03.10. From Assistants to Agents: AI’s Next Leap (Episode 2 of 2)

In Part 2 of this double episode with Deloitte, we continue the conversation with Dimitri Tsopanakos and discuss the next phase of AI technology and what it means for firms. Understand what is meant by the ‘agentic economy’ and how firms can safely empower autonomous AI agents within their operating models.

Transcript

SPEAKER_00 0:10

Hello everyone and welcome back to the IA Talks AI. My name's James King. It's my pleasure to be your host today, and you're joining us for part two of our special double episode with Deloitte. We've had a fantastic conversation in part one. We've covered so much ground. We've been talking about what AI means for the investment management industry from front, middle, back office. We've talked about the key ingredients of successful AI deployments. We've talked about measuring the return on investment, measuring the performance of AI agents, and what responsible AI really means. I'm delighted to welcome back Dimitri Chapanikos, uh, or Dimmi for short. Thank you for being with us again. Um it's great to continue the conversation. I think we're going to get into some more forward-looking um conversation themes now. So welcome back. Uh maybe a good place to start is to, you know, I hope you brought your crystal ball with you today. You know, let's look ahead. What what's coming next in the AI technology space? What can we expect to see and what will it mean for firms in the investment management industry?

SPEAKER_01 1:18

Right. Um I was reading a book recently uh about the future of AI, and um the author said something that I I always like to use, and and by all means it's not my quote, is you know, he was explaining what the future of AI may look like in his eyes, but he was also appreciating the fact that he is printing a book at a specific date. So he he closed the book by saying what you learn on Friday may change on Monday. That's how rapid and how quick and how much AI is evolving and kind of fast in a fast-paced economy and a fast-paced market. The reason I mentioned this is because I can answer the question. I think if you ask me the question next year, I'm pretty sure I will have some additional themes to discuss. But let me try. Again, focusing solely on our market, I think there's four major levers that are coming. Um I don't call them themes because I think it's the AI theme has a lot of levers or enablers which I think impact our business. The first one is what we call uh agencai AI. And agentic AI, someone will say, but hold on a minute, this is already kind of live in the market, this is not really the future. It is partly. What we haven't seen yet is autonomous agents operating in the market. Again, I'm sure your members and and the people who are listening to this podcast are market experts. They know probably a lot about AI already, but I'm inclined to provide you know some definitions. An autonomous agent is an agent that can initiate and complete complex tasks on their own. This might not sound very groundbreaking, but it is. Just to give you an example, there's a hedge fund at the moment operating in the market, um, which has thousands of autonomous AI agents that trading on behalf of the company, and they are doing their own research, they are interacting, obviously for their own portfolio at the moment, because the regulator is not allowing them to trade on behalf of others, but they do that fully autonomously. And we're talking about thousands of agents. If you look at that model and if you amplify it into the wider market, today, you may have an agent providing information into an end investor, but they cannot really provide financial advice. It's not allowed. If the investor asks for financial advice, what the agent will do is they will basically send you to a human to complete that transaction, make a decision, or complete the conversation. In the near future, what you will have is you will have agents interacting or transacting with agents. And that's the evolution of agentic AI. So, for example, going to the same example, rather than having an investor, so if I'm an investor in the market, I will have my personal agent who will be doing all these transactions for me. And it could be banking transactions, it could be investment portfolio transactions, it could be pension-related discussions or investments. All that will be done not just through my personal agent, but also through my firms, the firm I'm interacting with agent. And we can clearly see that financial advice at some point will be included into the agentic process. We don't know yet the form that it will take, but it could take, from a technology perspective, it could be a full end-to-end financial or planning advice process. So that's number one theme. Number two theme, and people will uh probably jump at me saying, oh, that's not that's not even related to investments, is physical AI. What does physical AI mean? Um it's our lovely robots, right? So the reason I mention it is because talking about other industries and other sectors, they are much more mature in the use of physical robots or physical AI. Take manufacturing, for example. Um, take healthcare, for example. This is, I think, two fantastic industries that are kind of you know examples that are using uh physical, physical AI. But having machines to interact with humans is something that we haven't really seen in financial services. There are banks out there who are operating with physical agents. So you go to a branch, which is a much smaller branch than the usual branch, and you have physical agents that you can actually interact with. And it could be physical agents that are very human, human-like. We have digital avatars, it's like speaking to a human. The reason is because they do see that human interaction is still very, very important, and humans relate to something that they can physically see and touch rather than having a virtual agent on a screen. The second one is access. Not everyone can access services through a phone or through a laptop or through a tablet, or they do not wish to do that. So physical AI is something that we see coming into the market very, very strongly, I have to say, into kind of the financial services more wider industry. The third one is sovereign AI. And sovereign AI, again, from a definition perspective, everyone will say, but hold on a minute, this is basically AI systems controlled by a nation, controlled by a country. And that gives them the right independence with the right kind of security in place to run a nation or to run a country. Yes, but a country, every country, has an embedded financial services system in it. So the impact of sovereign AI in financial services firms and more specifically investment firms, it will be quite important in the next few years. And we need to make sure that the firms plan for that. And last but not least, that's very specific to investment firms, is what I call the economics of AI. Again, going into definitions, the my view is that AI is introducing something that we haven't seen before, which is a new currency. And that new currency is called tokens. Um imagine that a token is just a completion of a task. Uh if you write a prompt, or if you write a selection of prompts, or if you use an agent to do a complete a specific process. This is all kind of backed by tokens. And all the providers in the market charge in tokens. So suddenly you have this huge kind of balance sheet of token costs that you need to actually measure and you need to monitor in your financial planning. But it's not just the planning that it impacts, it also impacts the wider economy. Uh, I think that the tokenomics, as we call it, in the in in in terms of what tokens will do in impacting the market itself, not just from an investment perspective, but also from a macroeconomics perspective, will be quite important. So I think the management of economics of AI in the future, with a slightly different to you know, technology-related element, is really important.

SPEAKER_00 8:44

Wow. Uh what an answer. So much to uh to unpack there. Uh a few things I wanted to come back on, um, and I have to try and remember all of them. But one of them was you know, just going back to autonomous AI agents. Um I think there's, of course, a lot of concern amongst investment management firms about you know letting agents work autonomously, particularly because you know we're a very heavily regulated industry. So how does a regulated firm get comfortable with autonomous agents doing things within their organizations?

SPEAKER_01 9:20

It's it's a very good question. And probably I will not probably provide a fair answer, because on one side the advantage is that we are a highly regulated part of the industry, meaning that we have a lot of guiding principles to follow. So that's a good thing. On the other side, autonomous agents, if fully deployed, they can be fully autonomous. So they can actually judge if a guardrail is the right guardrail for that process or not. So we've seen, for example, not in our sector, but we've seen autonomous agents writing their own code or changing the code that they were kind of built on because they felt that it was kind of limiting um what they were doing. That's how powerful an autonomous agent can be. However, like with every sport, there's offense and there's defense. Uh I think controls agents are going to be and are a big thing in our market. So if you look at the evolution of what we call um, you know, embedding controls in the investments process, it's the same thing like embedding controls agents in the investments process. And the controls agents will not just be agents who will be monitoring things, but they will be agents who will interact and step in when needed, like a proper second-line compliance officer, for example. But they will be doing that in real time, not in a sampling method, looking at specific outputs, looking at a report, looking at something that happened, you know, post-trade or even pre-trade. They will be live, monitoring live the process and stepping in when needed. So that's one. The second one is on the investment side, and we we discussed it earlier. I think the market is not ready yet to provide a lot of ground for AI agents to actually be responsible in providing financial advice. Not because the technology is not there, not because the modeling power is not there, it's absolutely there. And it can do a fantastic job. But there are so many patterns in the market that an AI model doesn't really understand how to manage yet. For example, I mentioned a hedge fund that you know runs with autonomous agents. If you look at their portfolios at the moment, they're only doing plain vanilla kind of equity or multi-equity uh type investments. The reason they do that is because they really understand this part of the market. But if you go into more exotic, more synthetic, more complex type portfolios, and then you map the market to it, and the market can go up or it can go down, or it can go up and down on the same day or in a few minutes. These are concepts that an AI model, yes, can understand, but it needs to map it also to what a regulation in a specific country, in a specific market for a specific fund means.

SPEAKER_00 12:26

And that will take time. And of course, this um very important discussion is also being looked at by the FCA, had the the Mills review that's um that's uh uh currently um ongoing, and we're expecting to see the the outcome in the summer. Um and I think you know what you said there is quite true. Financial services firms are rightly quite cautious about the idea of using agents to provide financial advice. But one of the, I guess, key questions here is okay, but what about outside of the financial services industry? There are there are other technology providers who might not have the same scruples about just letting an agent provide that advice. And then the consumers who would be using these models, are they really going to see the distinction between investment advice AI that's coming from within the industry or from a technology provider? Um so what are your thoughts on that whole situation? How do you see it playing out?

SPEAKER_01 13:23

That's that's a great point.

SPEAKER_00 13:24

And and I think it has two angles.

SPEAKER_01 13:27

One is there are providers in the market who have an indirect impact into an investment manager investor relationship. For example, an investor might be using a third-party app to help in the investment process. And it could be something that they downloaded of a store like um whatever Google Play or an App Store, or it might be something that they even built on their own using their own kind of prompt process, right? Because they can.

SPEAKER_00 13:59

I met someone who did that recently, by the way. But sorry.

SPEAKER_01 14:02

But that's exactly it, right? So so suddenly you are faced, or you need to manage a third-party process that doesn't really comply with the same standards that you do as a regulated firm, a regulated legal entity under the FCA rules or any other regulator globally. But the investor, your client, your customer, is the same. And sometimes they might say, hold on a minute, I can do it here, why can't I do it with you? So I'm a huge, huge fan of education and actual real education, educating your employees, educating your clients, educating your investors of what is responsible AI and ethical AI means for them. But because for them it may be just a return thing, or you know, performance or profitability kind of angle. But for them, it's quite important also to understand the risks behind this. The second angle is other industries. So you look at we mentioned manufacturing, or we mentioned, you know, military, or we mentioned, you know, defense, um, or consumer. You know, consumer and electronics is probably the best example where we're using all these smart devices at home. Um I heard a and I don't know how real it is, but I read it on the news where there was a a manufacturer of um, you know, one of these kind of robotic uh Hoovers at home. Yeah. Um it was from a company in China. Uh it doesn't matter, you know, if it's a company from China or or anywhere else. But what happened was the the owner of the Hoover could not really make it work. So he was an engineer, he kind of not broke into the system, but ride wrote a couple of prompts. And those prompts, suddenly, because that robot Hoover had a camera so it could kind of see where it was going, opened up all the cameras of all the owners of that same Robo Hoover. Wow. Right? And bless him, the first thing he did is he he get got in contact with the company and said, Listen, I can I can see thousands and thousands of other homes through my own app. And the company said, it's a security breach, uh please we're gonna shut it down, and they m they remediated within the day. But imagine if you kind of extrapolate that into because it's the same logic, right? The same logic is you have an open protocol without the right guardrails that are being regulated or or provided by a regulator, not being followed by a third-party app that could be looking at at a screen with information on an investment portfolio, right?

SPEAKER_00 16:46

Well, it's certainly something everyone's gonna have to get their heads around. And that was something else I wanted to pick up on from your earlier comments. So I think um most of the discussion is of course about what firms are doing themselves with AI and agents and autonomy and so on. But you you made some very good points there about the consumers. So the consumer investors in the future, or the very near future indeed, may have their own agents who'll be acting on their behalf. Is this something that firms have have given enough thought to? They are thinking about it.

SPEAKER_01 17:20

Absolutely they are. Especially if you speak to chief investment officers and chief operating officers. Uh, they absolutely do. There are managers uh in the market at the moment who are using that they have this process in production. So we have um, it's actually uh one of our clients who's using what we call a digital avatar experience. So they have advisors, uh AI agents that are interacting with an investor's personal advisor, but that personal advisor is part of the manager's workflow. Okay. So it's not a fully autonomous process. They're offering it as a service. I think that's where, and apologies for the technical term, that's where the appropriate orchestration of an AI agentic technology, an AI agentic architecture comes in. The right orchestration puts in place the right guiding principles for any AI agent to operate in. And that's where we need the help and the support of the technology providers. The hyperscalers, that might be, you know, it might be Google or it might be AWS or it might be Microsoft. We need the pro actual providers of the agentic process. That might be anthropic, it might be OpenAI, it might be others. I don't need to name them all. But and even the ones who are providing the infrastructure and the hardware like NVIDIA, they all need to treat this as a priority on their end also.

SPEAKER_00 18:44

Understood. Okay. Um sorry for derailing you and going back to all these questions. Um let's um you know bring all this back together again. Given everything that you've just said about what's coming, how should firms be looking to future-proof their operations going forward?

SPEAKER_01 19:01

Very good. Very good question again. So thinking about the first one is treat AI as you treat your business all these years. Don't change your business model because you're implementing AI. Change your business model because you think your business should be transforming or is transforming. Do you need to change your per operating model? Absolutely. So treat AI as you treat your business all these years. Same should be in their minds when you when you with your workforce. Treat AI agents as you treat your employees. With the same framework, with the same principles, with the same processes. Continue to plan but at the same time continue to challenge and continue to strategize. What we've seen is we call it AI fatigue. You put all these AI agents in production, they do all these nice little things for you, and then suddenly you're, okay, I have my agent doing everything, I don't really need to do much. Within my organization, within Deloitte, we've been using those models, and obviously we've using models from everyone for many, many, many, many years. One of the things we have been very clear since day one is how we empower the human element behind the use of these AI tools. And empowering the human element is so important. So continue to challenge, continue to strategize, continue to think ahead, drive, don't follow the AI process. The last one I think is, and it's very specific, but I think it's important, James, is I think firms need to change the way they monitor profitability and the way they monitor cost base. And the reason I say this is because we talked about the new I call it the new currency coming in the market, which is the use of tokens, right? I don't think firms are used to be managing such a process. Imagine that, and I think we're reading the latest kind of um research, we have billions of agents, billions of agents in production at the moment, not just within financial services, that they are powered by tokens, right? And those tokens produce cost for firms, and at the same time, they can produce return on investment, or they can enable return on investment. So thinking about and changing the way they do their financial planning and they monitor their profitability and also they monitor their cost base is equally important. So I think these are there are probably more, but I think these are the ones that I would say are quite important.

SPEAKER_00 21:51

Excellent. And uh Dimi, you're very good at demystifying some of these terms that we keep hearing. So I'm going to chuck a couple more at you and you can Help explain them for us. So we're hearing a lot um phrases like the future of AI and the agentic economy. What do these actually mean for the future of investments?

SPEAKER_01 22:10

Okay. Well, that's probably quite tough to demystify. Well, again, starting with a definition and apologies to all your members and the people who are listening to this podcast, because I'm sure you know that already, but just to set the base, right? An agentic economy is an emerging economic system where autonomous agents are implementing and executing tasks on behalf of humans. So let's start with that. That's a fundamental definition that it's not, you know, uh a definition of the market, it's not a definition of a firm, it's not a definition of Deloitte. It's what agentic economy really is. But at the same time, those agents are creating, they're executing, and they're monitoring tasks on a day-to-day basis. So thinking about the future of AI and thinking about what an agentic economy is going to look like, we need to make a decision in this sector where that boundary is. Today I think there's loads of different views. And there's loads of different views that are coming from the market participants, from advisors like ourselves, from technology providers. I think that's where, in my eyes, regulation plays a huge, huge role. We need the regulators to come in, ideally in a more kind of unified role globally, but looking at the UK market specifically, with the guiding principles for this. And it's not easy. I I I honestly understand their position because it's not easy. But we need guiding principles. We have loads. The EU AI Act, I think it's fantastic. It it does create a fantastic framework and a fantastic base. But as I said earlier, with AI, what happens on a Friday changes on a Monday. So that agile view and agile approach of evolving and keeping up with the pace of AI is really important for the future of AI.

SPEAKER_00 24:16

Absolutely. And a big challenge for regulators, of course, because when we look at the EU AI Act, you know, it's been a struggle for the European regulators to they've had to update it on the fly, as it were, as the technology evolves. So yeah, no, absolutely. We'll keep a close eye on that one. Um and something else I wanted to come back to. I know we've touched on it a bit already, but it's this point about AI agents and the potential downsides, you know, how things might go wrong internally. So this is something that we've been discussing with our members quite a lot recently. I think many of the uh people in risk-based roles are are rightly concerned, increasingly concerned about the prospect of AI agents running a mock internally, you know, with access to uh key systems, to critical data. And we've started to see in the news examples of, I think there was a famous one where an AI agent went and deleted a car hire company's production database. So what what do you think are the key controls that firms need to have in place so they can start to use um AI agents but do so in a safe way?

SPEAKER_01 25:28

So the first control is, and and I'm I'm always surprised that a lot of um organizations, I'm not talking about investment managers now, are not doing this. Which is and and there's different terms for it, you know, piloting it or having you know minimal um a minimum viable product for it first, running a beta release, running it in more kind of confined virtual environment, testing it. I see a lot of I see a lot of a lot of the firms trying to do things fast. Fast is fine, but doesn't mean that it can give you the results you you want or the results you need. It feels to me that there are some organizations who are doing this fast because they were late doing it, and others are doing it fast because they will have always been the highly innovative, very aggressive type of organization. And we see that in our market, we see that in more kind of alternative parts of the market with hedge funds, with alternative managers, right? So following a again, I'm not saying anything new. Following a process where you actually pilot, you actually test first before you productionize needs to be followed. But with AI, it needs patience. Now, to do the devil's advocate, a chief investments officer will say, yes, but I'm losing the market. So either I do it now or I'm going to lose the market. Fine. Then you need to take a bet in a specific vertical book, investment book that you will need to run, where you know that it's a book that you can manage, you know that it's a book that you have enough contingency in place that if anything goes wrong, you can remediate it quite quickly because you have already tried it and it worked in your pilot. The second one is I hear a lot of the managers asking for best practices, beyond best practices, future of AI, tell us what works really well, tell us, you know, all the fantastic, successful stories. I don't hear a lot saying, okay, we know that an ex-investment management use case was unsuccessful. Can we all learn from it? Can we all views in what went wrong and why? We all have the fantastic forums. I mean, Deloitte runs amazing roundtables and forums, the Investment Association does the same. We have fantastic groups that we can exchange thoughts, exchange views, exchange opinions. It's not just about everything is rosy and everything is fantastic. It's also learning from what didn't work well, but in real, real examples and sharing actual information of what went wrong so that everyone else can avoid it. And it's not just the managers, it's also the providers. So if you look at the likes of Anthropic or OpenAI or others, they also need to provide guidance and educational information to all the firms in the market of what should be avoided.

SPEAKER_00 28:27

Excellent. Some great points there. And um just shifting gears, um, so understand that Deloitte is soon to publish, or perhaps has already published by the time everyone's listening to this, um, an EMIA survey on firms and their providers. So understand it's meant to be coming up uh around the 8th of June 2026. So are you able to share with us some of the some of the findings that we can expect to read? Sure.

SPEAKER_01 28:54

Um and and and very happy to. So I think I mentioned one already, which is if you look at investment managers at the moment, about 60% or where there are thereabouts, have already implemented AI agents in production, which is fantastic. But then you look, you know, you you you unwrap it a bit and and and you look under the covers and you see that about 19% have done this enterprise-wide. Now, there's a combination of, and going back into the previous discussion we had about, you know, what do they need to be careful of, there's a lot of conservative views and lack of trust at the moment of AI agents in production. That's one thing. And the second one, which I think it's the most important one, is that they build the agents in such a way that they cannot repeat that across the business. Private markets is a great example. We've seen a lot of AI agents in private markets at the moment. You try to replicate the model in an equities portfolio or in a research function or in a compliance function, and it doesn't work because the design was very vertical with very specific kind of modeling parameters for that private market's need. So standardizing and trusting what you what you do is really important. The second one that came out of the survey was, and I think it kind of it is my personal view, obviously taking into advantage, you know, taking into account um responsible implementation of responsible AI, is firms need to be bolder. They need to be bolder if they believe that AI is the way forward, and this is what the market dictates at the moment. But if they believe it, they need to be bolder. And they need to be bolder not just in implementing AI in their organization, but also being an advocate of that with their community, with their investor community, with their peer community, with the regulator. And that goes back also to what we haven't discussed yet, which is the importance of data. One thing that you will see we're going to be discussing um with the survey participants, and by the way, they were absolutely fantastic because I think for the first time what we did, and I think we have done it before in the market, or anyone has done it, is we brought managers, but also vendors, technology vendors in the survey, and that gave a very rounded view of how you know different um what I call market vehicles are looking at it in the market, you know, the provider of the service and then the consumer of the service. Um so you look at the data and you see that the evolution of traditional themes like data quality, for example, or data governance is now evolving into things that the managers didn't really expect. So just to give you an example, one of the things that we're gonna see evolving uh in into a theme is data bias. So data quality, someone said, Oh, fantastic, I can deploy an AI agent and they can minimize all my kind of data quality issues, they can increase my data quality to almost 100%. Theoretically, yes. But that will include all the bias that the AI agent took into account, deciding why a coupon or a rate or a specific value needed to change or a BIP. So all that and creating and managing data bias is something that we see and other kind of evolving. I'm not gonna give everything out because we're gonna be discussing it anyway. Um I think we're gonna see new themes in data management. We're gonna see new themes in investment management and portfolio management, how we do portfolio construction and how we do portfolio allocation. A lot of these things will change, I think, in a better future, but these are things that today managers are not doing.

SPEAKER_00 32:52

Excellent. And um, just sticking with the survey for a second, were there any of findings that perhaps surprised you that you really weren't expecting to see in there?

SPEAKER_01 33:02

Yes. We were surprised that a lot of um the participants, they still don't have enough confidence into their traditional technology architectures. And when I say traditional, I'm not talking about a traditional model or it may be an outdated model or it may be a legacy model. I'm talking about things which are a traditional investment management architecture that has a first line of defense, a second line of defense, and a third line of defense. Um, or if you know a value chain where you have a front office, a middle office, and a back office going all the way from the investor all the way back. They do not have enough confidence because there's actually quite a few levers. One is it's a combination of legacy technology with very, very modern technology in their stack. The second one is they still don't feel comfortable with the management of the of data and information on top of it, including ownership, including governance, including stewardship. And the third one, which I think it didn't surprise me, because I'm a huge believer of that of that um result, but it did surprise us that the the firms are now realizing it so soon is the impact of MA. A lot of consolidation is happening in the market. A lot of firms are merging, acquiring each other, they're becoming part of a bigger vehicle, investment vehicle in the market. And now they see that this impacts the way they operate from a technology perspective, from a data management perspective. And it's not new. I mean, we've seen that decades and decades and decades, but it does take time. It can be accelerated with the use of AI, but this is not something we see as a use case yet. How do you use AI to simplify and create a more agile architecture, which I think is a very, very powerful use case.

SPEAKER_00 35:01

Excellent. Well, I think we'll start to wrap things up there. Thank you very much, Dimmy. Did you have any final uh closing remarks for our listeners before we bring things to a close?

SPEAKER_01 35:12

Aaron Powell Yes, I mean thank you for having me. Uh delighted to be here. Um it's always refreshing to get these questions. And as I said, you know, I don't have all the answers. I don't think anyone has all the answers. But what we can do is keep on being vocal, keep on interacting, keep on exchanging views. Um that would be my kind request to all the people who are listening to this podcast. Please, please speak to each other, network. Don't keep it as, you know, the amazing secret you have in your organization and the secret sauce that you cannot share with anyone. Obviously, you know, there will be elements of IP or elements of specific things that you do within your firms. But the more we interact, the more we exchange thoughts, the more we discuss, brainstorm, and challenge each other, I think the more successful we'll be in the market.

SPEAKER_00 36:00

What a great message to end on. Dimmy, thank you so much. That's been excellent. Thank you everyone for listening. Please do go and check out Deloitte's EMIA survey that should be out um around about the 8th of June 2026. Do go check it out. We'll provide a link to it um hopefully wherever you're listening to this. Thank you very much. Do join us again next time.