Artificial Intelligence (AI), especially Large Language Model (LLM) technology such as ChatGPT, is helping financial institutions (FIs) save money and launch products faster. An AI-first integration approach also unlocks operational efficiencies, scalability, error reduction, flexibility and adaptability, simplified upgrades, security, and more. Here, we’ll share some of the ways we are helping our FI customers harness the power of AI to connect with fintechs faster, stay ahead of technology, and keep pace with customer demands.
Overview and benefits of AI-first architecture
The recent LLM AI breakthrough has already demonstrated its incredible potential for translating human languages. In many ways, computer languages are similar to human languages, with grammar, structure, syntax, and semantics. Therefore, using an LLM AI agent for “translation” (akin to “transformation” in traditional computing) is a natural application. This approach provides three distinct advantages:
- Writing conventional data mapping code requires significant development time. FIs can eliminate this costly step and accelerate development time significantly by implementing AI-first practices.
- With an AI-first approach, system upgrades are cheaper and faster because FIs can minimize developer involvement. As endpoints evolve, agents simply need to be re-trained and re-tested.
- Conventional data mapping is rigid and inflexible. Therefore, any unexpected data variation will lead to exceptions requiring time-consuming manual involvement. With AI, mapping can programmatically evolve to automatically retry errored-out transactions with generative algorithm improvements.
Starting with AI agents dedicated to banking services
AI agents dedicated to banking services are AI algorithms trained to perform message transformation between a given set of APIs for a clearly defined set of use cases. For example, an AI agent can be trained to receive an inbound JSON message and concurrently transform the message into XML for a banking core, a CSV file for a document storage system, and an event for a message queue.
The training involved is twofold: First, the agent will learn about the schema of the inbound message and the API for the downstream systems via common schema definition files such as Swagger and WSDL. Then the AI agent can be trained by parsing through historical interaction data (i.e., the AI agent accesses records that show the desired outcome for a given message). Again, the ratio of outputs generated by an input does not need to be one-to-one; FIs should train AI agents to consolidate inputs and generate multiple results. It is critical to provide the AI agent access to a sufficiently large volume of past integration data to develop the pattern recognition knowledge that will enable it to transform the message and invoke the downstream APIs accurately.
Once trained, the AI agent can replace conventional data mapping and routing code in an integration application or API backend code.
Building an agent-to-agent ecosystem
As more financial services AI agents become available, the integration architecture can evolve to enable primary communications between AI agents (i.e., one leg of the communication will be from one agent to another, and the other leg will be from the agent to the specific system it’s fronting). Such architecture will provide the fastest time to market for the introduction of new integration services because agent-to-agent communication is quicker to train, and generative advancement is faster to achieve. To ease the transition to an agent-to-agent architecture, FIs should adopt an industry-specific canonical data model for messages among AI agents to standardize agent training and reduce training time.
The power of dynamic error correction and retry
A key advantage of utilizing an AI-driven method to integrate data is the ability to quickly and automatically rectify errors arising from data exceptions. Suppose an integration transaction fails due to a data exception. If the error is caused by use cases not covered by mapping logic (a very common occurrence), the application will break and require developer time to make changes to the mapping. This process includes more than just modifying the code; it also requires proper regression testing and redeployment of the altered code. This can take a significant amount of time since any changes to the code must go through regression testing and redeployment.
With the AI-first approach, FIs can adjust and retry data transformation errors immediately as the AI agent parses the error messages and reacts accordingly. Additionally, with every successful retry, the agent’s capability will generatively improve.
Prerequisites for AI-first architecture
For the implementation of AI-first integration to become a reality, an FI must fulfill certain requirements.
Private AI agents
The sensitivity of the data involved and the rapid response required for financial transactions means that the AI agents used for integration should be dedicated private agents. Fortunately, the market for private AI agents exists, and all of the primary AI LLM service providers offer private instances of their AI platform.
High volume and diversity of training data
For the agents to properly transform the messages on the fly, they must be well-trained with numerous records, including a wide variety of data combinations.
Standard data model
For agent-to-agent interaction, it is essential to have a common data model for messages among agents. PortX developed the Open Banking API model for exactly this kind of application. Additionally, PortX is a member of the Financial Data Exchange consortium, which supports the FDX standard for exchanging information in the financial services industry.
Agent certification authority
For the industry to widely adopt AI-first integration, a secure and trustworthy mechanism must authenticate and validate AI agents over the internet – a system that currently does not exist. However, by fronting AI agents with a simple cloud-based API, FIs can employ proven authentication mechanisms such as OAuth2.0 and mutual TLS to ensure the security of transactions.
How PortX Can Help
As the integration solutions provider dedicated to the financial services industry, PortX is uniquely suited to help implement the AI-first integration approach in the following ways.
- AI Agent Training – With years of experience in integration projects, PortX has accumulated a vast collection of sanitized messages crucial for training AI agents to perform the necessary transformations.
- API-Fronting AI Agent – PortX is a leader in API development with an extensive library of pre-built APIs that FIs can quickly deploy to work with AI agents.
- AI Agent Governance – Deployment of AI agents requires proper monitoring and governance. PortX has developed expertise in this area to ensure a robust governance framework will be in place to manage the AI agents.
If you want to learn more about how we are helping FIs build AI-first integration architecture, don’t hesitate to contact us for more information and schedule a consultation. We are always happy to help and answer any questions you may have.