.jpg)
At Unipaas, we recently tackled a persistent challenge in the payments industry: converting declined payments into successful transactions. Traditional payment systems typically respond to failures with generic messages like "Payment failed. Please try again." This is frustrating for customers, takes away from an enjoyable payment experience, and, in many cases, can lead to lost transactions.
We saw an opportunity to dramatically improve this experience by providing customers with personalised, actionable guidance. This article explains how we used AI and vector databases to make this possible.
The slowed response times of too much information
When a payment is declined, every second counts. Customers need immediate, accurate, and personalised guidance to complete their purchases. To provide these responses, we began using OpenAI’s GPT.
Read more on how we use AI to drive payment conversion
However, there was a catch: GPT doesn’t store any memory between requests, it’s “stateless.” This required resending the entire knowledge base for each and every request, including:
An action that was both costly and resulted in slow response times for customers, presenting us with another challenge. A problem we solved using vector databases.
The solution: retrieving the right information with vector databases
So, here’s roughly how it works. When a payment fails, our system doesn't just display a generic error message. Instead, it aims to help the user resolve the specific issue. To achieve this, it converts all the payment guides, support resolutions, and documentation into a numerical representation called an embedding. These embeddings are then stored in a vector database (Chroma DB).
Dive into the technical details of how we work with AI and vector databases
When a payment is declined, the system converts the error message into an embedding as well, and the vector database is used to retrieve the most relevant information based on the semantic meaning. This relevant information is then sent to OpenAI's GPT model which crafts a clear, detailed and actionable message to help the user resolve the issue.
The impact: increased conversion
The above-described Intelligent Messaging System enables us to provide personalised messages while minimising response times by avoiding sending entire documents to the GPT model with each query.
It provides quick and accurate explanations for payment decline reasons, offering actionable advice to help users successfully complete their transactions. For example, “It appears your billing address does not match the information on file with your bank, which may be causing the ‘Payment Not Completed’ error. Please update the billing address in your wallet,” is significantly more helpful than, "Payment failed. Please try again."
Hence, it’s no surprise that this system has resulted in increased checkout conversions.
A word on Apple Pay
Nowhere is the importance of intelligent messaging clearer than with Apple Pay. In the last few years, we’ve seen a significant increase in transactions using this payment method.
Converting declined transactions by providing accurate messaging over the generic variety proves particularly difficult for traditional solutions. These rely on many parameters, such as the user's personal details, device info, purchase history, and the like to verify transactions. If any go unmet, the result is a generic payment failure message. Yet intelligent messaging provides the necessary information for the customer to complete the transaction and therefore can be a big conversion booster for platforms offering this essential payment method.
Bridging technology and customer experience through intelligent messaging
Our implementation of vector databases and AI demonstrates how combining these technologies can solve a specific, common payment challenge: converting declined transactions into successful ones. By transforming generic decline messages into personalised, actionable guidance, we've come up with a system that prioritises customer experience while maintaining fast response times.
The combination of OpenAI's GPT models with vector databases proved particularly effective, allowing us to deliver immediate, relevant responses to payment failures while eliminating the need to send entire knowledge bases every time. This creates a maintainable architecture that can incorporate new payment scenarios and documentation as needed. And at Unipaas, we'll continue to refine this system based on new decline patterns and customer feedback, ensuring our users receive the most helpful guidance possible when payment issues arise.
.png)
