Visa’s AI edge: How RAG-as-a-service and deep learning are strengthening security and speeding up data retrieval

Global payment giant visa Operating in over 200 countries and territories, all of which have their own unique, complex rules and regulations.

When policy-related questions are raised, its customer service team must understand these nuances—such as “Can we handle such payments in this country?”—but it is simply impossible to know all of these answers to top.

This means they usually have to manually track relevant information, which is an exhaustive process that can take several days, depending on how they are accessed.

When generative AI appears, Visa treats it as a perfect use case, Search for a generation (RAG) Not only do you need to increase the information to a faster speed, you can also reference it back to the source.

“First of all, this is a better quality result,” Sam Hamilton, Visa’s data and AI signature, told VentureBeat. “It’s also a delay, right? They can handle more cases than before.”

This is just one way Visa uses AI generation To enhance its operations (supported by a deliberately built layered technology stack), while managing risks and maintaining fraud.

Security Chatgpt: Visa’s protected model

On November 30, 2022, Chatgpt was introduced to the world, and history will become a critical moment for AI.

Shortly after that, Hamilton pointed out: “Visa employees are asking, ‘Where is my Chatgpt?’ “Can I use chatgpt? ” “I can’t access chatgpt. ‘I want chatgpt.'”

But as one of the world’s largest digital payment providers, Visa naturally has concerns about its customers’ sensitive data, especially, it remains secure, not in the public domain, and will not be used in the future. Model training.

To meet the needs of employees, in balancing these issues, Visa introduced what it calls “Security Chatgpt,” which is located behind a firewall and runs inside Microsoft Azure. Companies can control inputs and outputs through Data Loss Prevention (DLP) filtering to ensure that no sensitive data leaves Visa’s system.

“All hundreds of data, everything is encrypted, everything is safe at rest and transport,” Hamilton explained.

Despite its name, Secure Chatgpt is a multi-model interface that offers six different options: GPT (and its various iterations), Mistral, Anthropic’s Claude, Meta’s Llama, Google’s Gemini, and IBM’s Granite. Hamilton describes it as a model service or rag-as-a-service.

“Think of it as a level of abstraction we can provide,” he said.

Instead of building their own vector database, they can choose the API that best suits their specific use cases. For example, if they only need some fine tuning, they usually opt for smaller open source models like Mistral. In contrast, if they are looking for more complex inference models, they can choose something like openai O1 or O3.

This way, people are not restricted, nor as if they missed something ready to be available in the public domain (which could lead to “shadow AI”, or use unapproved models). Hamilton explained that a secure GPT is “just a shell on the model.” “Now they can choose the model they want.”

In addition to the secure Chatgpt, all visa developers can access GitHub Copilot to assist with their daily coding and testing. Hamilton noted that developers use vice presidents and plugins for various integrated development environments (IDEs) to understand the code, enhance the code, and execute unit tests (to make sure the code runs as expected).

“So the code is overwritten [identifying areas where proper testing is lacking] Since we have this assistant, it’s greatly increased. ” he said.

Rag-As-A-Service in action

One of the most effective use cases for secure CHATGPT is to deal with policy-related issues specific to a given region.

“You can imagine that in 200 countries with different regulations, there could be thousands, thousands of documents,” Hamilton noted. “It’s really complicated. You need to be nailed, right? And it requires a thorough search.”

Not to mention, over time, local policies have changed, so Visa’s experts must be up to date.

Now, with a powerful rag of reliable, latest data, Visa’s AI not only provides quick answers, but also provides citations and original materials. “It tells you what you can or can’t do and says, ‘This is the file you want, and I’m based on the answer,” Hamilton explained. “We already know what’s in the rag.”

Often, an exhaustive process will take “if not hours, days” to draw a specific conclusion. “Now I can get it in five minutes, two minutes,” Hamilton said.

Visa’s four-layer “Birthday Cake” data infrastructure

These features are the result of visas making substantial investments in data infrastructure over the past decade: Hamilton said the financial giant has spent about $3 billion on its technology stack.

He described the stack as a “birthday cake with 4 layers”: the foundation is a “data platform-AS-A-Service layer with “data”, AI-Service, AI and machine learning (ML) ecosystems, and the top of the data services and product layers.

Hamilton explained that the data platform-AS-AA-Service is basically an operating system built on a data lake that aggregates “hundreds of tons of data.” The above layer, data serves as a kind of “data highway”, with multiple lanes traveling at different speeds to power hundreds of applications.

The third layer is the AI/ML ecosystem, which is the way Visa continuously tests the model to ensure it should be taken and is not susceptible to bias and drift. Finally, the fourth floor is where Visa makes products for employees and customers.

Blockade of $40 billion in fraud

As a trusted payment provider, one of Visa’s top priorities is to prevent fraud, and AI is playing an increasingly important role here. Hamilton explained that the company has invested more than $10 billion to help reduce fraud and improve cybersecurity. Ultimately, this helps the company Attempted to defraud $40 billion Only in 2024.

For example, a new Visa Deep authorization tool provides transaction risk scores to help manage card useless (CNP) payments (such as when users pay via the web or mobile app, as we all do with daily exercises). This is powered by a deep learning recurrent neural network (RNN) model based on context data. Similarly, real-time, accounting account payment protection (thinking through digital wallets or instant payment systems) can achieve instant risk scores and automatically block bad transactions by deeply learning AI models.

Hamilton explained that Visa used a transformer-based model—a neural network that learns context and meaning by tracking relationships in data—to enhance these tools and quickly identify and block fraud. “We want to do this based on the transaction,” he said. “That means we have less than a second of response time, and I should say milliseconds.”

Synthetic data also provides value for fraud prevention: Hamilton’s team augments existing data by synthesizing data to simulate around updated fraud enumerations. “This helps us learn what is happening now, and what can happen in the short and long term, so we can simulate and train the model to capture the data,” he said.

He noted that fraud is an arms race – the entry barriers for threat actors are very low. “We need to be ahead of this and predict and stop them,” Hamilton stressed.