Visa’s AI edge

Global payment giant Visa is active in more than 200 countries and territories. Each has its own complex rules and regulations.

The client services team of Visa must be able to answer questions about policy, such as ‘are we allowed process this type payment in this country’. But it is impossible to remember all the answers.

They’ve had to manually track down relevant information, which can take up to a week depending on the accessibility of the information.

When generative AI was introduced, Visa saw it as a perfect application. They applied retrieval-augmented creation (RAG) in order to not only retrieve information up to 1,000X quicker, but also cite its sources.

Sam Hamilton, Visa SVP for data and AI told VentureBeat that “first of all, the results are better.” “It is also latency,” said Hamilton. They can handle more cases than before.

This gen AI is one way Visa uses to enhance its operations, supported by a deliberately built, tiered tech stack – while managing risk and preventing fraud.

Secure ChatGPT: Visa’s protected models

The day ChatGPT is introduced to the world on November 30, 2022 will be remembered as a pivotal event in AI.

Hamilton noted that not long after, “employees of Visa began asking, ‘Where is ChatGPT? ‘, ‘Can I Use ChatGPT? ‘, ‘I do not have access to ChatGPT. ‘, ‘I want ChatGPT. ‘”

Visa, as one the world’s biggest digital payment providers, was naturally concerned about the sensitive data of its customers — specifically, whether it would remain secure,

To balance employee demand and these concerns, Visa launched ‘Secure ChatGPT’, which runs on Microsoft Azure behind a firewall. The company can control inputs and outputs via data loss prevention screening (DLP), ensuring that sensitive data does not leave Visa’s system. Hamilton said that “all the hundreds of petabytes are encrypted and secure both at rest and in transit.”

Despite its name, Secure ChatGPT offers six different interfaces: GPT (and all of its iterations), Mistral, Anthropic Claude, Meta’s Llama, Google Gemini, and IBM Granite. Hamilton called this model-as a service or RAG as a service.

He said, “Think of it as a kind abstraction where we can provide a layer.”

Instead, people can choose the API which best suits their use case. If they only need to fine-tune the model, they will typically choose a smaller, open-source model such as Mistral. However, if a more sophisticated reasoning model is required, they may choose OpenAI o1 and o3.

In this way, people won’t feel restricted or like they’re missing out what’s available in the public realm (which could lead to’shadow’ AI or the use unapproved models). Hamilton explained that Secure GPT was “nothing but a shell over the model.” “Now they can choose the model they want to put on top of that.”

Visa developers have access to GitHub Copilot for their daily coding and testing. Hamilton said that developers use Copilot plugins and various integrated development environments to better understand code, improve code and perform unit tests (to ensure code runs as intended).

He said, “The code coverage [identifying areas where proper testing is lacking] has increased significantly since we have this assistant.”

RAG-as a service in action

Secure ChatGPT can be used to answer policy-related questions that are specific to a particular region.

As you can imagine, if you are in 200 countries, each with a different set of regulations, the number of documents could be thousands, tens of thousands or even hundreds. “That gets really complex. You have to nail it, right? Visa’s experts need to be up to date on local policy changes, as well.

With a robust RAG based on reliable, current data, Visa AI not only retrieves answers quickly, but also provides citations. Hamilton explained that the AI tells users what they can and cannot do. It also provides a document to use as a basis for its answer. “We have narrowed the answers with the knowledge we have built into RAG.”

In the past, the exhaustive process could take “if not days, then hours” to reach concrete conclusions. Hamilton said, “Now I can do that in just five minutes or two minutes.”

Visa’s four layer ‘birthday-cake’ data infrastructure

According to Hamilton, these capabilities are the result from Visa’s heavy investments in data infrastructure during the last 10 years. The finance giant spent around $3 billion on the tech stack.

He describes that stack as a “birthday cake with 4 layers”: The foundation is a ‘data-platform-as-a-service layer, with ‘data-as-a-service,’ an AI and machine learning (ML) ecosystem and data services and products layers built on top.

Data-platform-as-a-service essentially serves as an operating system built on a data lake that aggregates “hundreds of petabytes of data,” Hamilton explained. The data-as a service layer is a “data highway” that has multiple lanes moving at different speeds, allowing hundreds of applications to be powered.

The third layer, the AI/ML eco-system, is where Visa tests models continuously to ensure that they perform as they should and are not susceptible of bias and drift. Visa builds products in the fourth layer for its employees and clients.

Blocking $40 billion of fraud

As a trusted payment provider Visa’s top priority is fraud prevention. AI plays a greater role in this area. Hamilton said that the company had invested more than 10 billion dollars to reduce fraud and improve network security. This ultimately helped the company. Block $40 billion in attempted fraudulent transactionsby 2024.

A new Visa deep authorization tool, for example, provides transaction risk scores to help manage card not present (CNP) payment (such as when users make payments via web or mobile apps, as we all do every day). This is powered by an RNN model that uses petabytes worth of contextual data. Deep learning AI models are also used to enable real-time account-to-account payments protection (think digital wallets and instant payment systems). These models produce instant risk scores, and block bad transactions automatically.

Hamilton explained Visa used a transformer model — a network that learns meaning and context by tracking relationships within data — to enhance and thwart these tools. “We wanted to align that with the transactions,” said Hamilton. “That means that we have response times of less than a millisecond, or less than a second.”

The use of synthetic data is also valuable in fraud prevention: Hamilton’s group augments existing data by using synthetic data to run simulations around the latest enumerations. “That helps us understand what’s going on now and what might happen in the near future and the long term so we can train and simulate the model to capture the data,” said Hamilton. He said that fraud is a race to the bottom — and that there’s no barrier to entry. Hamilton stressed that we need to be one step ahead and anticipate and stop them.