Agentic AI —Too Wild For Wall Street? | DC Fintech Week 2025

Great. Well, hello everyone and obviously thank you so much to Chris for such a warm welcome and I think, you know, before we kick off our panel on a gen tech AI, of course, a thank you to Governor Waller as well because I feel like he set up our panel so well. Um, but you know, just to give you guys a quick intro and an overview, I'm Elise Succi Watts. I'm the executive director of global digital finance and it's privileged to moderate this for you today. And before I go to my panelists to introduce themselves also with a quick rapid-fire question, I just want to set the scene a little bit as to what we're talking about when we say a gen tech AI and also why this matters for finance and for Wall Street.

And so, when we say this, we're talking about AI systems that can really reason, plan, and act autonomously. So, not just generating outputs, but actually taking the actions themselves and they combine that large language model reasoning with memory, tool use, and a feedback lens to actually execute multi-step tasks. And so, in finance, this matters a lot, obviously, which our panelists are going to dig into, but it's for portfolio optimization, compliance monitoring, risk management, payments execution. We're going to hear a lot more about those use cases, but why this matters right now is because it really makes a shift from advisor to actually operator within our financial systems. And so, to kick us off, I'd like each of our panelists to just introduce themselves and just in a very quick sentence, what a gen tech AI means to you, either in your line of work, in your personal life, or how you're thinking about it developing for financial services.

So, Bob, can you go first?

Yeah, you bet. Uh, Bob Bench, co-founder and CEO of Radius. Uh, we build linearly scalable smart contract execution networks. Uh, in short, we make money move extremely fast. Um, AI agents for us are we try to think of them simply.

They're applications on top of probabilistic models uh, that based on the current path, they'll be getting more and more accurate every single day we work on those probabilistic models.

Todd.

Okay, so the short answer for what is Agentic AI? It's the future of commerce. I'm Todd Fox. I lead the Visa Institute. I've been working on uh, payments policy at the intersection of innovation, finance, and regulation for the last 13 years.

And uh, it's really awesome to be part of this conversation today.

Sarah.

Hi all. Sarah Morgenstern, I'm a venture partner at Flourish Ventures. We're an $850 million global early-stage fintech fund. Uh, we back entrepreneurs that are transforming the financial services system for the better. And my three words are groundbreaking but early.

Love it. Kendall, last but not least.

Hi, I'm Kendall Howl. I'm an associate Davis Polk. I'm in our financial regulation practice where we advise uh, banks, broker-dealers, other financial institutions on really anything um, including AI. Um, three words or less for Agentic AI for me is efficient delegated authority.

I love that as well. All really good. And I like that we all had different answers, too. But I want to come back to Sarah first because I think as we kind of set the scene and think about, you know, what are the opportunities? What can be done with this?

I'd really like for you to tell us a bit more about, you know, for a consumer or a small business, given what you do, what would their financial life look like if it's being managed proactively by AI? And you know, what are the early signals you're seeing in that fintech pipeline that can kind of point us towards that that future vision?

Thanks, Alissa. You framed this up so beautifully. Um, I think over the last decade we've seen mobile-first players like Chime really reshape consumer finance and we've made banking simpler and more affordable for millions and we've driven down the share of unbanked Americans by nearly half. Um today with Agentic AI we are potentially on the cusp of the next wave. And it is a world where finance is not just mobile-first but increasingly self-driving.

And as you described, that means we've got software that is reasoning and acting on a user's behalf in order to reduce the friction and cognitive load of managing money and in doing so potentially really help them to build financial health. So, what could that look like? For a consumer, that could mean uh an agent that is automatically sweeping idle cash from a checking account to a savings account overnight, right? Sort of quietly maximizing yield without any manual effort. It could be a refinancing agent that's monitoring interest rates and restructuring debt when cheaper credit becomes available.

For small business owners, there's equally powerful use cases. Uh so, imagine a agent that can anticipate inventory and supply needs, source trusted vendors, negotiate better terms, and automate purchasing. You're effectively giving a small business owner enterprise grade procurement power without the overhead. Uh or an accounting agent that can reconcile transactions and sort of close books in near real time. All of this is in and around CFO in my pocket that's helping to do the heavy lifting and is especially helpful for households and entrepreneurs that have very little margin for error.

So, that's all really exciting. Um and while there is tremendous promise here, we are also coming back to the theme still very early, and there are a lot of unanswered questions, many of which we're going to unpack together as a group, around bias, model drift, liability issues. So, in short, inspiring, but we got a lot of work to do to get from here to there and to do it right.

And I love what you said in particular about small businesses, because obviously another topic that we've heard a lot about from various panels and keynotes today is stablecoins. And a lot of those use cases that you mentioned, especially, you know, on thinking about inventory and supply chain management, actually using that kind of agentic portion can supercharge what you can actually do with programmable payments. And so, they really do go hand in hand, but it's, you know, especially pertinent for small businesses who might not have had access to these type of tools or to the people that would have executed them before, but they can hopefully soon afford them at a better price point. But, obviously it is, as you said, very early, but it's also not all rosy. And I think we would be remiss in our scene-setting if we didn't also talk about some of the risks and what this could mean for market infrastructure.

So, maybe I could come to you, Bob, next to talk about what this might mean in a stress scenario. So, if we have an, you know, economy increasingly run by agents and having agents peppered throughout, what happens if something breaks? And, you know, what breaks? And what part of the human financial infrastructure do you think, you know, won't be able to necessarily handle integrating these agents at scale?

It's always great for an early-stage startup founder to talk about pessimism. I find it's really good for the business model. Um, so, what we think about is less the kind of what breaks in in the sense of bad things are going to happen, but what type of infrastructure is necessary to be built to take advantage of this architecture. You know, when we think about AI agents doing everything you just spoke about, that's not a single line item a single execution of code, right? To say for example my automated asset manager needs to make a new allocation decision of cash management, that could quite literally be tens or hundreds of thousands of direct API calls for unique forms of data that may or may not be public or private.

So, how fast does that software execution system need to be to make 10 or 100,000 API calls to get the data necessary from the sufficient data feeds to say we are going to go into a 4.5% government money market fund. And what is the execution system available right now in the world that can do that? The answer is none. At least until we go live in 2 months.

Um, and and I and you we have Vinnie here on the table. I think about Visa a lot, right? When when Dee Hock thought up the Visa network, there was nothing in Bank of America's systems that could do the transactional to transactions necessary for the Bank of America card system to work. So, Dee Hock in his genius went and found the airline reservation system called Sabre, which had the only system that was distributed and had enough latency sufficient latency to enable his vision for Bank of America which later became Visa to occur. When we built Radius, we thought about, okay, we're going to ask an agent to the classic example is make a flight to Singapore with the hotels for Singapore Fintech Fest, which is a great event by the way.

Um,

This one's really good, too.

Yeah, Chris is well, Chris is this is the best This is the

This is the best one.

the best one, but you know, I we all we've already made our reservations to DC when you go to Singapore. That could be several thousand data requests for hotel reservations, for food reservations, for your hotel. How does your artificial intelligence agent on top of its probabilistic model do that? It's going to be quite literally several thousand microtransactions. There is not a system right now that can do that at that scalability.

You know, we talked about, you know, Visa does 23 233 billion transactions a year. Uh it does that um with about 80% operating margin, which is incredibly impressive. Um on our network, we can do that the entire annual Visa volume in 21 hours and 7 minutes for $23,000. So, that's where we need to go because when Cloudflare launches their stablecoin and they're running interference for the entire public internet, they're going to need several millions of transactions a second just to even keep up. So, what what do we need there?

And what is what systems exist on Wall Street um that will not be able to keep up and will they break, right? Will the netting get too large? Uh will will they get too backed up? That's the thing we think about a lot. But I I I want to flip this around.

Right now, Nvidia per market cap is worth more than every publicly traded bank in the G7 countries combined. So, all 4,200 American banks, all three or whatever Canadian banks that exist, every bank in France, every bank in Germany, every bank in Japan, every bank in China, combined is worth less than Nvidia on a market cap basis. So, it's not necessarily what AI do wild for Wall Street. Will AI save Wall Street? That's what I want to think about.

It's a great question, but I think also another really good point that you raised is it's not just about what can the agents do and what are they doing, but what about the underlying rails and infrastructure and the data and the processing of that data that they'll need to operate. But I want to stick still with the risks for a minute. Maybe come to Kendall, of course, with your legal hat on is it obviously for a lot of the, you know, actions that they are taking, could you walk us through some of your views on what some of those unique risks, additionally, not necessarily underlying infrastructure, but what are some of the other unique risks that might be introduced from, you know, compliance and legal perspective, introducing these tasks that are used to be done by humans, now done by agents into the system.

Yeah, that's I think it's a really great question and you know, I I think back to Mickey's presentation from earlier when you looked at sort of agentic AI and the use case within financial institutions. And I think one thing he said as he, you know, started to frame the decision around integrating agentic AI within a financial institution is the importance of trust driving adoption. And I think that what that does or what he's saying at least to me is is that underscores the importance of risk management and governance, right? There has to be a sort of risk management oversight over agentic AI as financial institutions choose to deploy agentic AI throughout their financial institutions. In terms of like a primary risk that that sort of comes to mind for me, I think you hit the nail on the head, right?

You said that agentic AI, you know, it it can do reasoning, planning, and acting all sort of on its own. And I think that what that means to me from a legal perspective is the existing risk that we sort of associate with gen AI or large language models. Like we know what those risks are and I think that agentic AI is just going to supercharge those risks because of the autonomous nature of each agent, right? And so, I think from, you know, when we think about this sort of at a granular level, you know, a lot of my clients, right? They're subject to examination, right?

And so, the first thing that a regulator is going to ask a lot a lot of times when they see an issue is, "Well, tell us how how did this issue occur, right? What is the what is the sort of audit trail?" They're going to look at independent audit. They're going to talk to folks within the institution to try to figure out how we let, you know, what actions led to this particular issue, right? And so, when you think about agentic AI, if you have audit trail deterioration, right?

Where you have agents making decisions autonomously but not keeping a sufficient audit trail that led to those specific decisions or actions, I think that that can lead to challenges in terms of explainability, right? Managing the hallucinations and managing a lot of the risks that we've talked about or that you mentioned earlier. Um so, I think that's the primary risk that comes to mind for me is just the lack of a comprehensive audit trail, which then allows, you know, financial institutions to um, effectively govern as well as provide oversight and then also in turn, you know, just limiting the ability of financial institutions to issue spot and then communicate with the regulators.

I mean, I think that's absolutely right. And and obviously we've heard a lot from various regulators and public sector officials today about importance of balancing, obviously, you know, managing risk and innovation. But critically, regulators also need to know who to go to and what to do if something goes wrong. And that's why that audit trail is very important. But I think, you know, the other piece of this is it obviously also all comes down to trust.

And you know, Todd you mentioned in the beginning, you know, this is the future of commerce. And so when we're thinking about that, the future of payments, maintaining that trust is critical, especially for consumers and consumer data. And so what does it take to maintain that trust and integrity as we move to an agentic environment?

It's all about trust. I mean, I said the future of commerce intentionally because I look at the way that payments interact with the real economy. It's when consumers go and buy something from a merchant. Consumption's a big driver of GDP that in this economy and globally, if we make it easy for a seller and a buyer to trust each other in cyberspace, even if they don't know each other, they're going to hopefully transact and that's going to create some economic growth. Now, the trust issues that we have in the agentic world are quite monumental.

And so I'm going to go a little bit off book here. Since you already talked about Visa's network and the and the power that we can deliver within the financial ecosystem, let's talk about the merchant side of the story. If you have a website, you've spent the last 15 or more years fortifying that thing against bots. If some agent comes into your e-commerce environment with bot-like behavior, you're training your system to block it because you don't want to have a data breach. Guess what?

Agentic AI means that your first best customer might be an agent working on behalf of a person that used to come into your website, browse around, comparison shop, put something in a shopping cart, think about it for a few weeks, come back, and click to click to pay. That may all happen with much more velocity. That's why Visa just yesterday announced this announced the creation of the trusted agent protocol. The idea there is to create some kind of a protocol layer so that we can embed this trust in the in the merchant ecosystem. Because it's just so important that when you're making a payment, it's not just that the money moves from point A to point B.

It's this entire system of operating rules and governance that attach to that payment so you understand if I see that Visa brand, I know what my rights are. I know that if there's fraud, I have some recourse here. It's also for the merchant something that they don't need to know who I am to trust the Visa brand, and they're going to go ahead and let that transaction go. And so, one of the big opportunities that we have if we can solve this trust issue for agentic commerce, we can drive not just a lot of friction out of the current system. Hopefully, we can do it where we're actually adding security.

Let me just tell you one more thing about Visa that, you know, is is maybe something that a lot of people in this room weren't even born in 1993, but that's the first year that we enabled AI in the security workflow for Visa. And the thing that solves 30 billion or more in fraud per year for our financial institution clients is that we can interrogate this data through an extremely robust AI uh program. So, this is not something that we're just dabbling with. This is core to us, and it's all about really solving that trust issue for the ecosystem.

And I think that, you know, that identification piece that you mentioned is really interesting in the protocol level, which obviously congratulations on the announcement, hugely exciting. And I think as you say, in a system that has been designed to prevent bot activity, we now have to actually allow for trusted bot activity. So, we only have like about 5 minutes left. So, I want to go and do some rapid-fire question round one for each of you. My first one actually is going to be coming back to Sarah and still building on that theme of like agent identity.

And you know, building on what Todd said, how do we think about agent identity in this new future and so that we know which agents to trust? Are there no new infrastructure needed? New standards? Your quick answer, please.

Now, this is the million-dollar question, and I'm preaching to the choir based on the conversation we've already had. I do think one of the most important and underappreciated questions in this next wave is identity. As a field, we've spent decades sort of refining our frameworks for know your customer, know your business. As agents start to take action on a user's behalf, we're going to need the equivalent of know your agent, right? If an agent is going to manage money, sign a contract, execute a transaction, everybody involved needs to know who authorized this, what rules was that bot operating under, and what data did they use to make that decision.

If we don't have visibility into that information, we lack the verification trail that you're describing and auditability. And and that's the linchpin of trust. Um, I'll just add one thing which is I do think some of the technical building blocks are starting to reveal themselves. So, we are seeing startups working on decentralized identifiers, uh, provenance tracking, machine identity verification uh, building blocks of the agentic identity stack. Uh, but it this is going to require more than technology.

It will be private public collaboration to align around standards uh, on many of these issues.

I completely agree and I think we've heard a lot about public private collaboration today and the importance of it. So, Kendall, my last question for you. So, you know, thinking about that, the standard setting, the regulation and accountability, what more should we be doing to think about from a regulatory and legal perspective how to hold those high standards and think about how do we help regulators think about evaluating those decisions?

I think that's a really great question and you know, I think you hit the nail on the head in terms of public private partnership. I think a lot of times when we think about public private partnerships we're looking for formal formal forums in in which to engage, right? So, Treasury, other banking regulators have talked about public private partnerships, but I also think it starts, you know, with the the sort of relationship with examiners. And so, we have, you know, a lot of my clients, you know, are subject to regular supervision and examination and I think part of, you know, just getting regulators up to speed is just making sure there's a open conversation, transparency in terms of how agentic AI or AI is being used within the financial institution and starting the conversation at the ground level, um, which I think in turn will will help sort of regulators, you know, learn on on the fly, which I think is happens on on a daily basis in other forums and other contexts.

Totally agree. Todd, to continue on that theme, you obviously talked about what you are doing. What would be your advice what for other financial institutions looking to prepare for the gen tech commerce world, what should they be doing?

A, Kendall's right. You have to be in conversation with your regulator. But this is a global phenomenon and we are in a fragmented geopolitical environment. So I think one of the big tragedies that we need to preempt is that just on the cusp of a revolution that completely changes commerce and inclusion for the better, we end up in a second best outcome because the systems are so fragmented. So in addition to talking to your own regulator, I would pay close attention to what's happening outside of your home market because when this infrastructure is built out and the trust is there, every small business could be a multinational corporation.

Every consumer could be using their agent to go out in the world, not just their own local marketplace to find the best possible value for their money. And every government could potentially be benefiting from that incremental economic growth. So we can't let fragmentation stop this progress.

Completely agree. And I don't want to go over too much time, but last very quick question then for Robert to end on. What's the remaining role then for humans in this evolving world? You've already said you think agentic commerce can save Wall Street. Where does that leave us?

Yeah, so uh again, I'm not going to overrun the chair of the SEC. Um uh it's big picture, it's if we actually let agents do all this data work, uh we can get rid of the ad based economy. And when we get rid of the ad based economy, we can start spending time with our kids again. I have three kids. Phones are terrible.

They just are. Instagram who is is a great company, Google's a great company, um but the ad based economy that's been existing for the last thir- 27 years as an accident of history because we refused the payment companies allowed refused to allow uh open money on the open internet uh has led to an ad-based economy. If we let agents start doing this work for us, uh we can start doing a lot more human things again.

I love that. So, thank you so much.