The rise of smaller LLMs and agentic AI in payments transformation

Financial services providers are in the midst of a major modernization push, with AI playing a growing and transformative role. The potential impact is particularly evident in payments, where AI is driving progress in critical areas such as risk management and fraud detection.

Financial institutions face many challenges. They need to implement new regulatory changes, reduce risk, and combat financial crime. The adoption of AI can be a critical tool in this process, and many financial services companies are currently in an evaluation or proof-of-concept phase. They are exploring the potential uses of AI, though many have not yet made a final decision to deploy it.

This is not surprising, as much of the latest in AI development, especially generative AI, is still emerging and requires a well thought-through approach for implementation. In implementing AI, as with adopting any new technology, it is important for an organization first to understand what goals it wants to achieve and then develop a strategic approach to deployment. No major investment should be made without detailed planning to ensure alignment with business objectives and regulatory requirements.

Wide range of applications

Several use cases are currently being actively explored in the payments space that leverage large learning models (LLMs) and Agentic AI. 

One area of focus is the use of AI for intelligent payment routing, providing the ability to manage and distribute any type of payment into the right system seamlessly. The system can analyse the payment and dynamically optimize its flow to maximize transaction success rates and minimize costs. It can analyze various factors like customer data, geographic location, and risk profile to determine the best path for each transaction. 

Another critical area is risk, with many companies exploring AI-driven enhancements to their risk assessment and underwriting systems. For example, AI can be used to better and faster identify anomalies in fraud detection or money laundering. AI algorithms provide the potential to analyze large amounts of data and identify suspicious transactions in real time. At the same time, more accurate pattern recognition can also minimize the false positive rate. 

AI could also support the automation of manual and error-prone processes to reduce costs, increase speed, and improve the quality of workflows. One example is the correction of payment data, whereby providing a pre-validation and assessment support step minimizes the need for manual intervention. If manual intervention is required, the customer service agent can collaborate with generative AI to simplify the process of remediation. This holds great potential to accelerate the remediation process. 

The opportunities of AI are immense. However, it must also be understood that payments is a highly regulated market and that the security, transparency and reliability of the systems used must be given the highest priority. This is especially true with regard to the regulations and guidelines resulting from the EU's AI Act. In this context, three approaches in particular are gaining in importance: open source, smaller AI models, and hybrid cloud environments.

Transparency through open source

Open source policies, technologies, and solutions are about collaboration and transparency. In the context of AI, for example, this means easily accessible and reliable data and transparent decision-making, preferably with the ability to contribute back to a model. This forms the basis for "trustworthy AI", whose guiding principles are explainability, fairness, robustness and controllability of AI models. This type of AI is essential for financial institutions, as the EU AI Act, BaFin (German Federal Financial Supervisory Authority), MaRisk (Minimum Requirements for Risk Management) and BAIT (Supervisory Requirements for IT in Financial Institutions) already set important guidelines in this area.

Advantage of smaller models and agentic AI

Financial services providers also face a challenge when it comes to selecting the right AI model. It is becoming apparent that institutions will not be solely reliant on LLMs, but are looking to smaller models (SLMs) that are trained for specific use cases. For example, SLMs with billions of parameters versus LLMs with trillions of parameters. 

Smaller models offer multiple advantages. They can be implemented using fewer resources, helping keep down costs and energy usage, and support the continuous integration of new data, especially institution- and domain-specific data. As a result, training runs can be performed much faster. Finally, by using open source-licensed smaller models, firms can gain greater supplier-independence and have more flexibility to choose solutions that offer visibility to algorithms, training data and model weights.

At the same time, agentic AI is developing as a new paradigm in the implementation of AI models. These AI agents could analyze trends, make split-second decisions, and dynamically adjust strategies based on real-time data and news events. This could lead to more efficient systems.

Hybrid cloud as the infrastructure foundation

Payments is an inherently security-critical area as it involves money transactions and the storage and management of highly sensitive financial and personal data. For this reason, corporate-owned infrastructure and on-premises servers have often been favored to host the data and payment software. However, this approach has its limitations, including a lack of flexibility to introduce innovative technologies such as AI. In response, the cloud market has evolved to offer the best of both worlds: the hybrid cloud. The hybrid cloud is typically enabled through an abstraction layer and central management platform, meaning organizations can continue to use on-premises storage for sensitive data while taking advantage of the scalability benefits of the public cloud. They can move between public cloud providers based on their own policies, market demands, or regulation, while continuing to meet operational resiliency requirements.

A hybrid cloud approach hinges on having portability of workloads with consistent and secured management throughout the application lifecycle, from AI model development and training to AI model integration into a payment application. Financial institutions have the option to develop and train an AI model in a public cloud using publicly available data or synthetic test data and then embed it in the on-premises application. Or, they can train the models with confidential data in their own data center and then run them in a public cloud.

Acknowledging that AI comes with privacy concerns, investment costs and the challenge of setting a solid strategic direction from the outset, the potential role it can play in the payments industry is vast. It can streamline processes, increase efficiency and drive innovation. Underpinning it all is the need for a flexible, hybrid infrastructure that is built for greater transparency and collaboration, and that can handle whatever twists and turns the fast-moving AI world has in store.