Moving AI From Pilot to Production With Broadridge

[00:00:00] It's not your fault that your agentic AI systems are acting outside of compliance. There are just too many data sets to guardrail them all. But with Denodo, you can now organize your hundreds of data sources into one layer and govern your agents with a single approach. Try it now with Denodo by visiting denodo.com to learn more.

[00:00:28] What separates an AI experiment from a production system that can be trusted with regulated financial work? Well, my guest today is Tom Carey, President of Broadridge's Global Technology and Operations Business. Let's just say my guest Tom was ahead of his time. He was studying AI in the 1990s when finding a commercial use for it was considerably harder than finding another reason to use a mainframe.

[00:00:58] But today, he's helping financial institutions connect AI with the platform's data and operational workflows that keep markets moving. So we will discuss why firms should simplify a process before automating it, how human review fits in production workflows, and why smaller agents might be easier to control than one agent that is responsible for everything.

[00:01:26] And Tom will also share lessons from Broadridge's work on everything from email traffic, token costs, and agent governance. So what does it take to move AI from impressive demo into dependable work? This is just a few of the things we're going to explore today. So enough from me. Let me introduce you to Tom right now. So thank you for joining me on the show today. Can you tell everyone listening a little about who you are and what you do?

[00:01:56] Well, first of all, Neil, thanks very much for the invite. Great to be here. I'm Tom Carey. I'm head of our product and technology of Broadridge. Probably should introduce Broadridge a little bit as well. So Broadridge is the leading fintech. We are the infrastructure around governance, capital markets, and wealth. So for any of you working with us, our clients and the market, we are that back end of the infrastructure.

[00:02:19] We do sort of things at scale, and we try and help the industry modernize, electronify, and be as efficient as possible in what we do. So my role, I actually have two caps. I lead product and technology, and I also run one of our divisions as well. So I look after our capital markets and wealth management groups.

[00:02:39] So I sort of have a good sort of what I call sort of intersection between what's really happening in the markets and what's happening in sort of the technology evolution of what we do as well. Interesting fact for you, I'm obviously over 50. I have actually a degree in AI. So back in the 1990s, in a bizarre scenario, I decided I'd like to do AI as a degree. It was absolutely no use coming out of university to find a job in AI.

[00:03:07] I almost got a job at a very large automobile company, but they shut down an AI division because they couldn't find a use case for it. So it's been parked for like years and years and years, and it's been really interesting to see the sort of the evolution of compute power plus as well sort of the large language models and the new stuff coming out now that suddenly like reborn AI.

[00:03:27] Why? Because, you know, I'm sure most of our readers know, you know, the coining of AI as a phrase goes back to 1956 and the Dartmouth College Conference, which McCarthy sort of coined the phrase. And it's lived in the wilderness for 40, 50 years as unusable technology. But now the hardware side of things, the compute power plus the new models is like creating a fascinating sort of world.

[00:03:49] So I look forward to it. And we sit, you know, in this world of innovation now where Broadridge is, you know, powering forward innovation for clients really through a number of aspects. Platform at number one in our viewpoint of creating a ubiquitous platform for our clients to work against. AI number two. And then there's a big topic in our market around tokenization and digital assets, which we are leading in as well. So there you go. A quick little tour of me and Broadridge.

[00:04:15] What an incredible story. I've got to ask, did you get any cool dad points from your kids there? Because having a degree in AI before they probably thought it's just a thing that appeared three years ago. That's got to give you some kudos, right? You know, you obviously don't know my kids. No. I wish. But they were like, yeah, whatever. And they like, yeah, but you I mean, they did joke like, yes, but you used to use a big mainframe. We are my little iPhone now is better than your mainframe.

[00:04:44] So they're kind of condescending on that. But I do explain to the problem. It's a very interesting degree just that I've got a BA. I haven't got a BSc because it's actually in psychology, philosophy, linguistics and computing. Because you were learning about how machines allegedly think at that point in time. Yeah. So there's a very famous book, Computers and Thought, I think it's called, which was sort of the Bible back then as well. But yeah, no, I'm afraid I wish I got credit notes. I don't. I just get a request from my bank account details.

[00:05:14] Incredibly cool. I've been to 15 tech conferences this year from Egypt to Vegas. And the theme across every single one of them is agentic AI and agents in just about every industry. And we will have many people listening, attending these same conferences, but curious about how's it actually going to work in their world. And one of the reasons I was excited to get you on here today is Broadridge says agentic AI is live across post-trade and wealth workflow.

[00:05:43] So I'd love to bring this to life a little today. What does production ready mean when the software touches regulated financial operations and millions of transactions? I'd love to try and bring this to life today. Yeah, let's try and do that. I mean, it's definitely beyond prototype because obviously you talk to a lot of firms and they're sort of embedding either prototypes or basic models.

[00:06:06] Or they've, you know, not, again, this is table stakes, but they've embedded a chat agent or some kind of accelerator into their product. But we've really been looking at sort of the workflows of natural services and when you can actually apply agents and when you shouldn't. And we've been lucky in this because, or lucky or strategically sensible in this. We have, we're a technology company at heart.

[00:06:32] So, you know, the major part of our business line is, you know, SaaS-based technology that's driving, you know, multiple clients every day of the week. So we see a lot of flows and we see a lot of patterns. And obviously, you know, AI loves patterns and flows. But we don't take client data and use it at all. But we see the flows, we see the patterns. We're fortunate we also have an operations group. So that's people who actually touch keyboards and actually do operational work.

[00:07:01] And we do that for about 50 clients. So we can see the operational flows in capital markets and wealth and actually fingers on keyboards of what people do. So we can marry the technology with our operations and sit back and go, what are these real flows? And how can we actually look at those and automate those? And what is the benefit net for ourselves and for our clients as well? So we've been looking at that, you know, and we'll talk about it a little bit later, like email automation, for example.

[00:07:29] It's a classic use case that many firms have jumped on, effectively, yeah. But then where do you take it from there into the next step of taking that output from an email and automating the next level of response? And how do you trigger that effectively? So we've got agents embedded in our operational workflows already. And we've been able to test those ourselves. So that's why we're very confident to say that we've actually got, you know, production-grade workflows worked out. We're increasing the number we've got as well.

[00:07:57] So we keep actually building out the portfolio and looking at those. And we've been looking at productivity on scale. Maybe I step back and sort of give you the view of how we think about sort of productivity within Broadridge. Maybe that helps a little bit in terms of setting context. I've sort of heard it as four levels. One, you have this personal productivity. So you've got, you know, your favorite age, your favorite tool, be it, you know, OpenAI or Claude or some other model where you're, you know, personally going in each day and saying like, oh, I need help on researching this.

[00:08:27] And it comes back and gives you some great insights. More for you if you take that and just publish it because obviously it makes mistakes and it's got its own due points. But that's level one. And that's sort of basic productivity. Level two is where we start to see patterns in personal productivity that then we can actually start to actually replicate to other people. You say, this is actually a great tool. Start using it.

[00:08:50] And we have a great innovation function in our firm that says, here is now a tool or service or plugin, if you want to call it that, or skill that everyone can use. And again, that sort of starts to power and generate sort of productivity across the enterprise. You then get into production flows. And you've got my first use cases where you've got production flows where the human stays in the loop. Okay. And we're a big believer in the human's going to stay in the loop. We're not going to have like complete autonomous machines here.

[00:09:18] But you've got operational flows where the agent does something for the human in the operations. It does the work, brings the results back, and then the human verifies or checks effectively. That's my third level of sort of automation. And that's where we've kind of focused our agent policy at the moment in that regard. Then the final level is where you've got agents and you can just let them loose. And you can actually, that's a bad word in let them loose, but in a controlled way, put them into your production environment in a controlled framework.

[00:09:49] And the human is not in the loop because they're running every day. And just like you would run as code, they're running as code effectively. But within a platform framework that gives you all the controls and services you need to make sure that what they do is accurate, doesn't drift effectively, and is in place. So that's kind of what we've been doing. And I say we've been working with our clients on these use cases as well and then push them into our own BPO and operations group to prove they actually work.

[00:10:17] And for many people listening, they'll have operational processes that still depend on inboxes, documents, and manual handoffs. So the question I've got to ask on their behalf is how do you convert all that work into an intelligent workflow without removing the controls that those handoffs are carrying? Yeah, that's a great question. And obviously, it's people's concern a little bit around this whole evolution of control. And obviously, most of our clients are regulated.

[00:10:46] We have regulated pieces ourselves. So you want to be able to demonstrate from an audit control viewpoint that whatever you're deploying is accurate and correct. And I think I'll talk about it a little bit later maybe. Obviously, the concern on intelligent AI is there's drift, there's different responses, et cetera. But I think one of the things that I see is that AI is a catalyst for change.

[00:11:13] And what I mean by that, it's actually prompting people to really look at what they do. And in my experience to date, if we take a process that we do, take email, for example, as your first step there. When you truly step back as an organization and look at what you're doing, you start to actually realize that what you're doing isn't the most logical thing. And I'm pleased to all firms globally, I think. So the very first step, actually, and I'll call it to lean the process out.

[00:11:41] I don't mean going down a whole two-year lean project, but to actually step back and say, why in the world do I get 15,000 emails a day? And you will find, I think most firms will find, that actually there's a lot of waste or wonder in the term in what they do. So in our use case, we eliminated 30% of our email traffic into one of our groups simply by looking at what it was and realizing it was duplicates. It was non-response items.

[00:12:11] It was things that really we should have not been doing effectively. And so you get an immediate process benefit from that. And in my raw mass, and this is my mass rather than broadages, when you step back and you really commit yourself to look at AI-fying a process, you'll get about half your benefit, your true benefit from actually leaning out what you do today. You'll step back and realize, like, we have this four-wise process in this step, but then we do it again another step later.

[00:12:41] So you can actually get a really simplified flow from looking at it and actually looking at really what you're doing. The next step is there's a lot of automation that is classical AI, or is classical automation. It's not the final layer of intelligent AI. And you'll get maybe 30% from that effectively of your benefit. And then the last 20% is sort of the large language model,

[00:13:08] these sort of non-deterministic models in what you do. And that's where you get, you know, experiment with the really cool tech, et cetera. Yeah. But it's that my theory on this is actually that cool tech is forcing us to look back at what we do and forcing us to really rethink what we do and realize that, yeah, there's a lot more I can actually automate now with technology. There's a lot more I can lean out.

[00:13:34] And then you can apply, you know, your smart stuff at the very end for the really clever stuff. And then you can really dig into it and say, what do I really want? Our industry is mostly made out of rules. It's not made out of like, you know, human mind, deep thought around things. So it does lend itself to like more classical sort of agents that are deterministic rather than probabilistic in my viewpoint. So that's kind of what we're learning as we go.

[00:14:03] And again, no one can claim they have the answer because we're, what, 18 months into this journey. So it's kind of like everyone will learn a lot more and there'll be different models that come out that change our viewpoint. But that's what I'm seeing in the market. And I think for many people listening, they'll be scrolling down their news feed or doom scrolling and seeing so many horror stories around token spend and ROI from AI projects, etc. So I'd love to try and restore the balance in the universe here because I was reading before you joined me today that at Broadridge,

[00:14:32] you've had immediate operating cost reductions from your managed model. So where did those benefits appear first and how do you test the savings of not being shifted, of not shifted risk elsewhere? Yeah. And I don't think this is unusual in a new technology. Yeah. You know, I think I have a parallel to this in sort of the, you know, the cloud movement.

[00:14:57] People have deployed workloads in the cloud and just shifted workloads from a distributed infrastructure or a mainframe, shift into the cloud. And then we're shocked three months later that their then, you know, cloud bill was X multiples higher than their traditional run. And then what's happened? And the real reality is they hadn't architected for a cloud inbound. So they were pushing, you know, they're pushing terabytes of data back and forth between on-prem and a cloud.

[00:15:27] That's not how you architect a cloud. So I think there's a lot of step back into, like, how do you actually then do agentic AI and AI services in a way that's actually cost-effective? And as you said, you don't get the, I would say, the, you know, the operational saving, and then you contribute it towards your, you know, token bill at the back end. We think people have been defined as well as what is happening there.

[00:15:52] So first off, we have confidence because we have the operational group and we test it in the operational group in a lot of cases. So we can really go in there and say, you take this model or you take this new product line and you go and test it and we can actually see what happens with that model and see what happens. Baselining is always important. And it's been true for every business opportunity since I've been in business. You need to baseline your costs. You need to baseline what you do and have an accurate view of that effectively.

[00:16:23] If you don't have that, then yes, you'll wake up in 12 months time and you'll be facing like, hmm, it doesn't actually provide economic return on what we do. So we can accurately forecast that in our models. And I say more models to model basic, like in our sort of business-based templates, we know what to put into those models and see how they come out.

[00:16:45] And look, for our token usage, particularly for ourselves, we've got really maniacal on looking at token usage as well. Because it's so easy to suddenly have people putting the whole workflow through an agent and finding out the bill for that is higher than their salary. And you've had horror stories. I think you've heard them on developers running out of tokens on day two of the month.

[00:17:15] And then be like, okay, what do we do now? The answer is you can't send all your source code to this and expect it not to come at a huge cost. So we're very deliberate in that in terms of measuring our token usage. We're not perfect, by the way. So we are dealing with the same issue, but we're actively going through that and saying, okay, one of my groups last month, and this is technology, you could see that token spend was escalating.

[00:17:44] The reason, I'll give you a quick example, the default configuration in that group was to use the latest model from this vendor. This latest model is more expensive. It might be better, but it's more expensive. So we had to realize, okay, we've got to change the configuration in that group so that it stays fixed to the current model and we agree when to upgrade them or change them.

[00:18:06] And it's nothing, no fault with the developers or the technicians in that group, simply that our configuration allowed the latest model to be adopted as default. So there's a lot of learnings come from this, but I say, I don't have it any different to the cloud environment or other innovations we've done in the past. Our mainframe to distribute it is the same, by the way, of just have to get religion on managing your costs, looking at the patterns, understanding it. And once you're there, you'll manage it.

[00:18:34] And I say, we can talk about other things as well, because obviously there's going to be the OpenAI models and differences there too, but I think it'll evolve the cost base as well going forward. Yeah, great point. I think so much has changed and so much has stayed the same. And we will have people listening in financial institutions that could be mulling over the differences between a managed service and a platform it operates itself when they're exploring agentic workflows.

[00:18:59] Any tips or advice or things to look out for when making a decision like that? No, very good question. I think it's moved away from, it's not about control. A lot of the conversations are about, I need to control this. I don't think that's the sort of the language we're facing. It's also true that every company will end up with an agentic framework. Yeah.

[00:19:24] So I don't think there's any thought of any firm that I will outsource my whole agentic control framework to a third party. Everyone will have their own. So the interoperability between these agentic frameworks is going to be really key. Yeah. So, you know, you've heard about NCP services, but then there was A2A, ADK, et cetera. There's loads of things coming out. And our mission, we've got our platform. So we have a, let me step back a little bit.

[00:19:51] We have a core platforms in Broadridge that is our data model, that is our APIs, that is our UX layout, that is our control framework. We made the decision to embed our AI framework into that platform. And that's really important for us because suddenly AI services become part of our core infrastructure and not a separate utility on the side, effectively. So it means that anyone coming in, they can come in via APIs or they could come in via MTP services.

[00:20:20] The control of that and how it works is all in the same environment. And we'll get the benefits and scale from that. So, number one, I think everyone will have an agentic framework. So there's no point competing or please come to my agentic framework. But there is going to be how do our agentic frameworks collaborate and work. You then get into sort of what are your benefits? What are you trying to do?

[00:20:42] And a lot of cases, if you're doing work that every other firm in your sector is going to do, then that is kind of a bad news case in my book. And this is what I talk to our clients about because why do you want to apply a lot of intellect, a lot of IT knowledge, a lot of time to something that your competitor or your peer is going to do themselves on the same basis?

[00:21:07] So we sort of encourage the use case where you have something that's going to be at scale and industry-based, you should probably outsource that or use a managed service. Where you create alpha revenue or where you think it's truly differentiated or where you think you will do a ton of different changes and you're unsure of the model, that is where you probably should build yourself effectively.

[00:21:32] And that's kind of the proposition that I generally talk to our clients about of use the third parties and broadest as a great example of that and a trusted partner in doing that. And then think about where you want to really differentiate what you do and deploy your valuable IT resources in doing that. We thrive on being trusted. I'd say that's where our value prop is for broader issues. We do things in a very trusted way. We've been pushing on a transformation agenda as well.

[00:22:01] But people come to us because they know that we're accurate and reliable in what we do. And that's who you're looking for in a third party. The final point I make is none of this is cheap. While there's the 10x developer cost, the 10x sort of efficiency cost of like I can be 10x more efficient, the reality is it still costs a lot to innovate. We touched on token costs earlier. Infrastructure costs have been going up repeatedly.

[00:22:27] You know, that's been a really interesting phenomenon in this market that the cost throughout compute is actually getting more expensive effectively when you add it all up in terms of all the service costs behind it. So as you think about it, there's a whole ROI effect on this and you've got to balance choosing to develop and support for the long term versus trusting a partner to do it for you.

[00:22:49] And I'm curious from your experience and the journey that you guys have been on, what is the first workflow that typically gives a financial institution the best chance of proving value while also preserving human exception handling, accountability and regulatory evidence? Because I think that low hanging fruit, if they could get something off the ground quite quickly and prove value, then the buy in and adoption and everything will follow. But which is the best workflow to begin with now? Well, it's interesting.

[00:23:19] I mean, you need to, my vision on this stuff is you need to pick something that's relatively high volume because you need an updater. I mean, you're really boring and so you need to pick something that's highly rules-based, which kind of goes against the theory of like, I'm going to use intelligent AI to do all my stuff and think for them. I don't believe in that. I believe you actually want to experiment with something that's really tightly controlled and repeatable. Yeah.

[00:23:43] You need to learn and you need to actually experience how your agentic model will work and how your framework will work. And there's nothing better than having a deterministic agent in my book, eventually. And I'll talk about this maybe coming up later as well about how smaller agents are better than big ubiquitous agents in the world, I think. Do you choose something that's not got a complex decision model in terms of how it works?

[00:24:09] You look at something where the human can verify what the outcome is or you can write more code to verify what the outcome is. So do you know what the input is and do you know from the deterministic flow what the output is? If you know that, you can build a test harness around it and actually verify your results.

[00:24:28] If it's so complex that the outputs are like so varied that you can't actually understand all the potential outputs and the humans can have to like really deeply think about outputs, really bad use case at the moment. Because you just don't, then you can't actually prove to yourself how it works. So that's how we think about it.

[00:24:49] That's where we focused on like high volume, deterministic or rules-based traffic, known outputs, put it to work, demonstrate it works. Then you can start to deep dive into the more nuanced cases where you think there's, you know, really opportunities for thinking models and other things. So, you know, post-trade, great resolution, validation of exceptions, account maintenance, top level customer inquiry.

[00:25:15] And I mean that's the top level in terms of, you know, oh, we get, I keep a very simple use case here. You know, we get a thousand password resets a day. So, okay, well, some of that is, has to be good for a human because the person's actually caused a cyber exception, not kind of cyber exception, but you know what I mean, a security exception. They need to be re-verified. But the vast majority of it is the person just really can't get through the current manual or current automated process for resetting. So we're dealing with that.

[00:25:44] So those kind of ones are great because you can actually cut your teeth on those and then really prove your framework works. But what a lot of people are having to do at the moment is prove their agentic framework works. Okay. So it's less about those, that it's about, do I have a kill switch? Do I have an ability to actually monitor? Do I have audit and control? Do I have risk exceptions? Do I have policies? And this whole sort of industry around sort of your framework is actually probably the harder piece coming out than actually doing some of these agents in their own right.

[00:26:15] And the last thing I'd say is I am finding somewhat bizarrely that some people are putting a higher standard to AI than to humans, which I think is interesting. So I've had lots of questions, not internally, but with some of my dialogues in the industry around how do I know the agent is 100% right? And my response back is, how did you know the human was 100% right? So there's sort of that interesting dialogue going on around that.

[00:26:44] And again, I agree with the hypothesis of you need to control the agent, you need to be audited and see it. But I think there's a different standard been set to a certain degree in some places around that. Again, I'm not suggesting we should let agents loose, but it's an interesting dichotomy of human proof versus agent proof.

[00:27:03] And when moving from not just one agent, but having several agents all collectively contributing to one operational process, how do you preserve that clear chain of responsibility for decisions, exceptions and customer outcomes? I suspect this is another question you get a lot too. Yeah. And it's back to the genetic framework again. Yeah. I'm afraid it really is having your framework in place. I touched on it earlier.

[00:27:29] We've got a big belief in creating relatively small agents. So we don't believe in the sort of someone taking a big workflow and codifying it into a big agent, if you like, that does multiple steps. We think that's very, very difficult to administer. It probably doesn't have a long shelf life. And your audibility of it gets very, very difficult, particularly if you are using LLM in the middle of it to make determinations, et cetera.

[00:27:56] So we're much more into building smaller agents that we train together in their operations. And if you think about it, I'm going to go for a bizarre scenario here with you. If you trained an agent to make your coffee in the morning, okay, don't write an agent that's going to do every step in the process and go through boiling the kettle, filling the cup with a bit of coffee, pouring the coffee in, adding milk, et cetera.

[00:28:25] Build each step with an agent. And the reason why I say that is because you then can have control over what happens. So if suddenly you're, suddenly, let's keep it to coffee, for example, because it's bizarre, is your coffee comes out lukewarm rather than hot. You know it's boiling the kettle that's probably the root cause. Yep. Versus, I don't know what happened now. Somehow it's come out that way. And then you have to start introspecting with the agent of what happens. So we like that.

[00:28:53] Then, I know we're going to bizarre coffee examples now, but I've heard this one. Maybe when you get a bit more advanced, you can put a large language model in to learn what coffee type you like in the mornings. So on a weekday, you want a really strong double espresso, effectively, yeah, because you want to be wired for your start of your day. At the weekend, you want the more mellow thing. It can probably learn that over time and actually have the intelligence in it. But then there could be drift in that.

[00:29:23] Or you may change your personal preferences as well. So the ability to sort of codify the agents and have them at a sort of a granular level where you can actually see what they're doing and inspect them, I think, is the big thing that we're looking at ourselves. And then, of course, you've got controls behind all that as well in your ability.

[00:29:41] And that's going to be the huge thing for this marketplace is that when we do have, which is undoubtedly going to happen in the global industry, not just financial services, issues with agents. And you've seen the ugly face dialogues around what happened there. You're going to be able to inspect a lot better than you were in the past. And you'll be able to shut things down or change as well. So that's our theory on how you get agent flow and how you build sort of smaller, more component-based agents.

[00:30:11] Brilliant. And we started our conversation today talking about when you got your degree in the 90s in AI. And we're now coming full circle. So I'm going to ask you, I'm going to pull out my virtual crystal ball and ask you to look forward five years now. And I realize when I say that out loud just how crazy that is because the amount of technological change we've seen in the last three to five years is insane. But if you did look into the future, where do you see AI in financial services?

[00:30:40] Well, it's very difficult, as you know, because if you'd asked about agentic AI in, let's call it, September last year. Yes. I'd have given you a very different answer. Yeah. You know, the way the model's evolved since November, I'll call it. And we all know what happened in November in terms of certain third-party release. It's certainly changed the mindset of what goes on. But let's jump forward. I think a few things on my mind.

[00:31:08] One is that, you know, the classical use cases of chatbots, search routines, report agents, they'll be ubiquitous in every application going. Okay. Because it's just going to be embedded in and you'll just, it'll be much more like your experience you're used to on your iPhone in terms of, yep, it's standard. Why wouldn't I expect that? So the differentiation that was created maybe two years ago is people launching a load of services around sort of that kind of service will disappear entirely.

[00:31:37] There was a conversation around the future of UX as well. So the user interface experience, and also it's a lot of firms advertising a headless experience now in terms of converting their services to headless. And I see that as really something that will evolve because one of the things that AI has become very, very good at is actually taking UX from prototype to production really quickly. We've seen that ourselves.

[00:32:04] So the UX used to be the mammoth job of a firm to like, if they wanted to re-skin it or re-change it, was always a big, big exercise. We're seeing that really been a power enabler in sort of AI. So we think there'll be much more choice in the UX experience that people would choose, or there'll be agent-to-agent conversations rather than UX-to-UX conversations as well. So I think that you'll see evolve a lot in that space. The big question probably for me is where regulations would go.

[00:32:33] So, you know, you might theorize there's a lot more happening on AI behind the scenes with the big players than we know. But, you know, the models we're using, you know, they have other models and they have more advanced models. So you've got a real question around government sort of regulations, be that in the US, be that in Europe, et cetera, and what that'll mean for constraint on what we can and can't do.

[00:32:58] And it's probably going to take some issue to actually really push that agenda as well. So you may see over the next five years of sort of like a acceleration in some markets and a step back in some markets too, as they realize that we have to have more constraints on what's happening. And that will come properly, I think, if there's an issue, you know, if there's suddenly a, you know, a financial issue with the use of AI in what it does. I think you just have to see. The market's always gone that way.

[00:33:27] And I'm not talking about the financial market, but all markets have gone that way in terms of it takes some stimuli for people to look at it and go like this, this needs more regulation or this needs more control and what happens effectively. I think the last piece we'll see in my view is the models themselves. There'll be a democratization of the models. So models, the models won't be the competitive advantage. They've been sort of like in an arms race of late, if you'd like, in terms of who's best at what.

[00:33:54] When you actually look at the stats, they're getting very, very close in their behaviors. It's going to be about how you actually use the model in your business and what you actually have as specialist data in your knowledge fabric, if you'd like, and how you've actually exploited that data for the best advantage for your clients and for the industry as a whole.

[00:34:16] So we're really thinking about putting AI to work within the enterprise, sort of connecting it through our platforms, protecting data at all costs. And that's got a real key thing for how we think about that and delivering that responsibly. And I think that responsible delivery, the evolution of models, regulations, and a much more move potentially towards agent-to-agent conversations between firms is kind of the way it will shift. Five years sounds like a long time. It's obviously not a long time in the markets in reality.

[00:34:45] But we've seen a rapid acceleration in technology, you said, over 18 months, let alone three years. Yeah, and I think that is a thought-provoking moment to end on. And I cannot thank you enough for joining me today, especially for walking people listening in financial institutions, how they can move from AI beyond experimentation into meaningful enterprise-scale deployment. So many big takeaways.

[00:35:12] But for anyone wanting to learn more about you, about Broadridge, continue the conversation you started today. Where should they go? Well, great question. Thank you for that as well. And by the way, thank you for hosting as well. I think broadridge.com has our corporate website and gives you a lot of information around that. It provides a good source for that. We are available to talk as well. So we're having a lot of dialogues with our clients and the market around that. So please reach out to us.

[00:35:39] You can reach out to me directly or to our links as well. Our LinkedIn site has a lot of information too. And we're here to help power the industry. Our role here, as we see it, is to help firms get more efficient, more resilient in what they do, and to share the knowledge about what we're doing. And as I said before, we do see a world of agentic frameworks sat in every firm. And it's how we collaborate on what the best models are, the best approaches are. So we're here to help and we're here to serve. Awesome. And we have covered a lot today.

[00:36:09] And I think the way that you've brought the topic of moving from AI pilots to AI at work making a real difference and what agentic AI looks like in production, some of those big savings that you had on your journey as well will really bring this to life for people listening, resonate with a lot of people. So I will add links to everything that you mentioned there. I encourage people listening to check that out. But again, thank you for bringing all this to life in a language everyone can understand. Appreciate you. Great. Thank you very much. Pleasure.

[00:36:37] I think Tom's four levels of AI productivity offer leaders a useful way to assess where they are today. Shared tools can spread these gains and controlled agents can begin handling repeatable workflow steps while people can verify the outcome. Full automation, all that can come later once the controls, monitoring and evidence can finally support it. And another lesson I think is that the cleverest model is rarely the whole answer.

[00:37:07] Because firms need to understand the process, establish a cost baseline, protect their data and decide where building something themselves can create a real advantage. And sometimes that first win comes from removing duplicate work before an agent even touches it. You can learn more about Broadridge at Broadridge.com. Find Tom on LinkedIn. Big thank you to him for joining me.

[00:37:34] But where would you place your organization on those four levels of AI productivity? Let me know at techtalksnetwork.com. You'll also find links to everything that my guests left there. Let's continue this conversation. But I'm afraid we're out of time now. Time for me to go. I'll be back again real soon with another guest. Thanks for listening as always. Bye for now.