Global payment giant Visa It operates in over 200 countries and territories, all with its own unique, complex rules and regulations.
The customer support team must understand these nuances when questions related to politics appear-an example “Can we process this type of payments in this country?” -But you simply cannot know all these answers.
This signifies that they sometimes needed to manually track the relevant information – an exhaustive process which will last days depending on how available it is.
When generative artificial intelligence appeared, Visa saw it as an ideal use of use, using a generation of recovery (RAG), not only to attract information about the temperature as much as 1000 times faster, but back to its sources.
“First of all, it’s a better quality of results,” said Venturebeat Hamilton, SVP Visa and AI. “It’s also a delay, right? Many more cases can deal with zo than before. “
This is just one of the ways in which the visa is used by the AI gene to enhance its activities-supported by intentionally built, multi-level technical stack-at the same time managing the risk and stopping the fraud.
Secure Chatgpt: Visa protected models
On November 30, 2022, the day when chatgpt was introduced to the world will fall in history as a key moment for AI.
Soon after, Hamilton noticed: “Visitors asked:” Where is my chatgpt? “” Can I exploit chatgpt? ” “I do not have access to chatgpt.” “I need chatgpt.”
However, as one of the world’s largest digital payment suppliers, Visa naturally had concerns about the confidential data of its clients – in particular that she remained secure, except for the public domain and wouldn’t be used for future models training.
To satisfy employees’ demand during balancing these fears, Visa introduced what she calls “Secure Chatgpt”, which is behind the dam and works internally on Microsoft Azure. The company may control the input and output data through screening in the scope of stopping data losses (DLP) to make sure that no confidential data leave VISA systems.
“All hundreds of petabytes of data, everything is encrypted, everything is safe at rest, as well as in transport,” explained Hamilton.
Despite the name, Secure Chatgpt is an interface of many models offering six different options: GPT (and its various iterations), Mistral, Anthropic’s Claude, Meta Lama, Google’s Gemini and IBM’s Granite. Hamilton described this as models-as-a-service or RAG-AS-A-Service.
“Think about it as a layer in which we can provide abstraction,” he said.
Instead of individuals building their very own vector databases, they will select and select the API interface that most accurately fits their specific case of use. For example, if they simply need a little tuning, they sometimes select a smaller model of an open source, resembling Mistral; However, if they are looking for a more sophisticated model of reasoning, they will select something like OpenAI O1 or O3.
In this fashion, people do not feel limited or as if they lose what is easily accessible in the public domain (which might result in “shadow AI” or the use of unprofessional models). Safe GPT is “nothing more than a shell at the top of the model,” explained Hamilton. “Now they can choose the model they want.”
In addition to secure chatgpt, all developers have access to Github Copilot to assist in on a regular basis coding and testing. Hamilton noticed that programmers use Copilot and plugins for various integrated programming environments (IDES) to grasp the code, improve the code and perform unit tests (determining that the code works as intended).
“So covering the code [identifying areas where proper testing is lacking] It increases significantly because we have this assistant – he said.
RAG-AS-A-SERVICE IN ACTION
One of the strongest cases of use for secure chatgPT is support for questions related to the policy specific to a given region.
“How can you imagine, being in 200 countries with various regulations, documents can be thousands of thousands, hundreds of thousands,” noted Hamilton. “This is really complicated. You have to nail it, right? And it must be an exhaustive search. ”
Not to say that local policies change over time, so Visa experts should be current.
Now, with a solid cloth based on reliable, current data, Visa’s artificial intelligence not only quickly regains answers, but provides quotes and source materials. “He tells you what you can do and what you can’t do, and says:” Here is the document you would like, I answer based on this ” – explained Hamilton. “We narrowed the answers with the knowledge we built into a rag.”
Usually a exhausting process would require “if not hours, days” to attract specific conclusions. “Now I can get it in five minutes or two minutes,” said Hamilton.
Four -layer infrastructure of “birthday cake” Visa
According to Hamilton, these possibilities are the result of huge investments in data infrastructure over the past 10 years: the financial giant has spent about $ 3 billion at a technological stake.
He describes this stack as a “birthday cake with 4 layers”: the foundation is a layer of “data platform as a service, with the AI and Machine Learning (ML) ecosystem, services and products and layers of products built upstairs.
Data-Platform-AS-SERVICE mainly serves as an operating system built on the lake of data, which aggregates “hundreds of petabayes of data”, explained Hamilton. The layer above, data as a service, serves as a sort of “data highway” with many belts with different speeds to provide a whole lot of applications.
The third layer, the AI/Ml ecosystem, is a place where Visa is continuously testing models to make sure they do how they need to and are not prone to prejudice and drift. Finally, the fourth layer is a place where Visa builds products for employees and clients.
Blocking $ 4 billion
Being a trusted payment supplier, one of the most vital visa priorities is to stop fraud, and AI also plays an increased role here. Hamilton explained that the company has invested over $ 10 billion to assist reduce fraud and increase network security. It finally helped the company Block $ 40 billion for a fraud attempt Only in 2024.
For example, a latest VISA tool for deep authorization ensures a risk assessment to assist manage card-not-present (CNP) payments (for example, when users pay via the web or mobile application, identical to on a regular basis practice for all of us). It is driven by a model of a recursive neural network (RNN) based on petabytes of contextual data. Similarly, real-time payments, account account (think using digital wallets or immediate payments)-is enabled by deep AI learning models, which produce immediate risk results and routinely block bad transactions.
Hamilton explained that Visa used a model based on a neural transformer-which learns context and meaning by following data in data-to improve these tools and rapid identification and thwart of fraud. “We wanted to do it according to transactions,” he said. “This means that we have less than a second, I should say milliseconds, reaction times.”
Synthetic data also ensure the value of fraud prevention: Hamilton’s syndrome extends existing data to synthetic data to simulate newer calculations of fraud. “It helps us find out what is happening now and what can happen in a short -term and long -term perspective so that we can simulate and train the model to catch the data,” he said.
He noticed that fraud is a arms race – and there is a very low entry barrier to threatened actors. “We must be one step ahead of this, predict and block” – emphasized Hamilton.