Agentic AI Powering Instant Intelligence

Hi everybody. Thank you for joining us gentlemen. Um I'm really excited for this conversation because I think agentic AI is quickly becoming one of those things that is thrown around a lot and can mean a lot of things sometimes can mean nothing. and and so I'm hoping that this group of panelists can put some details around it and give everyone some um some actionable intelligence to move forward. Um Louie, we are in your house.

So we'll we'll start with you um chief digital officer at ADQ. Can you actually just give us a your definition of agentic AI versus traditional AI and machine learning and maybe one use case today and maybe one aspirational use case?

Sure. Can you hear me? Well, so um if I want to oversimplify it, an agentic eye is an equivalent eventually to a full-time employee and that's how we're trying to measure it. So when we build an agentic tool, it has to replace almost a full-time employee. Um in in in very simple terms uh one use case that is very obvious is you know running due diligence for example on on transactions that uh we undertake.

Uh it's fairly one of the best use cases that you can use today beyond a couple of others in procurement and so on and so forth. But this is going through tons and tons of data analyzing the data and making sense of it uh for um you know for for whoever is going to make the decision on that transaction eventually right so synthesizing it and you can create multiple agents for different topics be it uh you know the commercial due diligence be it the legal due diligence be it the technical due diligence so on and so forth and those can actually eventually be orchestrated to create a common outcome on an aspirational basis. I think the way you would want to look at it is a bit more iterative and over the over time you want to actually create more of integrated agentic solutions that can work together from different parts of the business. Uh and that allows for a more cohesive workforce that is the hybrid workforce that everyone speaks about for the future. uh and that's aspirational today because it means everyone has to build those tools and allow them to kind of interact and and work together.

Thomas, a similar question to you. You know, we've seen these huge AI announcements just gobs of money about data centers and investments in LLMs. And one thing we haven't seen quite so much is the kind of enterprise spending and and you're mostly a a B2 a B2B business. What are you seeing from the end customers? where are they really spending money?

Where do they see the promise in agentic AI?

Yeah, thank you. Thanks for having us on the panel. And I think on Lewi's point, what's aspirational is when you can get the AI and the human work forces to be together, the processes happen. What has happened in the enterprise very often we looked at it from a cost point of view and that makes it very hard to price the value of AI right but I think we are seeing organization ask ourselves and even countries and governments say how can I push the GDP further? How can I push the top line further?

And Joffrey and I were just discussing this. If you look at value creation as the price of AI, then you first need to start with a telemetry of thinking what do you want the AI to do. If I'm in a bank and I want the wealth management portfolio to grow by 50%. I'll take a percentage of that to pay for the intelligence that I want to get out of it. So the questions around where apply AI is not in risk, not just in risk, not just in compliance, not in manpower savings, but really what revenue are you going to get out of it?

How much more territory will you get out of it? And I think that is where we're seeing customers start to price AI.

Godfrey, you're at Visa, which processes something like 65,000 transactions a second, um, and has some obligation to try to spot fraud before it happens. And I'm curious, you know, how are they using how are you using AI not just to to spot it, but to, you know, self-directed stop it in its tracks? Are these sort of this is a smart contract, it sees something it doesn't like and it shuts it down without a human intervening?

Uh, so we've been using AI in our payment systems and across the ecosystem for about 30 years. Uh and every time a transaction is made in the 200 milliseconds on that transaction completing between your card and the associated bank, we run risk and fraud tests on that transaction and send that data back uh to the participants in our ecosystem so they can decide on a transaction and whether it's a good transaction or or or not. What we're seeing now is that we are on the cusp of the next era in commerce. Uh we had the e-commerce era starting in the 1990s. We shifted to mobile commerce.

Uh and now we're at the cusp of aentic commerce. Uh which we think is a really exciting transformation in the way in which people shop. Uh and so what we recently launched building on our AI capabilities is something called Visa intelligent commerce. So that we can take the trust of the Visa ecosystem and imply that to the the many AI agent firms who are operating in our world today whether this is Chat GPT or Grog perplexity or or the others because what what people are doing we're seeing is we're you they're using those engines to search and browse for transactions but they're not completing the payment transaction. They're going into an offline process, taking out their card, entering in a 16-digit number.

Now, we've integrated with Visa intelligent commerce all of our trusted technology to ensure that you can complete an end-to-end agentic transaction, ask Chat GPT to to purchase a weekend away with flights and hotels and restaurant bookings, etc. uh but have the trust and security of the Visa network every step of the way with the customer authentication and approvals that go in so that you can use your card throughout that transaction are operated by the by the agent and we still have the dispute processes the fraud management processes that sit behind that that if there is a problematic transaction you can contact your issuer and resolve that

but is there a difference from a fraud detection perspective when you know we get all the way to the end of the process and I didn't dig out my credit card, there's there's some human fingerprint on that transaction that's missing that is going to spook the issuer, the merchant, the rails.

So what we have done is to replicate all of the trusted processes that we have in the traditional e-commerce environment and replicated those in the agentic environment. So we have a series of five APIs which include payment instructions. So you as the consumer is always put in charge around completing the transaction or authorizing your agent to complete the transaction. We have a payment signals API which provides the issuer with even more data around the nature of that transaction. So if it is in dispute they can they can remedy that with the appropriate merchant.

Otto, a kind of a similar question on on accountability. You know, PayPal, I think, is the most consumer-facing business that's represented up here. Um, in addition to payments, you also extend credit. I mean, these are decisions that when the person receives them, they feel pretty strongly. How do you think about speed versus explanability?

How comfortable are your users with the explanation? The computer decided, the AI decided?

Thank you, Le. Uh it's great to be here this incredible event today. Uh we cannot overemphasize or ignore the fact that agency commerce AI will be happening. I think we've just recently seen Visa data that says 47% or so of the US uh consumers used agentic tools to do their shopping over this Black Friday and Cyber Monday holidays. We're seeing the data.

We've seen uh Morgan Stanley data that says by 2030 10 to 20% of ecom will be agentic commerce and that's about 190 billion to over $300 billion in payments will be agentic. We have to address agentic commerce and whether we like it or not it's happening. It will continue to grow. Consumers prefer it and they want to interact with those tools. Now the question is how do you actually make it trusted?

If you think about it the success or what it'll require for a gentic commerce to be successful are three things. Its trust, its identity and its orchestration. Now you know think back 25 years ago when PayPal was founded and we were solving this trust issue on the internet. where you wanted to make a purchase from an eBay seller cross border. You don't know that seller.

You like what you see online. You've not communicated. The seller does not know you, but you are willing to transfer funds to that seller and trust that you will receive a package at home in a week. And the only reason why you did that was because you trusted that PayPal has identified that seller and knows that seller and the seller trusts that PayPal has identified you, the buyer, and knows you. And so if they ship those those goods, they will get paid.

And if you make that payment, you will receive your goods. That is the trust framework that we've spent the last 25 years building on the internet. That trust framework doesn't exist for Aente Commerce today. We actually have to rebuild that. And that's the partnership that we've entered with Perplexity, with OpenAI, with Grock, is to build that trust framework because how do you if you're a consumer that does an agentic search and makes a purchase, how do you know that that's a legitimate merchant?

AI agents are creating millions of merchant websites and receiving payments without shipping goods. If you're a legitimate merchant that shows up in AI search and you receive an order and ship goods, how do you know that that is a legitimate buyer? You have AI agents that are taking card numbers of the dark web, clicking on checkout buttons on merchant sites and submitting fraudulent orders. And so there is that trust framework is actually broken in the aentic commerce space today and that's what we're rebuilding. And the way that we're doing that is saying we've got over 40 million sellers globally.

We process 25% of the world's ecom transactions, $1.7 trillion. We are going to surface the inventory and cataloges of all of these merchants in the agent commerce framework so that they are trusted merchants and trusted inventory. And we are going to submit orders from trusted consumers and identified consumers into those merchants order management systems so that they can trust that those are legitimate buyers as well. That is how we're rebuilding that framework and this is extremely important otherwise when once trust breaks down, commerce breaks down.

Tom, can you pick up that thread a little bit? We've seen bad actors increasingly using AI. Are we headed towards some good agents versus bad agents arms race and how do you spot that? How do you try to interject friction in the right places without gumming up what is, you know, promising to be a pretty frictionless environment? Yeah, I think what Otto has done is clearly describe what the trust ecosystem should look like and when you expand that to where enterprise users are I think trust remain that that element and we said AI needs to be invisible if not it's not fictionist so on the biasness where there's good agent and bad agent and and data I think a big part of it starts with the the design of it so data where it sits data security ethical use of AI that's how we developed the first set of agent if you zoom out data sovereignty like what UA has done is what we take to guarantee the source of the data that's learning the agent itself is in the areas of where it's applied.

So I give an example when we team up with HSBC and FAB and the Amoris Institute of Finance what we try to do is not decide what agents what technology what large language model but really to define what is the outcome and use cases they want to do. It goes from risk compliance to to wealth management to operational intelligence. When you define the outcome well, it's like identifying different types of FTE employees in the mark in in in the in in your environment. When you do that definition well and you know where the source of information is coming from, that's really where you pay attention to opposed to getting into there's good agent and bad actors working to disrupt a trust system. Understanding the source and the outcome that you derive, you will then see the processes that you need to align with.

Yeah, I think Otto articulated it really well that trust is the key to unlocking this next wave, this next era of commerce. Uh what we have introduced as part of our uh efforts in that direction is now we're familiar with KYC and KYB. We now have KYA, know your agent. And so there's a partner program of those agents which are trusted agents which consumers can can engage with and they have the trust and security of the of the Visa network behind them. Now it's also interesting on the merchant side because you see players like Amazon who say we don't want to be disintermediated.

We also want to know who those end consumers are. Uh and there's a case at the moment between Perplexity and Amazon. And so as we build out that ecosystem, those merchants need to also have trust in the ecosystem that they know who they're dealing with and the loyalty points that they accumulate for their customers can be uh be shared directly with them. So the other component which we added on as we were developing this is trusted agent protocol that allows those merchants to see behind the agents that end customer and continue to maintain that relationship and share loyalty points and rewards and uh and these types of things. So it's a rapidly evolving space and we need to pivot and uh and come up with solutions to make sure that that trust component remains u yeah remains there.

I mean, it's interesting as consumers, part of the promise of agents that I don't have to do that and that I don't have to have, you know, I can buy what what I want based on what I tell without having to give away my data or get a bunch of, you know, direct mail that I didn't ask for. So, that last mile is is really interesting. You know, Louie, what does this all mean for work forces? And and everyone says, you know, you talk to CEOs and they say, "We're not we're not firing. We're probably not hiring.

We're slowly we want to sort of upskill people, turbocharge talent." Where do you see the weak spots? Where do you see the places where this can really help um people be more productive?

It's very easy to pick up on the same trail of um trusting agentic tools. When we're building agents, what we're doing is we're training them on the jobs that we currently do to some extent. Uh we're giving them uh a JD, but a JD that's much more detailed and descriptive about exactly what they should do and they shouldn't do. So when we think about the current workforce, what is needed is out of that current workforce is to train those agents and hence they need to be able to abstract their work and put that into some sort of guidelines for the agents on how to operate and that will differ significantly from one agent to another. But at the end of the day, you're trying to kind of get an agent to do part of your job and you need to train it as if it was a new recruit to deliver that part of the job and you allow yourself to do a bit more strategic work at the end of it.

And that's the reskilling that's needed. I mean of course there's some technical skills on how do you code agentic tools, how do you chain agentic tools, how do you manage agent and so on and so forth but I think the biggest transition is how do you allow people to kind of think about their jobs as frameworks and code that into agentic tools that are coming online.

Tom, I wanted to ask Godfrey was talking about kind of off-the-shelf agents versus proprietary agents who do a lot of work um with governments. Are there going to be are there going to be sort of open architecture agents that central banks around the world can plug in? Are where does security come in here? You know what should be proprietary? What should be behind the wall and in front of it when you think about you know connecting these systems?

I think that's a great question because what for instance what Abu Dhabi want to do is create a billion agents to support the workforce. But I I just want to pick on what the way said and your question. The agentic AI workflows and what is bringing to society. It is not a manpower replacement, right? It is an enhancement and and accelerator to outcomes we want to achieve.

So it's not about how much workforce you can replace. It's the speed of which you get the insights. Had you had a bigger workforce, right? So whether is the government or the enterprises, I gave you an example this morning. I said if there was a wealth manager, relationship manager, right?

Then typically you pair your RM with one associate. Associate finds the content of the client's whereabouts, what their preference are and you make one call in the morning. Imagine now you multiply that with a team of aentic uh customer associates who does a lot of work for you 24/7. When you make the first brief at 9:00 a.m.

You know exactly what what your client wants to do. You know exactly what the marketers behave. It is not about the cost of the benefit is six customer associate. The cost of the benefit is having that information precisely at the time that you need. Now you extrapolate that to your question on what a central bank and what the government would do.

It's about how do we create a workforce that essentiates what we have today. At the same time on the earlier point of sovereign data infrastructure with what UA is doing now large language models that are sovereign to the country but then you still need to train in on the data that you trust right so it's a combination of technology that pushes agentic AI to adoption but to me the main driving force is the outcome that you can quantify that is beyond manpower. So if we if we start our journey to say that oh this is the number of people I'm going to replace and therefore the cost of AI is X then that's the wrong path we're on. We should look at if I had this information two weeks, two months ahead of time, what is the gain I'm going to get out of this? It could be for my GDP, it could be for bid.

I think that's the true cost and the true value creation of your first question. How do you repric AI with all this capex going into the infrastructure?

Then Otto, you know, you oversee Middle East and North Africa, a lot of countries in the PayPal ecosystem. For a regulatory front, maybe one thing that you think is needed and one thing that you think regulators really should not do. I I would say that regulators should really be focused on making sure that small businesses are not left behind. If you're a small business today, you can sell more by investing in SEM, search engine marketing, SEO, and gain visibility. But how would you as a small business show up in a chat GPT prompt or perplexity prompt today?

Many do not know. Regulators should be really focused on this. How do we ensure that small businesses are not left behind? We are making sure that those small business cataloges are made available in those environments. That's what we are committed to.

If you're a small business, you don't have access to data, to infrastructure, and to data scientists. How do we ensure that you have access to AI tools to be able to do smart pricing, to be able to find customers, to be able to do smart ads, and to be able to do smart receiping. These are all AI tools that can transform small businesses and help them sell more. Regulators should be really focused on that, and that is one area that we're very focused on and that we're bringing tools for.

Well, thank you all for joining us. We'll do this again next year with a an AI agent in my chair and see how it does.

Excellent.