ChatGPT can do so much, can it route payments?

Artificial intelligence (AI) and machine learning (ML) are the talk of the town today, and the most frequently mentioned example of using both is ChatGPT from OpenAI. Effectively an intelligent chatbot that mimics human interaction, this technology offers an intriguing opportunity to automate the decisioning process that can route transactions to specific payment service providers (PSPs.)

Behind the scenes, ChatGPT uses three core technologies:

• Natural Language Processing (NLP): The system accepts the words that are typed in by the user, then processes them to turn them into something a computer can understand. So, for instance, a question like “which are the top token vault vendors right now” may be turned into a SQL statement the computer could user to search a database, or into a well-structured search request that could be plugged into Google or Bing.
• Machine Learning (ML): Under the covers, machine learning creates a series of algorithms that find trends and patterns in data. This allows the system to answer the original question, or prompt, albeit with data that may not be ready for human consumption.
• Generative AI (GenAI): A Large Language Model (LLM) is developed using specialized ML to understand, and be able to mimic, large volumes of language data. The GenAI module uses that LLM to frame the answer the underlying ML algorithms were able to find, converting it into human-like language.

A key thing to remember about the middle step of this process—finding the correct information to respond to the original prompt— is that AI, unlike traditional programming, works on a probability model. In practical terms, think of it as answering questions with answers that are likely, where traditional, or deterministic programming will only offer an answer that is definitively correct. 

This not only leads to GenAI sometimes producing incorrect or misleading answers (often termed hallucinations), but also provides a useful guidepost for how to best utilize AI: it is very well-suited for applications where trends and patterns are more important than down-to-the-last digital point accuracy.

AI as a Decisioning Engine 

ChatGPT itself is already an application built on top of AI models: it’s a chatbot, and as such not necessarily ideal for use as a backend decisioning engine. The models themselves, though, very much are.

For multi-processor merchants, AI presents a truly intriguing opportunity to make in-the-moment decisions on where to direct transactions as they are processed. A merchant might, for instance, set up their own orchestration engine by:

• Uploading all their PSP contracts, including fee schedules and volume commitments, into a custom AI model. They could add daily volume numbers over time, or, if available, create connections to PSP records so that the model can access real-time numbers through Retrieval-Augmented Generation (RAG).
• Building a standard prompt, that takes key details from the transaction (customer location, transaction size, etc.) and asks the AI model which PSP the transaction should be directed to.
• Setting their payment system to always request this information before sending the transaction on to whichever PSP is best suited.

The benefit of this approach is that it is an ideal use for AI: it can see trends, develop algorithms that ensure the highest number of transactions are successfully processed at the lowest fees, and constantly be learning and adjusting to patterns as they emerge. This stands in stark contrast to a deterministic, or traditional, program, which would have static, human-developed decision logic, which would need to be manually updated in response to analyses completed separately. 

Importantly, this is also using AI in a manner that is destined for success: this notional decisioning engine is looking for trends and patterns and making decisions that are the most likely to be correct. 

In other words, there may not be a perfect answer, so a system that is more probabilistic than married to 100% accuracy is ideal.

Return to Top

AI + Token Vault 

The most common approach to gathering payments by merchants is a single step: once the consumer selects what they want to buy, their payment is processed through a full-service PSP like Stripe, or through a platform’s provided service like Shopify. 

This approach, however, eliminates the opportunity to:

• Run third-party, or self-developed, security checks to avoid sending ‘bad’ transactions.
• Collect and store consumer payment information that they retain access to, rather than having it secured by the PSP.
• Select different providers based on payment method or geographical location, increasing the likelihood of successful transaction completion.
• Opt for the PSP offering the lowest processing fee for each particular transaction.
• Arbitrage currency conversion rates for their own benefit, rather than allowing the margin to be captured by the PSP.

With a programmable token vault, by contrast, merchants separate the steps of the payment process so that they can

• Have payment information securely collected and stored - without bringing the core payment system into PCI-DSS scope - for future use, reducing friction for repeat costumes.
• Run additional security checks prior to submitting transactions for processing, reducing chargeback risks and improving security for customers.
• Select the right PSP for the task, increasing the likelihood of a successful close, reducing the processing fees payable, and offering the opportunity to manage and benefit from the currency conversion process.

Using an AI agent to select the appropriate PSP for each transaction automates the process of making the right choice, without having to write complex deterministic codes or maintain manual algorithms.

Adding an AI decisioning engine to a multi-processor payment system, using a programmable payments vault, isn’t just good sense: it’s the future.

Return to Top