Artificial intelligence in banking is entering a new stage. Traditional AI is normally designed for a defined task: calculate a credit score, recognize a suspicious payment pattern, or answer a customer question. Agentic AI goes further. It can interpret a goal, plan a sequence of steps, use approved tools and data, take an action, assess the result, and escalate when human judgment is needed.
In a fraud-management environment, this could mean much more than generating a risk score. An AI agent could collect signals from the transaction, customer profile, device, location and payment network; compare the activity with normal behaviour; request additional authentication; temporarily hold a payment; open an investigation case; collect supporting evidence; and notify the correct analyst. The value is not simply faster analysis. It is the ability to coordinate a controlled response across several systems.
Why banking is a natural use case
Banks process enormous volumes of events while customers expect payments to be completed instantly. Fraud teams must therefore make high-quality decisions within very small time windows. Static rules remain useful, but criminals continuously change devices, identities, channels and transaction patterns. Agentic systems can combine rules, machine-learning scores, network relationships and contextual information, then select the next permitted action based on the risk.
Consider an unusual mobile-banking transfer. A conventional model may label it high risk and send an alert. An agentic workflow can go further: check whether the device is new, compare the beneficiary with known mule-account networks, review recent password changes, trigger step-up authentication and prepare an evidence summary for an investigator. Low-risk cases may continue automatically, while uncertain or high-impact cases are routed to people.
Real-world systems already show the direction
Mastercard’s Decision Intelligence Pro demonstrates the speed now possible. According to Mastercard, the technology examines relationships between entities surrounding a transaction and improves the risk score in less than 50 milliseconds. Mastercard’s initial modelling indicated an average 20% improvement in fraud-detection rates, with higher gains in some circumstances. The company has also reported using generative AI to accelerate the identification of potentially compromised cards. These are vendor-reported results and the capabilities are not necessarily fully autonomous agents; however, they demonstrate the real-time intelligence layer on which agentic workflows can be built.
Visa provides another practical example. According to Visa, its AI-powered fraud systems screen transactions within milliseconds, while Visa Advanced Authorization evaluates more than 500 risk attributes per transaction. Visa’s Account Attack Intelligence Score also applies generative-AI components to identify and score enumeration attacks in real time. These are Visa’s published product claims; they illustrate the movement from periodic review toward continuous detection and immediate, risk-based intervention.
Feedzai’s RiskOps approach connects behavioural, device, transaction and third-party data to produce persistent risk scoring across the customer journey. Its public discussion of agentic AI also highlights an important reality: fully agentic fraud prevention is still emerging. Mature machine learning is already widely used, while autonomous multi-step decision-making requires stronger evidence, explainability and operational controls.
What an agentic fraud workflow could look like
- Observe: Continuously collect transaction, customer, device, behavioural and network signals.
- Reason: Evaluate risk, identify relationships and determine what additional evidence is required.
- Act: Approve, challenge, delay or block an activity within clearly defined authority limits.
- Coordinate: Create a case, gather evidence, notify the customer or analyst, and update connected systems.
- Learn: Use confirmed outcomes and investigator feedback to improve future decisions under model-governance controls.
The benefits for banks and customers
The clearest benefit is speed. Coordinated decisions can happen in milliseconds instead of waiting for separate teams and systems. This can reduce fraud losses, shorten investigation time and improve compliance with payment-service-level requirements. Better context can also reduce false positives, meaning fewer genuine customers experience unnecessary declines or account blocks. At the operational level, agents can handle repetitive evidence collection and case preparation, allowing fraud specialists to focus on complex judgments and emerging threats.
Autonomy must come with accountability
Banking cannot treat an AI agent as an unrestricted digital employee. A wrong decision can block a customer’s money, approve fraud, create regulatory exposure or damage trust. Every action therefore needs defined authority, traceable evidence and an accountable owner. High-impact actions should require human approval, particularly when confidence is low, the value is high, or a vulnerable customer may be involved.
Strong implementation should include explainable reasons, complete audit logs, segregation of duties, maker-checker approval, model monitoring, data-privacy safeguards and tested fallback procedures. Banks must also defend the agents themselves. JPMorganChase has warned that agentic systems blur the boundary between data and instructions, creating risks when external content or persistent context influences behaviour. Access should be limited to the minimum tools and data needed, and every action should be reversible wherever practical.
A practical adoption path
Banks should begin with supervised use cases rather than full autonomy. A sensible first stage is an investigation assistant that gathers information, explains risk signals and recommends an action. The next stage can automate low-risk, reversible tasks such as case creation, evidence collection and customer verification requests. Only after performance, fairness, resilience and auditability have been proven should the bank allow narrowly defined automated interventions.
Success should be measured through fraud losses prevented, false-positive rates, customer friction, investigation time, analyst productivity and the percentage of decisions that require human correction. A controlled pilot using historical and live-shadow data can reveal whether an agent improves outcomes rather than merely producing impressive demonstrations.
The future: collaboration, not replacement
Agentic AI will not eliminate the need for fraud analysts, risk officers or customer-service teams. Its most valuable role is to connect signals, carry out repetitive steps and respond at machine speed while people retain authority over sensitive decisions. The banks that benefit most will be those that combine real-time intelligence with careful governance. In fraud detection, speed matters—but accountable speed matters more.
Sources
· Mastercard, “Mastercard supercharges consumer protection with generative AI,” 1 Feb 2024. https://www.mastercard.com/global/en/news-and-trends/press/2024/february/mastercard-supercharges-consumer-protection-with-gen-ai.html
· Mastercard, “Accelerates card fraud detection with generative AI,” 22 May 2024. https://www.mastercard.com/us/en/news-and-trends/press/2024/may/mastercard-accelerates-card-fraud-detection-with-generative-ai-technology.html
· Visa, “AI solutions for fraud prevention and detection”. https://corporate.visa.com/en/solutions/visa-protect/insights/ai-fraud-detection.html
· Visa, “Visa Announces Generative AI-Powered Fraud Solution,” 7 May 2024. https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.20661.html
· Feedzai, “What is Agentic AI: From Black Box to Strategic Partner,” 5 May 2025. https://www.feedzai.com/blog/agentic-ai-financial-services/
· JPMorganChase, “Securing the next generation of AI agents,” 23 Mar 2026. https://www.jpmorganchase.com/about/technology/blog/securing-agentic-ai
Janaka Ratnayake | Solutions Engineer – Special Projects (eGov)