Skip to main content

Blog

Agentic AI and AI Governance in Banking: Risk Management and Supervision

What Agentic AI Means for Banking

Agentic AI refers to artificial intelligence systems designed to act autonomously toward specific goals. In banking, this means AI that can make decisions and take multistep actions for customer onboarding or mortgage approvals with minimal human oversight. These systems operate across complex workflows, learning from data and adjusting their strategies as conditions evolve, with key capabilities including autonomy, adaptability, and coordination across a bank's ecosystem.

Exciting insights from Day 2 at PulseCore Events on Digital Banking Excellence, centered on the theme of Digital Identity and Digital Onboarding!
Exciting insights from Day 2 at PulseCore Events on Digital Banking Excellence, centered on the theme of Digital Identity and Digital Onboarding!

Unlike traditional AI tools that assist with tasks like document retrieval or fraud alerts, agentic systems plan, reason, and adapt without constant human direction. They pursue defined goals through sequences of actions, making autonomous decisions within guardrails set by the institution. This marks a fundamental shift: rather than human operators telling the system "assist me," the agent operates on "do it for me" instructions and executes them independently.

The Risk Management Summit 2027 (25–27 May in Amsterdam) features sessions on agentic AI and supervision, bringing together senior risk and compliance leaders to examine how these systems reshape oversight and governance.

How Agentic AI Differs from Traditional AI

Generative AI and traditional automation both have important limitations that agentic AI overcomes. Agentic AI systems consist of autonomous agents possessing agency—the ability to take initiative and execute actions in pursuit of defined outcomes, built on large language models (LLMs) and augmented with techniques such as retrieval augmented generation (RAG).

Traditional automation runs a fixed sequence for defined inputs. Generative AI generates content and may retrieve data but typically does not coordinate sequences of steps toward an outcome. Agentic AI adds the ability to interpret context, choose among paths as conditions change, and execute entire workflows from start to finish. For example, a traditional AI chatbot might retrieve a customer's balance and suggest a savings plan. An agentic system could autonomously create that savings plan, monitor spending patterns, adjust the strategy in real time, and handle exceptions—all without human prompts.

Service representatives at financial services institutions report using a median of 10 technologies to support a single customer interaction. Agentic AI can unify and simplify those workflows, reducing handoffs and accelerating execution.

Where Banks Deploy Agentic AI

Agentic AI applications span front, middle, and back-office functions. In the front office, generative AI and agentic systems allow banks to provide tailored experiences at scale while guiding customers through complex tasks, with customers interacting with personalized financial assistants that understand conversational nuances rather than navigating menus and forms.

Middle-office use cases include credit scoring and loan processing automation, enhanced risk assessment and compliance monitoring, and fraud detection. Agentic AI evaluates creditworthiness using multiple data points, automates approvals and funding decisions, monitors real-time regulation updates and transactional anomalies, and flags suspicious behaviors before they impact the customer or the bank. In anti-money laundering, a multi-agent system could work on AML investigations where one agent reviews an alert to understand what rule has been violated, another reviews transactions, a third documents findings and recommends steps, and finally a human validates the report before an agent autonomously files suspicious activity reports with regulatory bodies.

Back-office applications automate document processing, exception handling, and large-scale data analysis. Agentic AI identifies patterns, generates insights for faster decisions, and turns forecasting trends into actionable intelligence.

The Governance Challenge

Oliver Wyman's 2026 AI Risk and Governance Market Study found that fewer than half of surveyed financial institutions have a board-approved strategy for managing AI risk, highlighting how governance is becoming a critical differentiator. This gap is critical because agentic AI introduces fundamentally new risks compared to static models.

Traditional model risk management was designed for relatively stable models with predictable inputs and outputs. Agentic AI changes that assumption. Rather than producing a single response, AI systems can retrieve information, interpret documents, invoke tools, draft communications, calculate results, route work, and execute actions—sometimes without human oversight. Risk therefore exists not only in the model itself but across the entire execution path: how the agent retrieves data, which tools it invokes, what authority it exercises, and how it records decisions.

A customer response may appear accurate while relying on outdated information. A well-written credit memo may use unauthorized data or bypass required approvals. The quality of the final output alone no longer tells the whole story. Senior risk and compliance leaders examine these challenges in depth at the summit's schedule of 33 sessions, covering agentic AI, supervision, cyber, DORA, fraud, AMLA, credit, liquidity, and stress testing.

Risk Management Across the Agentic AI Lifecycle

Effective agentic AI governance requires a structured approach across six key stages: defining the AI profile and boundaries, standardizing risk identification, determining risk materiality and governance required, governing by system and envelope, industrializing risk oversight, and defining monitoring and observability.

Deloitte's analysis reveals more than 350 risks that can arise from autonomous or agentic behaviour, with many possibly posing a threat to banking systems and processes. These risks cluster into three layers: core reasoning risks (where overreliance on weak training data or memory defects distort how agents interpret tasks), tool and connectivity risks (where misconfigured permissions lead to unintended document changes or unauthorized data access), and ecosystem risks (where agents interact with external platforms, vendors, or other agents, introducing counterparty and propagation risks).

Banks need governance that documents and evaluates how AI reaches a decision, not simply the decision itself. Effective controls should determine whether an AI-generated action is appropriate, permitted, evidenced, and authorized before it progresses. Human oversight remains essential but its role shifts from reviewing every output to setting policies, defining thresholds, approving exceptions, and monitoring outcomes at scale.

Building Supervisory Frameworks

To scale agentic AI responsibly, banks should treat AI agents as active operators within their systems and design controls accordingly. Key foundations include treating agents as accountable actors similar to human employees, maintaining unique agent IDs and output tagging, implementing immutable tool-use logs, and ensuring real-time monitoring leaves clear audit trails.

Leading speakers at the summit include experts in AI risk, compliance and supervision who address how banks can expand their AI risk management frameworks to assess agents—not just models. Risk taxonomies should add new categories like tool misuse, action validity, and outcome monitoring. Banks can then maintain agent registries with metadata, risk scores, and tier-aligned controls that determine deployment approvals, permissible actions, and the requisite scope of monitoring.

Many banks rely on third-party AI and adapt it to their environments. Even when technology comes from a vendor, accountability does not. Vendor controls alone are not enough to manage reputational, regulatory, operational, and customer risk. The business should own use and outcomes, risk and compliance should provide oversight, and internal audit should deliver independent assurance.

Real-World Adoption and Readiness

Agentic AI adoption is in early stages despite high expectations. BCG's AI Safety and Governance in Banking Benchmark found that 73% of banks permit agents in production for select use cases, but at 88% of banks agents make up 10% or less of the AI portfolio, and only 8% of institutions have a comprehensive, standardized monitoring and control framework that explicitly covers agents.

Wrapping up Day 4 at PulseCore Events on Digital Banking Excellence, we delved into the fascinating world of Artificial Intelligence in the financial sector.
Wrapping up Day 4 at PulseCore Events on Digital Banking Excellence, we delved into the fascinating world of Artificial Intelligence in the financial sector.

Organizational governance maturity is becoming a differentiator. Most banks have established foundational AI risk frameworks, but significant gaps remain in translating these into consistent, scalable operating capability. Seventy-seven percent of institutions have a dedicated AI policy, yet fewer than half have a centralized, tool-based AI inventory, and none report a comprehensive human-oversight framework consistently implemented across AI systems.

Effective governance enables speed: banks in the top third for more mature risk management are 2.5 times as likely to be advanced AI adopters as those in the bottom third. Banks using preapproved AI patterns are twice as likely to move lower-risk applications into production in under six months. The path forward requires not isolated pilots but an operating model that includes a centre of excellence uniting technology, operations, risk, compliance, and product specialists.

Frequently asked questions

What is the difference between agentic AI and a standard chatbot?

A standard chatbot responds to individual queries and may retrieve information or suggest actions, but it does not execute multistep workflows independently. Agentic AI plans a sequence of actions toward a goal, executes those actions using available tools, monitors results, and adjusts its approach as conditions change. It operates with autonomy within guardrails rather than waiting for human input after every step.

What are the main supervision challenges agentic AI introduces?

Agentic AI is designed to operate with minimal or no human supervision, which creates new risks. Supervisors must now monitor not just model outputs but how agents retrieve data, invoke tools, exercise authority, and record decisions. Traditional human-in-the-loop designs that place a person at every decision point may defeat the efficiency purpose. Supervision shifts to setting policies, defining thresholds, approving exceptions, and monitoring outcomes at scale using real-time dashboards and automated controls.

Are most banks ready to deploy agentic AI?

No. While 73% of banks permit agents in production for select use cases, it comprises 10% or less of their AI portfolio at most institutions. Only 8% have a comprehensive monitoring and control framework for agents. Most banks are building governance foundations but lack consistent, scalable implementation. Governance maturity is now a competitive differentiator; banks that establish clear controls and reusable approval patterns can move lower-risk applications to production much faster than peers.

How should banks approach third-party agentic AI platforms?

Banks adopting third-party agents (from vendors like Amazon, Salesforce, or Google) must recognize that accountability remains the bank's responsibility. Vendor controls alone are insufficient. Institutions should maintain agent registries with metadata and risk scores, conduct due diligence on model updates and memory mechanics, monitor inter-agent interactions, and define clear escalation rules. The Risk Management Summit 2027 offers in-person and virtual passes with documentation and recorded sessions covering these third-party governance challenges in detail.

Sources

  1. Guide to Agentic AI in Banking | Salesforce
  2. Agentic AI in banking | Deloitte Insights
  3. How Retail Banks Can Put AI Agents to Work | BCG
  4. CEO mandate must be to ensure effective AI risk management | Oliver Wyman
  5. Managing the new wave of risks from AI agents in banking | Deloitte Insights
  6. How Banking Is Adapting to AI Governance and Risk | BCG

Share this article

Blog