Join the Community

24,309

Expert opinions

40,842

Total members

355

New members (last 30 days)

216

New opinions (last 30 days)

29,356

Total comments

Join Sign in

International Fraud Awareness Week - How Agentic AI is revolutionising fraud management

18 November 2025 Be the first to comment

Oliver Tearle

Head of Technology Innovation

The ai Corporation

Agentic AI’s automated decision-making capabilities have demonstrated impressive performance recently. From automated software development to personal assistants, its use cases are rapidly expanding as the technology is incorporated into many business areas and products.

Agentic AI improves upon previous automation technology by using AI models such as Large Language Models (LLMs) and Reinforcement Learning to take actions based on information it has learned from its environment, rather than following predefined decision paths. Agentic AI can understand the context of the task and decide on its own how best to approach it using available tools and data. For instance, Agentic AI has widespread use as a personal assistant, capable of performing tasks such as booking flights without being specifically told which flight, or airline to book – it just books the best for the provided information (price, date range, etc).

Agentic AI is already presenting the payments industry with a significant opportunity to revolutionise how it manages fraud and optimises its fraud systems. Today, fraud management is an incredibly manual process, even with the use of specialised fraud detection models and tools. Handing this complex process over to an AI Agent is the next step in the evolution of fraud management.

Fraud management generally comprises four distinct processes:

Performance review of any current, live models and rules.

- How effective is the overall system at detecting fraud?
- How well are they identifying fraud?
- What is the false alarm rate? Has it risen recently?
- Have any economic or environmental disruptions caused significant shifts in payment trends?
Reviewing the latest data, including fraud cases and suspicious activity, to determine what changes are happening within each data region, and understand the movement of fraud trends. This process allows the fraud managers to understand the latest trends so that they can change rule parameters accordingly.
Developing fraud rules or models to keep up with moving trends
Optimising the fraud strategy. This process aims to keep the number of rules and models to a minimum whilst detecting the maximum amount of fraud, with the lowest false positives. Poorly performing rules or models, which once worked well, may no longer be effective due to ‘concept drift’ – where payment trends change over time, and need to be removed. The fraud manager can then integrate the new rules and models from stage 3 into the strategy; however, the system needs to be carefully balanced to ensure optimal performance.

Employing Agentic AI for each process ensures streamlined fraud management with minimal human intervention. The system continues to learn over time, and the human guides its decisions. This continual feedback loop enables the system to remain effective, even as payment trends change.

So, if we apply this idea to each of the above processes:

Performance review. Agentic AI will not only determine the performance of rules and models based on the standard metrics, but will also develop some of its own metrics, which humans may never have considered, such as how good a model is at detecting a fraud case quickly. Date review. Instead of following set processes for reviewing new data like most humans, the agent may develop its own data review process, which produces results that are more descriptive and useful than human-developed metrics, driving efficiency of later processes.
Developing fraud rules and models. When provided with data science tools to develop machine learning models, the agent can test which approach works best. For example, that might involve a manual rule or a deep neural network. The agent will then refine the model or rule until it performs as well as it can without human interaction.
Optimising the fraud strategy. The final stage is the most difficult for the AI agent, but it is where all feedback is received. The agent must select the optimal combination of rules and models for the current trends, whilst keeping false positives within a range suitable for the fraud management team to review. The agent in this stage can interact with the agent from stage 1 to design the best-performing strategy. Once the new plan is available, humans will give feedback by implementing changes. If specific rules are implemented, a positive signal is provided to the agent, and if changes aren’t implemented, then a negative signal is provided instead. This feedback loop forces the agent to re-strategise for better results the next time around. The last feedback loop is the result of the fraud alerts. If a fraud reviewer takes any positive action based on an alert, for example, blocking a card or contacting a customer, this positive feedback is sent to the agent. If the human fraud reviewer dismisses the case as genuine, this negative feedback is provided again to the agent, forcing it once more to improve for next time.

Agentic AI is a powerful optimisation tool, particularly driven by feedback for continual improvement. When provided with the correct set of tools and a team of fraud experts to ensure the AI is regularly reviewed, this technology has the potential to raise the bar and dramatically reduce fraud losses and increase acceptance.

External

This content is provided by an external author without editing by Finextra. It expresses the views and opinions of the author.

947

Report

Channels

/artificial intelligence /payments

Transaction Fraud Systems and Analysis

A community for discussion of Transaction Fraud systems and anlaytical techniques for bank card and financial services organisations.

Join group

69 opinions 27 members 18 November 2025

Comments: (0)

Oliver Tearle

Head of Technology Innovation

The ai Corporation

Member since

10 May 2021

Location

London

More expert opinions

Stanley Epstein Associate at Citadel Advantage Group