Rajat Taneja says Visa detected and stopped over $40 billion of fraud in the previous year using its AI models.
“Last year, we detected and stopped over $40 billion of fraud with the models we have.”

2 Apr 2025 · 24m · Rajat Taneja, Craig S. Smith · 19 quoted facts
Visa argues that it has been an AI company for over 30 years, and that its advantage is a huge, purpose-built data platform plus in-house predictive models. Rajat Taneja says these systems stop fraud, keep payments flowing when banks or networks fail, and now underpin a push into "agentic commerce," where AI agents shop and pay for people.
The central tension is between invisible, trust-first infrastructure and a more ambitious future role. Taneja says Visa is not an e-commerce platform and wants others to innovate on top of it. Yet identity, biometrics and agent payments are moving the checkout upstream, so Visa must authenticate people and software before a purchase happens. He stresses consent, privacy and compliance throughout.
Taneja says Visa trains, runs and manages its predictive neural-net models in production entirely itself. For newer generative AI it uses an abstraction layer that lets it swap among six models.
When a bank or telecom link fails mid-transaction, Visa's models mimic what the bank would likely have decided. Consumers and merchants are therefore not affected.
Visa rebuilt its data platform from scratch about a decade ago and has spent over $3.3 billion on it. Taneja says it now holds over 500 petabytes, and that older data keeps improving the models, like aging wine.
Without a physical card, Visa must verify people and their AI agents by other means. It points to tokens (about 13 billion), passkeys with biometrics, and a planned April 30 product announcement.
Visa scores the manner in which credentials are entered, treating typing as a unique signature. A suspicious pattern can trigger step-up authentication rather than a blocked transaction.
Rajat Taneja says Visa detected and stopped over $40 billion of fraud last year using its models. He describes the models as deep neural networks, a kind of AI that learns patterns from past data.
Taneja describes it as AI agents helping people discover, choose and buy products, including handling chores like warranties and price matching. Visa cares because every such purchase still needs a trusted payment and transaction layer.
For its existing fraud and risk products, Visa builds and runs its own predictive models. For generative AI, Taneja says it uses six different models behind an abstraction layer, so models can be added or removed as they evolve.
Taneja says Visa uses behavioral analytics, including how a person types, because it is a unique signature. If the pattern looks wrong, Visa can ask for extra verification or decline the transaction.
No. Taneja says Visa is the underlying platform supporting merchants and banks, and that other companies will build innovations on top of it through its 2,900 open APIs (interfaces that let outside software connect).
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Rajat Taneja says Visa detected and stopped over $40 billion of fraud in the previous year using its AI models.
“Last year, we detected and stopped over $40 billion of fraud with the models we have.”
Taneja says Visa serves 4.7 billion credentials on one side of its network and over 150 million merchants on the other.
“we serve 4.7 billion visa credentials. And on the other side, there are over 150 million merchants.”
Taneja says Visa processes over 300 billion transactions in real time in a typical year.
“in a typical year, we will process over 300 billion transactions in real time.”
Taneja says Visa runs at 99.9999% uptime ("six nines") and uses AI to detect and get ahead of operational issues.
“We run at six nines, 99.9999 % uptown. And so detecting and finding and being ahead of operational issues, AI is a big tool in that.”
Taneja describes enumeration attacks as attempts by criminals holding stolen card data to find out which cards are still active, and says Visa's models are built to counter them.
“Enumeration attacks when bad guys steal or have access to a lot of card data, and they're trying to figure out which card works and which doesn't”
Taneja says Visa trained models to mimic issuing banks' decisions so it can process transactions on a bank's behalf when communication fails.
“we have trained our models to be able to mimic what they might have done over years and decades of training our model on the data”
Taneja says Visa has been committed to AI for over 30 years, continuing to invest through the so-called dark winter of AI.
“we have been committed to and a deep believer in AI for over 30 years”
Taneja says Visa trains, creates, runs and manages its existing predictive AI models entirely in-house.
“We train the models ourselves. We create it ourselves. We run it ourselves. We manage the model in production 100 % ourselves.”
Taneja says over 100 Visa products use AI, mostly predictive neural-net models.
“over 100 products of ours have AI, these are essentially predictive AI models, neural nets.”
Taneja says Visa has worked with generative AI techniques since 2018 and currently uses six different generative models.
“As Generative AI has come out, we have been working with Generative AI techniques from 2018 onwards, actually. We have embraced six different models right now”
Taneja says Visa has spent over $3.3 billion on its rebuilt data platform, which holds over 500 petabytes of data.
“We have spent over $3.3 billion on this data platform in this time period. and we have over 500 petabytes of data.”
Taneja compares Visa's data to grape juice turning into wine, saying it has instant value and gains value over time as it trains and improves models.
“I liken it to fresh grape juice, which has great value, and then wine. Over time, it ages and becomes more and more valuable.”
Taneja says Visa planned a major agentic commerce announcement on April 30, a product drop led by its CEO and colleague Jack Forrestal.
“And April 30th, we have a major announcement of this that I hope you will join”
Taneja says Visa now exposes 2,900 APIs and handled about 40 billion API calls a month at the end of the previous year.
“We now have 2 ,900 APIs. We do about 40 billion API calls a month at the end of last year.”
Taneja says Visa has invested in payment tokens as a replacement for physical cards and has around 13 billion of them.
“we've already been investing in that with tokens, which we have around 13 billion tokens now, approaching 13 billion.”
Taneja says Visa announced Visa Passkeys the previous year so that biometrics can be part of payment authentication.
“We announced last year Visa Pass keys that allows biometrics to be part of this”
Taneja says Visa scores how a person types, treating typing as a unique signature, to help detect account takeovers.
“We are able to see how you type. and everybody's typing is a very unique signature.”
Taneja says Visa released a paper on the future of shopping with an outside institute on the day of the interview.
“We just released today a paper in collaboration with the Institute of Future that jointly with us is trying to imagine that.”
Taneja says Visa operates in 200 countries and territories and must comply with laws and regulations everywhere as AI agents develop.
“We operate in 200 countries and territories and so for us, as these develop”
Taneja introduces Visa as a huge payments network where trust depends on AI across risk, cybersecurity and uptime.
“in a typical year, we will process over 300 billion transactions in real time”
Explains the core technology: ensemble neural networks that separate good from bad transactions and catch attacks such as card enumeration.
“we mostly use deep neural nets”
Shows a less obvious use of AI: keeping payments working when bank or telecom links fail.
“trained our models to be able to mimic what they might have done”
Visa builds and runs its predictive models itself and uses a layer that lets it swap generative models.
“We train the models ourselves. We create it ourselves. We run it ourselves.”
Describes the rebuilt data foundation that Taneja credits for Visa's later AI innovation.
“decided to rebuild it from scratch for the AI era”
Taneja outlines Visa's vision of agents transforming shopping while Visa stays the trusted payments layer.
“shopping is still in the hunting and pecking and hunting era”
As physical cards fade, tokens and biometrics become the way Visa verifies who is paying.
“As you move to e-commerce and then agentic commerce, identity will move upstream.”
A concrete example of behavioral biometrics used to stop account takeovers.
“We are able to see how you type.”
Taneja closes on how agents will play a role only if consent, privacy and compliance are maintained.
“AI agents are going to play this part.”
Last year, we detected and stopped over $40 billion of fraud with the models we have. But we also take it forward. For example, a typical transaction will go from the merchant to their bank. It comes in real time to us. We send it to the issuing bank that has issued the credit card or the debit card to the consumer, and they then send back their signals to us. We process it and send it to the merchant.
Now, in all of that, there are occasions when the communication doesn't occur seamlessly, where maybe there's a telecommunication problem or the bank has issues on their side. We have trained our models to be able to mimic what they might have done over years and decades of training our model on the data to say, how can we stand in and process the transactions so the consumer and the merchant are not impacted?
Create an oasis with Thuma, a modern design company that specializes in furniture and home goods. By stripping away everything but the essential, Thuma makes elevated beds with premium materials and intentional details. I'm in the process of reorganizing my house, and I'm giving Thuma a serious look for help in renovating and redesigning. Thuma combines the perfect balance of form, craftsmanship, and functionality. With over 17 ,000 five-star reviews, the Thuma Bed Collection is proof that simplicity is the truest form of sophistication.
Using the technique of Japanese joinery, pieces are crafted from solid wood and precision cut for a silent, stable foundation. With clean lines, subtle curves, and minimalist style, the Thuma bed collection is available in four signature finishes to match any design aesthetic. Headboard upgrades are available for customization as desired. To get $100 toward your first bed purchase, go to Thuma, that's T-H -U-M -A dot C-O slash IonAI. IonAI all run together, E-Y -E-O -N-A -I.
So for $100 off your first purchase, go to Thuma dot C-O slash IonAI. That's T-H -U-M -A dot C-O slash IonAI to receive $100 off your first bed purchase. I usually start by having you introduce yourself to listeners, and then I'll start asking questions from there. Introduce yourself to listeners and tell us why Visa is at an AI conference.
My name is Rajat Taneja. I'm the president of technology for Visa. And we are one of the original practitioners of AI. And we are delighted to be part of the conversation here in HumanX, thinking about the future, helping design the future as well with AI.
And I know that you guys have used AI for a long time in fraud detection. And I was interested with Gen.AI expanding the attack surface for fraud attacks. How are you keeping up with that?
So AI is in the very fabric of Visa at some level and has been for decades. And if you kind of step back and think about what Visa does, we are one of the world's largest SaaS platforms. One side, we serve 4.7 billion visa credentials. And on the other side, there are over 150 million merchants. So in a typical year, we will process over 300 billion transactions in real time. And we process about $16 1 trillion through our network.
And so in order to ensure trust between the buyer and the seller, it's very important to make sure that they know that the transaction will be authorized correctly, that money will flow accurately. In the case of any dispute, Visa will be behind it, and we will manage any fraud and make sure we are fully compliant. So AI has helped us in all of these factors through the decades. Our first model was to manage risk in a transaction, as you mentioned.
But subsequent to that, AI is part of our cyber security program, is part of our operational infrastructure. We run at six nines, 99.9999 % uptown. And so detecting and finding and being ahead of operational issues, AI is a big tool in that. It's across everything we do and has been. And generative AI, of course, takes us into the next generation of AI. And we're excited and thrilled to be part of that.
Yeah, and specifically on fraud detection, is it pattern matching and uncovering anomalous features in a pattern? I know it's much more complicated than you can explain in a few minutes, but what's the basic tool behind fraud detection?
So we mostly use deep neural nets. And at the very core of those is patent matching and training the model to understand what a good transaction is and what a bad transaction is so that the good transactions flow through, the bad transactions are stopped, and the false positives and false negatives are very minimal. And we use different techniques in that, and typically our model is an ensemble model. And risk is one. Payment fraud, disruption is the other one.
Enumeration attacks when bad guys steal or have access to a lot of card data, and they're trying to figure out which card works and which doesn't, and then utilize the ones that are still active to do a lot of nefarious things. So there are called enumeration attacks where we are on top of that, e-commerce disruptions, fake sites are set up for e-commerce. So we are able to signal that. Last year, we detected and stopped over $40 billion of fraud with the models we have.
But we also take it forward. For example, a typical transaction will go from the merchant to their bank, the acquiring bank. It comes in real time to us. We send it to the issuing bank that has issued the credit card or the debit card to the consumer, and they then send back their signals to us. We process it and send it to the merchant. Now, in all of that, there are occasions when the communication doesn't occur seamlessly, where maybe there's a telecommunication problem or the bank has issues on their side.
we have trained our models to be able to mimic what they might have done over years and decades of training our model on the data to say, how can we stand in and process the transactions so the consumer and the merchant are not impacted? And that's another use case of a deep neural net that we have been using.
Is this your own model or are you fine-tuning an open source model or do you partner with
anthropic or somebody like that? So for the existing models that we have, and over 100 products of ours have AI, these are essentially predictive AI models, neural nets. We train the models ourselves. We create it ourselves. We run it ourselves. We manage the model in production 100 % ourselves. And we have been committed to and a deep believer in AI for over 30 years. during the dark winter of AI, where nobody sort of really believed in it, Visa continued to invest and had deep belief in it.
As Generative AI has come out, we have been working with Generative AI techniques from 2018 onwards, actually. We have embraced six different models right now, but we have created an abstraction layer that allows us to add, delete, remove models as they evolve and have an ensemble. And then we connect those to our own data where we need to. And then we enable direct access and through other application layers for various use cases.
Yeah. And, you know, so Visa, unbeknownst to the public, has been an AI company since the 90s, Yes, indeed. Is it frustrating that you're not recognized for your AI work or potential?
You know, we're a humble company. We run behind the scenes. Our goal is to uplift everyone everywhere by being the best way to pay and be paid. And as long as that mission is being accomplished, we are serving the community, we are serving consumers, we are serving merchants. we are okay with us developing quietly our technologies and manifesting itself in the seamless experience where trust is established and ubiquity is established for consumers and
merchants to do business in very nice ways. Yeah, you know I use, it's now called Credit Karma, It used to be called Mint. Yes. And you can connect all your different accounts, and it downloads your balance and adds it all up and does all sorts of things. As those kinds of apps proliferate, it just seems incredible that you can keep track of data flows. it's like this spaghetti bowl of connections and microservices that are talking to each other.
How do you manage that? Is there some visualization somewhere or is it all left to the AI or NAI system to keep track of?
You're absolutely right. Our back-end data platform is one of the largest and most sophisticated in the world. And we, in fact, a decade ago, realizing where the puck was going, decided to rebuild it from scratch for the AI era. And it was instrumental in allowing us to do so many innovations that we have done since then. We have spent over $3.3 billion on this data platform in this time period. and we have over 500 petabytes of data.
And every transaction brings data, every API that we expose. And companies like Credit Karma provide data to us in different ways, structured data, unstructured data, document data, graphical data. And we have a very systematic way in which we ingest the data, store the data, process the data. And that is processed by AI, by applications and the business logic of those applications. It has great value at the moment the data comes to us, but then over time continues to add value.
I liken it to fresh grape juice, which has great value, and then wine. Over time, it ages and becomes more and more valuable. So that's the way we have our data. It immediately has instant value. and then collectively over time continues to train and improve our models and our products.
Yeah, everyone these days are looking at agentic systems, multi-agent systems. Are you using agents in this system to track transactions or highlight possible fraud?
So a lot of our current products that are using generative AI are using the essential underlying technologies of generative AI. And we are seeing the benefits of that in both productivity and software development, in client services, in our business operations, and so on. But as you're pointing out, one of the more exciting things that we are envisioning for the future is agentic commerce that will have a fundamental profound impact on the way buyers and sellers discover each other and have deeper connections and relationships over time.
We are big believers that while a lot of great things have happened over the years in payments, shopping is still in the hunting and pecking and hunting era and it's not as easy and that will be transformed with agentic commerce the ability to take the mundane aspects of commerce like registering products warranty manuals price matches and things like that agents helping with that and then bringing the joys and pleasure of commerce discovering new things travel art music, so on, products in new ways.
But the fundamental common constant over there is payments and transactions, and how do you enable that with agents talking to agents or an agent representing you. And we are spending a lot of time on this invention. And April 30th, we have a major announcement of this that I hope you will join, your viewers will listen in where our CEO, Ryan Beckinerny, and my colleague, Jack Forrestal, are going to do a product drop. And it's going to be a very exciting day in commerce.
Yeah, but you guys are not an e-commerce platform. Are you headed in that direction?
We are the platform that supports. We are the platform that supports merchants and acquiring banks and issuing banks. And that platform has to provide the same level of trust and cybersecurity and payment security and ease and reliability complies. And so we are the underliking and continue to be the top that much.
But when you talk about matching and helping consumers find products, is that a function that Visa will play or that's the function that the e-commerce platform plays and you support?
I think there will be a lot of innovations in the ecosystem. And we love those innovations, and we expose our capabilities to open APIs, which was a big architectural investment we made, again, about a decade ago to open up our platform. We now have 2 ,900 APIs. We do about 40 billion API calls a month at the end of last year. And so the innovations will happen on our platform, and we will continue to provide seamless capabilities for the transactions to take place accurately, for the clearing to happen, for the settlement to happen, for the compliance to happen, risk to be managed.
But the great deal of innovation and inventions will happen by others in the ecosystem partnering with us, partnering with each other, and so on.
Yeah, I'm hearing a lot about distributed systems, blockchain identity, you know, registering identity on the blockchain and all of those things. Is Visa looking toward that? I mean, I still use a card or I use my Apple wallet, but it seems that the verifying the identity of the user and how the user is paying, whether it's through what modality, all of that is changing very quickly. How do you see that developing?
I think the change has already begun. I mean, in the face-to-face commerce, you have an identity that is manifested by a physical card and your ID that is needed. We have a chip. There's a lot of information that transmits for us to know you are who you say you are, and you can do what you say you want to do. As you move to e-commerce and then agentic commerce, identity will move upstream. You no longer have the physical card, and we've already been investing in that with tokens, which we have around 13 billion tokens now, approaching 13 billion.
We have been investing in mobile and commerce with different mechanisms to authenticate the credential, to authorize and score the transaction and ensure all the downstream processes work correctly. Now, take that to the next level with agentic commerce. That's going to be a big area of investment and inventions and innovation. Biometrics will play a big part over there. We announced last year Visa Pass keys that allows biometrics to be part of this, and then it's going to continue to evolve.
And that shows an authentication and identity, as you're pointing out, is a key area of emphasis for us to ensure accurate transactions that the buyers and sellers can trust. The platform will make sure that disputes or frauds and all the other things that are happening can be taken care of.
On this identification, verification, as you said, part of it's going to take place at the agentic level. but for personal ID with biometrics, how is that going to work? Are you going to use face scans, iris scans, fingerprints, voice prints?
It depends on the use case, but in many cases it is all of the above, depending on the device you have that is giving us some information, or we are able to collect it in different ways. I'll give you one example of an invention we have at Visa. This is for making sure accounts are not taken over nefariously by bad actors. By stealing your password and your user ID or by social engineering or phishing, doing that.
We are able to see how you type. and everybody's typing is a very unique signature. So we are able to score it with AI, behavioral analytics of actually not just the credential itself, but the manner in which the credential was entered and signal if there is a need for us to do a step-up authentication or to rather than the transaction with more details. And that's a manifestation of, in that particular use case, based on the signals we got, how do we protect the consumer?
Because our main session, I would call it the word, is to protect the consumer and the merchants and our ecosystem. And in different areas, it will be a combination of what we are able to get and add on to it with our own ideas and innovations.
Yeah. Okay. Well, I'm going to ask one more question. What do you see? How do you see the future for an individual who's browsing websites or walking through a store or just imagining something that they want to buy? Do you have a vision for how that's going to develop?
We just released today a paper in collaboration with the Institute of Future that jointly with us is trying to imagine that. How will it look like in the future? And it is a very interesting, thought-provoking future that is still being developed. and the role we play and how we'll adapt and shape the future is going to be one of the most exciting things around us. But one thing is for sure, AI agents are going to play this part.
How and what role they're going to play. Depend on every consumer's consent, every consumer's personal preference, every consumer's comfort with transparency, with managing the privacy of the data by knowing that the value they're getting, whether it is in browsing, whether it is in discovering, whether it's an actual purchasing for you. We're presenting you as a personal assistant. We need to make sure that the trust is maintained and it is fully compliant with every law and regulation of the world.
We operate in 200 countries and territories and so for us, as these develop that confides that trust that cyber protection the came and protection must paramount but i think it evolved in very exciting ways but it's important for us to be looking right now at things that may not work well yeah that could cause problems and have both the optimistic mindset and the pessimistic minds thinking through that you're shaking the future correctly but Many of the things you said will probably occur sooner than we think right now.
Create an oasis with Thuma, a modern design company that specializes in furniture and home goods. By stripping away everything but the essential, Thuma makes elevated beds with premium materials and intentional details. I'm in the process of reorganizing my house and I'm giving Thuma a serious look for help in renovating and redesigning. Thuma combines the perfect balance of form, craftsmanship, and functionality. With over 17 ,000 five-star reviews, the Thuma Bed Collection is proof that simplicity is the truest form of sophistication.
Using the technique of Japanese joinery, pieces are crafted from solid wood and precision cut for a silent, stable foundation. With clean lines, subtle curves, and minimalist style, the Thuma Bed Collection is available in four signature finishes to match any design aesthetic. Headboard upgrades are available for customization as desired. To get $100 toward your first bed purchase, go to Thuma, that's T-H -U-M -A dot C-O slash IonAI. IonAI all run together, E-Y -E-O -N-A -I.
So for $100 off your first purchase, go to Thuma dot C-O slash IonAI. That's T-H -U-M -A dot C-O slash IonAI to receive $100 off your first bed purchase.