990: Insights: What’s happening with agentic AI in financial services right now?

0:00Get business done with the new American Express Graphite Business Cash Unlimited card. With unlimited 2 % cash back on all eligible purchases, unlimited 5 % cash back on flights and prepaid hotels booked through American Express Travel Online, and a flexible spending capacity that can grow with your business, you'll have the confidence to keep building. Apply today and earn a welcome offer of$1 ,500 cash back after you spend$50 ,000 in qualifying purchases on your new card within the first six months of card membership. Terms apply. Learn more at go.amex.com. This episode is brought to you by Google Chrome.

0:33You think you know a browser, but Gemini and Chrome? That's new. It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50-page restoration block, or finally break down that long article you've had open for weeks. Gemini and Chrome is here for it. Ready to make anything online make sense? There's no place like Chrome. Check responses set up required, compatibility and availability varies 18+.

1:10AI has already delivered impressive gains, from streamlining back-office operations to powering conversational assistants. But agentic AI is a whole new frontier. These systems don't just respond, they act. They make decisions, they take initiative, and operate on behalf of users with a level of autonomy that could fundamentally reshape how banks and financial institutions serve their customers. But what's still missing? What barriers are holding us back? And most importantly, what needs to happen for Agentic AI to deliver real value for both customers and institutions? I'm David Barton-Grimley, and on today's FinTech Insider Insights, we're partnering with Gradient Labs, who recently worked with a major UK bank to implement Agentic AI in a way that transformed, not just automated, critical customer interactions.

1:59We'll also look at how innovators like Klarna are setting the pace and explore the major challenges around ownership, regulation, and trust. So is Agentec AI the future of customer experience or just the latest tech buzzword? Let's find out. I'm joined by a fantastic panel of experts who know the space inside and out. Let's dive in. So firstly, we have someone who is very familiar with the case study from Gradient Labs, their CEO, Dimitri Massin, CEO of Gradient Labs. Welcome Dimitri. Can you tell us a little bit more about your role at Gradient Labs? Cool. I'm really glad to be here. Thank you.

2:36Yes, I'm Dimitri. I'm the CEO of Gradient Labs, where we work with financial institutions to really automate any types of manual repetitive tasks that there might be in customer operations in particular. So we have built a collection of AI agents that do that work and in a fully autonomous and safe fashion. So really we start kind of from customer support experiences, which is kind of the easier tasks, and now expanding more and more into back office and specialist support operations. Yeah, that's us. Awesome. Great intro and welcome on, Dimitri. And joining Dimitri, we have Tim Elder, Chief Product Officer at Multiply.

3:15Welcome to the show, Tim. Can you tell us a little bit more about Multiply and the role in the AI space? Sure. Thanks for having me, David. Very happy to be here. I'm Tim, Chief Product Officer at Multiply. And at Multiply, we are making financial advice accessible to everyone. And when I say financial advice, I talk about that in terms of, it's a form of financial intelligence that is regulated. And so financial advice looks like being able to tell someone specifically about what they should do with their money, be it opening a product or using a product, financial product in a particular way.

3:51And historically, that's been done by human advisors. so highly qualified human advisors in this regulated space. Because it's been driven by humans, it means that the people who consume financial advice tend to be, well, basically it's not mass consumer. There's a cost associated with that advice. So our technology is really helping make financial advice accessible to everyone, make it a mass consumer service, and we're doing that in two ways. One is helping people that give advice today, humans that give advice today become hyper-efficient and the second way we're doing that is provisioning our technology for scaled automated personalized financial advice so that means you can use our systems to give advice end-to-end advice on a fully automated basis which is which is quite exciting so our tech looks like um deterministic advice engine and risk engines like working in parallel some modeling tools.

4:47But where we are today is that's been provisioned in a way for agentic use. So we're super excited about agentic use, super excited about LLMs because, you know, one of the things that human advisors have been really great at is talking human to human, using language and translating that into a financial plan. And that's where we are today. We're in a really exciting place where people are beginning to use LLMs to kind of talk about their finances in their own terms and use our systems to get a financial part. Awesome. Welcome on board. And completing our lineup, we have Aditi Subarau, Strategic Partnerships Lead at Instabase.

5:23Welcome back to the show, Aditi. Tell us a little bit more about what you've been up to since we last spoke. Great to be back, David. Thank you for having me. It has been a busy time indeed. So with all of the recent developments in the AI space, as always, the almost acceleration of agentic development and adoption, we have been kept on our toes quite so much. But interestingly, most of what we're working on is what you could almost term as backstage work in the sense of not really building the end applications or the final agents, but more so in the data layer where we're helping our clients develop the infrastructure and the foundations to develop, adopt, and also distribute these agents in a truly enterprise format across those organizations.

6:16So not as glossy, but definitely very critical work that has been keeping us awake. That's actually very interesting. A great intro and also a wonderful segue added to the first part of the podcast, where actually really one of the first questions that I'm going to pose for all of you is where are we with agentic AI? What kind of is that pulse check that status. And maybe just to recap what each of you said in your intros and also what I said at the beginning, there is absolutely a distinction between the broader technology of generative AI and its ability to compile and manipulate and assemble knowledge and all of that wonderful stuff, which I think we all know, but also agentic specifically, which is what we're talking about here, which is the ability for a degree of autonomy for action away from humans.

7:04So Dimitri, I'm going to come to you first. Where is it today? What is the status? I suppose that's quite a broad question of Argentica today. Yeah, I think there's certainly a lot of noise and hype around that area, I'm sure all of us are aware. So it's kind of hard to distinguish really, like is it a hype, is it a bubble, is it internet 2.0 type of thing? I don't believe so. And I think what I'm seeing from the trenches in particular is that we have crossed with a genetic AI, we have crossed a very important milestone. And that milestone is that we have actually already seen use cases which are proving to be very, very valuable in enterprises and in companies.

7:44And so kind of seeing inside out is just a matter of time now. So it's not the question of if this will ever be useful or transformational. We've seen those cases already. It's just a matter of how quickly companies will adopt it. And I think that would be the summary of what I'm seeing right now. That's really interesting. So it's a little bit like what Aditya was saying, is that people are beginning to sort of prepare those back-end systems, the sort of data operations and everything to prepare for that kind of agentic future. As well, there's a lot of that happening for sure. It's just every company operates at different paces, right?

8:22So I've seen companies that have implemented that already, but very, very few, but very successfully. So that shows us that it's possible and there's a lot of value. But most companies I'm talking about are much further behind. They're creating first policies and like which models can they use and not and how and like, how does it get us through compliance and whatnot. So there's just like a very big variability in terms of where companies are at in their maturity journey. There's lots of complexity. Tim, what's your view from the world of wealth advice? I mean, I know you guys also released a paper, I think on this recently as well, which I had a look at.

9:01I thought it was really good. But yeah, what's your thoughts on that bridging that gulf from generative to agentic? I think to your point, I think if we look at today, we've seen people looking to find often kind of internal efficiencies with workflows. And I think that's tied to, I think what Dimitri was touching on, which is tied to finding meaningful use cases but i think there's this aspect here dimension of risk so finding uh something where there's a valuable use case that people have the kind of appetite for and i guess the thing which is is is one of the things that's quite interesting is uh beginning to kind of uh look at how actual end users and customers interact with these services uh and like yeah obviously Dimitri's doing some really interesting work in this space.

9:53And I think there's question marks around what are the right use cases for this? So we're beginning to see people dip their toes with customer-facing experiences, which is quite exciting. And in the advice world, specifically, I think that looks like there's some lower risk things you can do. So for example, like non-advice, non-guidance, help people explore goals, for example, like what is it that they're trying to achieve with their finances? You can help them understand their goals. You can help them describe kind of financial products to them, provide information. So that's super cool. And then obviously you can kind of take like key self-serve actions, right?

10:34So if you're talking about like things you want to do, again, like I think some of the stuff we're seeing in markets, especially with Dimitri's work is looking really exciting on that front. I think where it's going in the future is stuff around providing guidance and providing advice. So for the audience listening and for people to understand in terms of levels of risk, if you get to a place of giving fully automated, holistic, financial advice, that's more risky than giving guidance and that's more risky than giving information. So I think that's what we're seeing at the moment. People beginning to dip their toes, taking key actions and looking at, you know, energetic systems through that lens at the moment.

11:11The way I think about it is we're kind of coming out of a period that I would define as agent washing. It's like everything is suddenly going from being AI enabled to being agentic, but it's actually not. a bulk of what we're hearing in the market or as stories or as research reports and news articles and blogs is essentially glorified automation. So it's what AI used to do with a little bit more, perhaps autonomy, but not really truly a little bit more kind of sophistication in the prompts and independence, more reasoning, but it's not really truly agentic. Whereas there are some spaces like Dimitri and Tim mentioned, where it is truly coming to its own and like customer service, etc.

11:54is an area there. And I'm now really excited to see how that evolves. There are banks like RBC, like JP Morgan, like BNY, for example, who are kind of doing truly agentic adoption. But I still think it is limited to spaces like customer engagement, like individual productivity, like perhaps some search functions, which then enable the first two. Yeah. And what would you all say are the barriers? I mean, maybe just to summarize, We've talked about risk barriers as well, right? So there's regulatory risk, you know, Aditi you've mentioned, you know, getting the back office systems working and all of that.

12:29Are there also barriers in some of these models as well? I mean, I know there has been some, well, there is quite a lot of controversy around there about the AGI debate, you know, how much can these, you know, foundational models scale in their reasoning. so some of our listeners may have read maybe a summary of Apple's latest paper, which I think is very interesting, which basically says that, you know, these models will not scale in reasoning. They're not going to be hyper-intelligent. That's never going to happen. It's a big kind of debate. And I sometimes, as just someone watching and listening this, I find it very difficult to understand, to filter through the signals in the noise here.

13:06Like what is actually required from the technology in order for an agent to fly and for it to work? And, you know, Dimitri, maybe it's a question for you. I mean, do you see that the bottles that you're using, for example, are able to do some of those more foundational, you know, agentic use cases, or is it actually not? There's still more work that needs to happen. Yeah, that's actually a great question because kind of every company is grappling with this. The example that I was always making is like, we hear from examples where the new SLM models are performing at PhD level. Yet, if you look at the seemingly very simple use case of customer support, it's still not good enough for that.

13:49So somehow those two things don't align, right? So you have PhD level models, but a simple use case and it still doesn't work. But the reason is not because the models are bad. The reason is because human brains can do a lot more than just the LLM model. So LLM models may be one reasoning element of what the human brain can do, but humans can kind of observe, learn from their mistakes. They kind of have memory and concept and they can save those things or like memorize, I guess, in human terms. And I think those modules of the brain is what's missing to have really good, robust application. So I'm going to make an example from customer support area.

14:27Like if you hire a human into that space, they will come in and they will observe more senior people. They will read conversations. They will go through some onboarding and they will build out a map of the company, of the processes, of the product inside their head. And this capability of doing that is actually not captured by alums. Alums play an important role in reasoning about what to learn and how to learn, what to memorize. But fundamentally, there are lots of those building blocks that are missing. And I think companies that are doing well in that space are essentially building blocks around our lamps.

14:59That's so interesting. I mean, Aditi, are you seeing some use cases over the horizon, over and above customer service? Or is it kind of customer service is the area where it... No, I think there are definitely tons of use cases. It goes back to Dimitri's point. I think a very simplistic way of framing that gap is effectively a current LLMs or Gen AI can't really read the room. So they don't understand the context. They don't understand the environment. They don't understand those nuances that are a combination of environment, previous customer data, or again, like previous organizational and market data and current interaction.

15:42And they can't really bring it together as intuitively as a human being can. That is stopping further rollout. Obviously, in addition to everything we spoke about in terms of data silos and like infrastructure and APIs and connectivity. However, in regards to those use cases, I think there are still significant areas within the financial services space and even just banking, not even wider FS, but even within banking, where there are a lot of use cases that are still relatively instructable. Again, to go back to the example Dimitri gave, if I were to hire a new employee, I could have a graduate or an analyst or an associate come up to speed very quickly on processes and functions like trade reconciliation, like fraud detection, like optimal overnight liquidity management, for example.

16:31There is a process of doing it. There is a logical framework that you can explain to somebody which still requires some tweaking, some adaptation, some on-the-fly decision-making, but it is still a framework that can be explained. And I think those are the use cases that should be next in line for agentic automation following kind of customer engagement as the first step. Yeah, and I imagine, Tim, in your world of advice, it's a similar thing, right? That delta between an LLM that can, I don't know, understand and intend to summarize a conversation in guidance, scaling into advice, I guess, becomes, I suppose to paraphrase what Dimitri and Edithi are both saying, it's kind of engineering, right?

17:17It's all of the stuff that sits around the LLM required to make it work. Yeah, sure. I think there are two immediate words that are being, or two concepts that are being spoken about. Use cases, really important to focus around. And then what you said yourself, David, around intent capture. So we're certainly looking at this in terms of, can we capture the intent of what someone's doing? And then if we have a clear idea of distinct use cases that we're seeing commonly, how do we make sure that we build around those use cases? And then in terms of tool usage, I mean, this is a really interesting point around, you know, financial intelligence, LLMs, origination of financial advice.

17:59So when I say origination, we care where the advice comes from. and at Multiply we're taking a reaction opinion of we want the advice to originate from our systems and then the LLM to use that. So hence the agentic usage of Multiply, so Device Engine, Risk Engine and Modeling Tools. But yeah, certainly in terms of that's kind of the architecture around or that's how we're currently architecting the system where we think about focused use case where as Aditi describes, we can kind of look at making sure we can evaluate performance around the use case rather than approaching it as this kind of catch-all oh, this LLM is going to give you total holistic financial advisor.

18:35That's not the approach that we're taking. Yeah, absolutely. And there's a strong risk aspect to that as well. I mean, Dimitri, I'd like to maybe just deep dive on risk a little bit because you're deploying your service into banks, customer services answering personalized information, all of that kind of stuff. How are you thinking about complying to regulations in an agentic world where it's taking action? What are some of the pitfalls? It's actually fascinating. I mean, the space is still evolving very quickly, but I would say most traditional legacy banks, they're still, from what I'm seeing at least, one to two years out from deploying fully autonomous AI agents into critical customer-facing applications.

19:21right so all banks currently are typically at the stage where they use co-pilots to kind of help their employees to do certain processes to kind of to know how to reply correctly and so on so it's certainly a process that is quickly evolving but the key thing of where I see between companies which progress quicker versus not is there are some companies who treat those new types of systems almost like a human brain right so they're essentially saying okay if we can prove that this human brain or this system can operate at a higher quality higher customer satisfaction fraction of a cost then we can essentially say it's better for the customer it's better outcomes for the regulator it's better for the company and we can prove that that system is better than whatever is that exists right now and we can apply the same standard to it like we apply to human teams.

20:19So that's one camp. And those companies typically, if they can make that argument internally, they can typically prove with the help of companies like ours that, yes, it is possible to perform better than the human teams. And that's one route. And then the companies who are still in the mindset, this is a software and we need to test it how we test software. And so they're trying to test essentially write tests and like, they're still stuck in this traditional software mindset and they can't overcome the fact that it's a non-deterministic system that you will not be able to test it fully because like the way the customer conversation can evolve is like in million, billion different ways and you just can't test every single part of it.

21:04And you kind of need to have this mindset, okay, it's like a human brain and that's why you need to test it also like this rather than a traditional software where you just write a bunch of tasks and they're 100 % sure that it's going to work. And that's what I'm seeing in terms of two types of companies and how they think about it. I almost think that's one of the biggest blockers to large-scale agentic adoption across the industry. And not just FS, it could be any industry, because I think we are still stuck in such incorrect ways of evaluating this technology and even evaluating the outcomes of this technology.

21:41it's like we're still stuck in the old ways of saying am I going to save cost am I going to reduce cycle time it's what cycle time you might have a completely different process from what you had before it's like somebody says well I've decided to sell my Ford Fiesta and buy a Ferrari but I go into the showroom and ask them how many gallons does it take like how many miles do I get per gallon and it is such a bad way of looking at this like we need to completely redefine those frameworks, we need to completely define what the objectives are that we are solving for, and then build the metrics, the monitoring processes and the governance that flows back from those outcomes as opposed to just existing systems on steroids.

22:23So, but that's a, I completely agree with that, Dimitri, huge bugbear for me. But it's so interesting, because it sounds like there are no frameworks. They just don't, like no one knows how to think about this. I mean, is it the case that there needs to be a universally agreed way of, yeah, Dimitri? In a way, we can assess humans, like every company does it, right? So if you have, let's say, people performing operational work, you still want to know that they're doing a good job. And typically what you would have is internal quality assurance functions, where you review the work and you assess it based on specific criteria.

22:59Is it correct? Is it process compliant? and does it kind of capture all the regulatory requirements? So that's the process, right? So we know then, okay, some humans are doing a good job and some are not. And we're essentially saying you need to apply a very similar mindset here because essentially what you're having here is an intelligence and you try to control that intelligence, but ultimately it will also make mistakes similar to how humans do it. Yes, and Tim, you know, your industry is very, very, I don't know, human, right? This is human to human advice. How are you seeing your industry and wealth advisors almost react to this tech coming on down the line?

23:40Is there an embracing of it? Is there a sense that we need to think about this as a productivity lens? Or are you also seeing that disconnect that Dimitri and Aditya are showing? So when our technology is being applied in the context of human advice firms, I think one thing that we've certainly learned over time is to, there's definitely a piece of education to be done. I mean, Dimitri Nditi described kind of how do you benchmark and evaluate performance in this way. And the classic thing we'd hear would be, you know, something from a compliance team along the lines of, oh, show us every possible permutation output.

24:23And obviously, you end up in a kind of a philosophical place of saying, well, actually, I'll be giving you an infinite set of potential results with continuous differences between different cases. But we made sure to be quite opinionated, to turn up with a set of test cases and obviously building out evals, educating the teams involved, ideally cross-functional teams, so making sure there's the correct representation across the business, senior stakeholders as well involved. And yeah, I think there's an education piece on kind of what to expect and get people aligned because I think we've, in the earlier days, we certainly kind of ran into challenges when a compliance team might make a kind of an old traditional request and then we've had to try and kind of meet that request in a way that it didn't really work for anybody.

25:18So certainly turning up with an opinion, educating a team, I think is the way forward. I think you said there's no framework. I think there is actually a framework, but that framework is evolving as a new one that didn't exist before. So as Tim was talking, they are evolved sets and they're all kind of various ways of kind of rolling out quick, slowly, a limited rollout or like doing offline validation. Red teaming is a very popular approach. but all those techniques are kind of evolving and new. And so it's an emerging framework, but the kind of the fundamental principle still applies. You kind of more comparing it to humans rather than to a soft traditional software.

25:59So that's the difference. Awesome. And on that note, we're going to take a quick pause here for a quick break and coming up, we'll dig deeper into agentic AI and in particular, some real life examples from Gradient Labs. It's all coming up after this super quick break. So don't go anywhere. Get business done with the new American Express Graphite Business Cash Unlimited card. With unlimited 2 % cash back on all eligible purchases, unlimited 5 % cash back on flights and prepaid hotels booked through American Express Travel Online, and a flexible spending capacity that can grow with your business, you'll have the confidence to keep building.

26:33Apply today and earn a welcome offer of$1 ,500 cash back after you spend$50 ,000 in qualifying purchases on your new card within the first six months of card membership. Terms apply. Learn more at go.pamex.com.

27:14Welcome back to Fintech Insider Insights, where we've been looking at where we are with agentic AI in the financial services industry. Now we want to take a look at what needs to happen to make agentic AI truly valuable for customers and institutions alike and take a closer look at some real life examples. So Dimitri, we're going to come to you and I'd love for you to tell me a little bit more about Otto and why it's different to AI agents. So tell us a little bit about what is Otto? Yeah. Otto is an AI agent, well, first of all, but it's purposefully built for financial institutions. And so what this means is that we could really tune it to the use cases that would apply to typical banks, lending institutions, savings, and so on.