accounting treatments across jurisdictions, calculate ratios, value collateral using real-time market data, and draft preliminary risk assessments. 

For established parameters, the entire review runs autonomously. For new or deteriorating exposures, agents prepare comprehensive analysis packages so that credit officers can focus on judgment and client strategy, not information assembly.

Cross-border operations and custody 

Luxembourg's financial sector, concentrated in cross-border funds and custody services, faces unique operational complexity. Agents can autonomously reconcile custody breaks, process corporate actions from announcement through cash and securities movement, and calculate net asset values (NAVs) for complex instruments. 

Governance: The condition for trusted autonomy

Autonomy without trust introduces risk. The EU AI Act, in phased implementation since February 2025, classifies many banking applications as high-risk—a status that demands human oversight, explainability, and robust risk management. The CSSF is equally direct: supervised entities must ensure rigorous governance for any AI system connected to financial services. 

The Digital Operational Resilience Act (DORA), effective since January 2025, adds information and communication technology (ICT) governance and third-party risk requirements that directly intersect with AI use. By August 2026, high-risk AI systems in financial services must be fully compliant.

Institutions must embed trust by design across three dimensions: 

1. Clear decision boundaries: Governance frameworks must define which transactions settle autonomously, which credit decisions require human review, and which compliance escalations demand immediate attention. 

2. Comprehensive auditability: Every automated action must be traceable, with full logging of which agent made which decision based on which data. For credit decisions, the rationale chain must be explainable; for AML alerts, the detection logic must be auditable. 

3. Ongoing regulatory alignment: Capital calculation agents must automatically apply Basel framework updates, and transaction reporting agents must incorporate MiFID II amendments. Maintaining an AI system inventory becomes standard operational practice, ensuring each system’s purpose, risk classification, and decision logic are fully documented. 

Autonomy does not replace humans; it evolves their work. This transition elevates professional roles by prioritizing high-value tasks like judgment, exception handling, and strategic oversight. 

Relationship managers freed from transaction coordination can focus on financial planning and client development. Compliance analysts can investigate sophisticated schemes rather than processing routine alerts. 

This adaptation requires deliberate role redesign: clarifying upfront which decisions remain human prerogatives and which become agent responsibilities.

A new era of advice and execution

Luxembourg is uniquely positioned to lead this transformation. As a jurisdiction built on regulatory excellence, cross-border sophistication, and client-centric service, it can leverage agentic AI to enhance these strengths without compromise. Through the CSSF Innovation Hub and the country’s early adoption of DORA and the AI Act, Luxembourg institutions are uniquely positioned to architect governance models that serve as a global blueprint.

Agentic AI and ZeroOps represent more than technological innovation: they redefine the very nature of banking operations and strategic advice. The future of financial services is now inseparable from how institutions design, govern, and scale this autonomy.