04.02. Building Trusted AI: Why Investment Operations Need More Than an LLM

How do you deploy AI in one of the world's most regulated industries without sacrificing trust, governance or control? In this episode, Lawrence Baker speaks to Davide Martucci, CEO and Co-Founder of Next Gate Tech, about the realities of implementing AI in investment operations. They explore why clean data and deterministic workflows remain essential, where large language models genuinely add value, and why the future lies in combining agentic AI with robust governance frameworks. The conversation covers data management, workflow automation, regulatory accountability, and how investment firms can move beyond experimentation towards scalable, trusted AI adoption.

Transcript

SPEAKER_01 0:11

Hello and welcome to another episode of the IA Talks AI. And I'm delighted to be here today with Davide Martucci of Nextgate Tech, and we're going to discuss um your eight-year-old company founded in Luxembourg, what you did when you started out, what you do today, and importantly, what you've learned about artificial intelligence and how we can use it and make most out of it in investment management. So perhaps starting at the beginning of your company and going from there, David.

SPEAKER_00 0:43

Of course. Hi, and uh thanks for having me. So the beginning of the company, 2019, um, Luxembourg. I am not a Luxembourgish, I'm actually Swiss, and I'm started, I started my career working for UBS, and I spent quite a lot of time, close to 10 years, working for UBS as an investment manager, so um more on the front office side of uh of the value chain. And I moved to Luxembourg in 2016 to set up the investment division of a single family office. So I was managing money for senior family office, moved to Luxembourg because of you know right place for setting up this kind of companies. And this is when I really discovered the main challenge that every company in the investment fund industry faced on middle and back office and how manual still nowadays and was back in the time, all the different processes. And um what I did is I met my co-founder, who is CTO and co-founder of NextGit Tech, and has a very similar background than mine, so more of quantitative portfolio management, but very, very similar background. And we started to look into why is steel so manual, like why the industry still manages Excel spreadsheets so much, and you know why all those controls are performed by people and there is so little technologies involved. And what we discover is the main reason is that the data is a mess. And when you look at um wherever you sit in the value chain, you are an asset manager, an asset owner, a bank, an asset servicer, you will have a mess of your data because you work with a lot of service providers, you have a lot of delegates, you have a lot of technology providers, everyone sending you data in different format with different data model. So the known ability of ingesting and harmonizing that data created over time this scattered, you know, uh, situation in terms of data that was the main challenge to solve before being able to automate processes. And uh NextGay Tech really came uh uh from from from the the idea of uh we want to solve the data problem first. That's the really mission back in 2019 to create this foundation to then build on top different type of automation and you know answer a different type of pain point for for the industry.

SPEAKER_01 3:30

Excellent. Um yeah, I I think that's a a story that will be familiar to lots of listeners who work in financial services. It's um we we have this patchwork of modular systems, don't we? And um and getting them to all talk to each other is something that still lots of human beings have to do to make judgments about what information goes where. And so um obviously you started out on this journey about how to do it. Uh technology has changed a lot also in in that time. Maybe you can describe a bit for us how how you would start it out trying to tackle this problem and then what's changed now that we've got gen AI and and things have have developed in that way.

SPEAKER_00 4:12

Of course. The um so for us we split the the the challenge of um data, like the big challenge of data, into three main steps. The first is the ingestion. So have a very broad, flexible, robust, scalable ingestion capabilities, which means I can ingest any type of files, any type of data, structured and unstructured, with any type of connectivity. That's kind of the goal. Um, and that's the first step of the challenge. The second step of the challenge is my data have been ingested. I have the data that is structured, not harmonized, but structured in my data warehouse, in my database. I need to identify what I have in my hands, right? So imagine I receive an accounting book of records. So from a fund administrator, I receive my nav pack, I am an asset manager, I receive it in SFTP, in Excel or CSV format, because that's the norm. The norm is this, or even sometimes email with attachment on uh on those emails. So API connectivity or data warehouse or solution that can be connected with API are really the minority today. So once I have this uh XLS that have been ingested, I need to determine for the different sheet after parsing what I have in my hands. What are the different columns, right? The fields, the attribute that I receive. And this is something that you do when you see for the first time a new source, but it's also something that you want to be able to do in a scalable and automated way because we do know that sources change those fields very often. You know, you have a column that appears, or suddenly a column that disappears, or I mean, we all know those kinds of problems. And so we wanted to have a very automated and efficient way of determining what is the data model. And this is the second challenge. And then the third and final challenge is once I have my data, which is structured, I know what is what in terms of attribute, I want to know what is what in terms of records. So I have my data, I have my file, I have my portfolios or my balance of account, I know what are the different columns, what are the different fields, this is a name, this is the maturity date, it is a strike, this is a nice in or whatever. I need to be able to know what are the different entities, the different instruments inside. And what we did in terms of technology is that we use different types of technology along the chain to answer those three challenges and the truth, those three steps. So the first, in terms of ingestion, I think is where we have been able to leverage the most recently the LLM race and particularly for the ingestion of unstructured data. So this, when we started, we were mainly focusing on liquid assets. And uh the reason for that is the data is not harmonized, but it is structured. At least you receive, you know, CSV, you receive text, you receive like files that have a certain structure. It's not a PDF, it's not an Excel S with you know plenty of shit and formula inside and like very, very dirty type of data. So we focus on the liquid, and today we are able to say we focus on whatever. Like send us whatever type of file, and thanks to LLM, we are able to handle unstructured data when before, when we started, only companies that had proprietary NLP were able to do. So that's like the first challenge, and we answer the first challenge that way. And then the second and the third, we have proprietary methodology that we still use and are the same that we started to develop back in 2019, which instead of leveraging LLM on that, we want to be able to be extremely certain about how do we define what is what in terms of attribute and what is what in terms of uh entities. And so we develop our own entity search methodology that use actually embedding methodology. So we put everything we receive from a source. So if I take again the example of this uh accounting book of record file that comes, I have my portfolios, I want to know what is what and what are the entities. And instead of leveraging on a you know referential mapping tables that I have somewhere, I take all the entities and all the fields that I receive for these entities, and then I put that in a mathematical spectrum with the vectorization, the embedding. And this is how I'm able to say this is uh equity, it's an option, it's a fixed income, it's a real estate investment, it's an SPV, because I am able to leverage on all the information that I get and automatically actually determine what's this entity. So we do see an evolution, of course, on the data management piece in terms of using LLM and Gen AI mainly into the ingestions, the first step. But where we really see a big, big impact is what happened after. So once your data is harmonized, once your data is clean and standardized, what LLM and Gen AI give you as a possibility nowadays that was not the case before, is to create very custom workflows on top of that in a very short time frame and being able actually to shorten the time to value on those customized workflows that you can build on top of data that have been processed in a more deterministic way, if it if it makes sense.

SPEAKER_01 9:58

Yeah. Um so a potential example then, using the sort of net asset value example that you gave, that that was something I think that you you started out doing a lot of your work with clients was on this this particular challenge, you know, that that um fund managers need to know, you know, at a moment's notice what the net asset value of their portfolios is. That entails basically trying to aggregate information from all of these patchwork systems to get a value. And there are frequent errors. That was that was one of the things that you first tried to address, you know, starting with the the rule in Luxembourg that that obliged fund managers there to actually have a proactive approach to this rather than just getting a slap on the wrist if they were found in a spot check to have have got it wrong. So you kind of dived in there and and and helped them structure that. Um the next layer on top, you're talking about the kind of now the more agile workflows that you can add once you have that foundation. So uh if we take, for example, a client who's interested in knowing if if they are particularly exposed to a company, you know, if if we call them Astro Y or something. So what what is our exposure to Astro Y, this this huge newly listed company? It's looking very volatile. I'm concerned about its place in our portfolio. And your you your tool can help them interrogate that question and provide an answer to that question without somebody having to go in and and resolve all of these data challenges, you know, to to answer this one query. Is that is that right?

SPEAKER_00 11:33

Yes, it is. Not only this type of question, but we started with in terms of use case and application, the ability of saying I have my data, it's there, it's clean, and I did it with my deterministic flow. So I am I can trust the results, and those are reliable. It's not you know pure LLM because today you could be tempted to say, who cares about the entire process I just described for 10 minutes before, just take a file and you know put it in whatever cloud and ask what is my exposure on that. And uh I think this is why we read now in the press so much that companies and enterprise are you know decreasing their use in terms of AI internally is because they have been unsatisfied about the results because it was not used correctly. And uh the ability for us of answering questions such as what is my exposure on this asset, or do I have an anomaly on the evolution of my NAB, or do I have a problem in fees calculation, or can I run a reconciliation between IBOR, ABOR and you know the custodian book record, or so all those different use cases, is to ensure that your data is in order without the entire process to not leverage uniquely and and and in 100% based on a pure probabilistic approach. So we do continue to run the full data management piece as I described before, and then on top of that, you can have indeed the interaction with an agent that you can create or use template agents that we have in our platform to either investigate, detect anomaly, or you know, have this kind of conversational approach of exploration of your data set. But I do believe it's super important to make understand the industry that it needs to be a mix between deterministic workflow and probabilistic LLM approach, and using LLM where they are strong, and not using LLM for tasks that they are not, you know, normally have been created for those types of tasks. So everything that is numerical, everything that is actually that require this uh scalability and and and trust in terms of the results require this mix between between the two approach and require to put a governance around around the the agent that you will use to interrogate and and do this conversational aspect of of with your data set so that you ensure that you can ground and you know decrease the level of of hallucination you may have if you just go wild on use of agents.

SPEAKER_01 14:41

So uh you'll correct me if m if I'm wrong, but my impression is that you've kind of arrived at this understanding through that exper that sort of years of experience you had preceding the introduction of gener generative AI, and then then what came afterwards. So you you know, in in lots of domains in financial services, we found before Gen AI where we could kind of automate or or tighten up processes, very manual processes, you know, using traditional software, we we did that. Uh, but it reached its limit. There was still there was still this threshold beyond which we couldn't, you know, as you say, unstructured data is something that's traditional software just uh never really was able to navigate. So you kind of learned the the benefits of having a really tight process for the things that can be manipulated deterministically. You saw the the value that that created. Then Gen AI arrives, and now you have a tool that you can put on top of that that can kind of work where you need that little bit of extra agility. The LLM is able to bring that agility and you can kind of you can then build this architecture that kind of leverages, on the one hand, the deterministic areas in all the areas you want it to be really tight and deterministic. But then where you need that agility, the LLM comes in, it provides that, you calibrate potentially the model temperature to make sure that it's not not going too wild, as you say, but you're able to layer it on top rather than coming in and saying, Oh, we've got these Gen AI agents now, they're very agile, they can navigate unstructured data. Let's just ask them the question and they'll come back with the answer. And I think a lot of us have used the tools know that that's uh it's inconsistent at best if you try to get it to do that.

SPEAKER_00 16:27

Yes. Yes, you you're right. Like the um on top, in within, you know, like really having the ability of using LLM, creating agent, interacting with tool skills, and and and being able to govern them in a way that allow us to be extremely flexible in work workflows creation. And I can I we we can give a very, very concrete example of that. Where agents are good and where we would then you know leverage an undeterministic workflow and next gate platform, and this is why today we we position ourselves as we are the ecosystem of agentic automation in investment operation. It's because what we provide is really this environment to enable the use of LLM and agent with the entire uh guardrail and and you know tools that need to be ingested and used and leveraged by agents to perform this in a way that you can trust. If um we look at up where we see that it's very good. So let's take an example, you are an asset manager, let's say small-size asset manager doing private assets. So you do private capital or whatever, like a loan, private loan, or whatever the asset you are doing. The LLM and the agentic approach will be very good into the full data unstructured data extraction. That that's good. Um and even before that, where we see that is very good is in the solution engineering approach. So what it means, it means I come with my set of files. Uh, I take uh, you know, like a whatever, like uh 10, 50, 100 files that are representative of what I have in my universe, right? A bit of accounting book of record, a bit of prospectuses, a bit of uh GP reports, a bit of uh trans transfer agent. I mean, I take all of that. And I give it to a solution engineer agent that looks at the different files and give you sense of it. And say, I see that you have those Excel there, those are the shit inside, this is the type of data, it matches with this, it doesn't match with that. So it gives you kind of sense. Out of that, the first step that you want to do is create a referential master data set so that everything that you do after in terms of struction of unstructured data, it's mapped through the structure and the um org chart, if I may say, that you created in this referential master data set of your entities. And why I'm saying that is to even in the unstructured data extraction, if you take documents that are very long, you know, prospectuses that have 100 pages with very 200, very complex data inside, the ability of giving to the agent already the structure of uh funds, subfund share classes, SPVs, different organizations that have been created by the solution engineering agent reduces the the truncate problem, the problems that you may have on using LLM on analyzing those types of data. So this is where using agent is very valuable in all the exploratory part, the deployment part, solution engineering parts, creating the master referential data, and extracting unstructured data. Once this is there, and then you have this data, you know, we we have capability in terms of data management to provide really classic, you know, PostgreSQL databases where you can extract the data from those Excel differentiat, parse it, put it in a defined data module. Then you can use workflows that are deterministic with trigger that are deterministic, with step and action that are deterministic, and inside ask some agent LLM to perform some actions. Once I have my data that have been ingested, and structured data that have been ingested, there I want to use a mix of deterministic and NLM. And a good way of understanding that is let's say I ingested the data and I want to start to perform controls, like what we were saying before, like nav anomalies. So I ingested my data in this loan, credit, private capital fund. I have weekly nav, so I have start to receive different data points of those nav. And I want to give those data that have been processed, as we say, to an agent to look for anomalies. Now that, if you just do like this, again you will end up to have something that it's not can you trust it or not. I mean, like in terms of reliability, it's difficult to stay in a full probabilistic approach. And what we found extremely helpful is not only to put the deterministic approach around, but also inside the LLM in in itself. And this is what everyone today is like kind of hyped with, which is tools and skills, you know, like we say. And what we saw is using what we coded in our deterministic workers in the last seven years to perform nav anomalies, to perform any type of actual use case. And create tools and skills out of that that can be consumed by agents and run deterministic code when it's needed, because that's what it is, right? If you give a tool to an LLM and you give him saying every time that you perform this particular independent field calculation, instead of inventing with your generic learning base, run that that script, actually, which is a really deterministic script to do, then you leverage on the ability of the agent to reason and on the strength of running a deterministic piece of code inside that knows how management fees needs to be calculated under a certain regulation or under a certain framework. And this is where we have a very, very powerful solution that's you know, kind of take, as you say it, um take out of the best of both worlds.

SPEAKER_01 22:53

Yes. Um I think because I think a lot of people will be familiar with um the great improvement that they've found using Claude skills, for instance. But Claude skills are a kind of system prompt. So they, you know, it's about you know where you need the LLM to retain fundamentally its agile nature, then the skill can help guide its behavior, but it doesn't determine, fully determine its behavior.

SPEAKER_00 23:19

It could still go off piste or yes, the skill, yes. The tool is different. So the the and then I think that there is still a bit of you know um naming convention around the industry that not everyone really uses the same naming. But skills, you're totally right. I mean, if I give us skills to uh an agent, this will kind of harness and and ground a bit the probabilistic approach because it will go and guide into the prompt before going into the generic, you know, fully uh wild, as we were saying, approach, but still it's a probabilistic approach. Using a tools is a bit different, like the tools give so we we use a lot of tools today to create agents that are I give access to my agent the ability to create other agents, or I give to my agent the ability of uh reading uh files that I put in in my data warehouse so that the connectivity, the you know, so those are are tools. I I give the ability for the agents to interact with my data management solution, what we were saying before. But in addition to that, you can give the ability to your agent to calculate a performance phase. And there it can be still getting the agent reasoning approach, context approach in terms of understanding and where the agents are strong to do. And then when he needs to calculate, really properly calculate, then you run a script. So that's really like a piece of code. So it's not you know something that you would prompt in terms of instruction, it's really a piece of code which is purely deterministic that you run and that they actually the agent run. And uh this is, I think, I mean, we are not the only one, right? In in the investment operation and in the financial industry, I think is not so broadly spread yet. But if you take other companies in you know the legal industry, or that this is exactly what how they approach, right? Like it's by saying, let's give those skills and those tools that enable my agents to perform better, because sometimes it goes deterministic, sometimes I harness the pure probabilistic approach by giving skills, and sometimes I just use agent. It happened that we use agent to build deterministic workflow, and actually the deterministic workflow doesn't call agent anymore. Like the agent does the unstructured data, the mapping, the full, you know, like solution engineering part. Because we have tools that we give to the agent to create the workflow, so we have a proprietary workflow solution, it does the workflow, pure deterministic, and then the workflow runs in a deterministic manner, so there is no ask agent until there is an execution that fails, and that called the agent to say, hey, the deterministic didn't work. Maybe something changed in the data, maybe something, and there you recall the agent. And that has plenty of benefit. Of course, as deterministic, I can trust it. And second, I don't consume token every time that I run my workflow. So imagine that I need to do nav oversight not on one fund, but on 10,000 nav a day. And we have clients that is what they do. If you call an agent for every single nav to perform the anomaly, that is something you can do deterministic. You just spend token for nothing. And you know, like that's not uh that's not the approach neither. So yeah, that's that's kind of the positioning today.

SPEAKER_01 26:57

Yeah, I th I think um a lot of a lot of people are sort of learning these approaches now that they are forced to pay the full c cost of tokens. There was this world in which and and in some ways, um I like the example about getting an agent to help you build a deterministic, fully deterministic workflow because the inclination of agents is to they they want to be helpful, they want to, they kind of are con they have the self-confidence of someone who who thinks, yes, I'm gonna help the user with this query. So they their their inclination, I tend to find is is not to suggest, let's let's be sensible, let's hard code this. Like they just say, Yeah, I can I can do that. And like um just give me the the information and I'll I'll have the answer for you in in no time. And um that's that's not always the best way. Indeed, I I've found kind of doing things myself, you have to be um you kind of have to be discriminating always about what um what should be deterministic, where you want the agility, what sources to use. I found that um I had an example recently where it was a a research project and I wanted it to query particular sources, but I hadn't specified that well enough. So it it went away and it it came back with the wrong sources. So I then more tightly specified the sources I wanted to look at, gave it a source hierarchy. And the problem there was that um it found its own previous draft among the sources, like in the in our internal files. And then it was like, it looks like all the information the user wants is here. So it basically just repeated back to me its own work and without you know, this source hierarchy I gave it, um, then it didn't, it didn't feel it needed that anymore because it had found, and so I was like, no, that's actually that's the wrong thing. So I had to delete entirely those files to get it to follow my source hierarchy. But I think that example shows in in highly regulated industries, you know, where um where the regulator needs to know and understand that that the proper processes have been followed, that the data querying has been been conducted in the right way, then you you have to be really focused on what the right architecture is, how to make it work appropriately. And I I think that takes us on to the kind of when you think about uh these questions architecturally, that takes us perhaps towards your your Spark platform. Perhaps you can describe a bit about how that works at at that level.

SPEAKER_00 29:15

Um yeah, I'm completely agree with everything you said just now. The um the approach that we took is the following. It's indeed to say we believe that the real success of implementing generative AI in financial operation will be a mix between everything we did in the last seven years, which does use, you know, like when I was saying before embedding, I mean we do use deep like machine learning and statistical approach in a lot of what we do on the data processes, but again, it stays a deterministic approach that I can explain and retrace and have this uh audit trail for regulator and for clients that we need, and then using LM for where they are strong. And this is, I think, is what the industry didn't got yet in terms of product. Um and this is why today there is a lot of confusion out there in terms of ah, but you know, I use ChatGPT, I use Copilot, and I'm not super the experience that I have, and I'm not super super helpful, and and so they put it a bit on the corner or projects are started. And so what we really wanted to give to the industry is everything that we built in the last seven years in terms of data management, um, in terms of which is the first pillar of our ecosystem, of our architecture. So everything that I described before, the ability of uh not only ingesting, harmonizing, enriching, but as well just you know very basic functionality such as storing those data objects, having those data models, being able to run data quality checks, data lineage, so everything around data management with a real skew on the vertical of investment operations. That's the first pillar. The second pillar is the workflow ability and capabilities. And this we were saying how important it is to run deterministic workflows that are visible and that you can understand exactly what happened and you can orchestrate in a way which you understand. And this is very important. Like imagine you are running reconciliation across multiple sources for huge volume, so like a big, big, big portfolios or a lot of funds, or even if you do it for a complex fund, or even if you have one, you want to ensure that your workflows run in a certain way, that the orchestration run in a certain way. If everything just goes into a big scheduled project that is running into cloud, you may not have that ability of transparency. So we have our own proprietary workflow solution, which is you know like an NA10 kind of, you know, like really like you can put your different action, you can ask agent within the workflow, but you can as well create those workflows from agents, what we were saying before. So you can use the agents to analyze the situation, create a workflow, and then running this workflow that would consume the data management pillars. And then the last pillar of our ecosystem is the interaction part of the business user. Because we saw that business users, so and when I say business user is in our clients, you have tech people, you have product people, and then you have business people. And business people do have a very diverse understanding of how to prompt, of how to you know interact with AI. So we needed to give them the ability to interact with a platform that they they are used to, which means I want to see widget, I want to see, you know, like a canvas where I have charts, where I have tables, where I have um and and this kind of dashboard ability, task management ability, because I receive breaches, I receive, you know, I want to know the trail, I want to see. So this is the third and last pillar. So is data, workflow, and interaction in terms of dashboard, canvas, task management for user. When you give those three pillars to the agentic capability, which is at the core, you end up to have this ecosystem that allows you to build and to answer in reality any type of workflow and any type of pain point in a very short amount of time. And I do I do sometimes demo or presentation where I show that from scratch, so I I take Spark empty, like really literally nothing is there, no agents are created, nothing. I create an agent live, I select our NGT auto as a model, which is actually an NLM that selects the best LLM depending on the task underneath. I give tools to interact with the different pillars that I just described in terms of data workflows and dashboard, and then I showcase how in five minutes I build from scratch a solution that ingest prospectuses, so PDF, and create a full master referential database, you know, stored in a PostgreSQL database with a dashboard visible for a user. And this you do it in 10 minutes, five minutes. Like, you know, when we started in 2019, there were a couple of companies that were trying to do that with NLP. Reading prospectuses is a challenge in the industry that have been, I think, forever there and never really answered by those companies because it was very hard to do in terms of the diversity of those prospectuses, how they are like difficult documents, big with tables, with and now seeing that you can build that, you know, like in five minutes, showcase the example of look with our solution engineer, which I think are key part of the entire process, which are actually human engineers that come and understand your problems, know our solution, and say, okay, this is how we will set up the architecture and the set of tools that we have to answer your need in terms of reconciliation, rebalancing, reporting, data quality checks, or helping you to just you know perform certain type of of checks.

SPEAKER_01 35:49

This is this is this is and what what you then get from that is a system that in which rather than just as you say, you know, we're c comparing with kind of Claude, just ask Claude for the answer and see what it does. And you know, Claude's a very capable tool. But we want we want systems that are more tightly defined. So Spark will help create an agent that then, you know, at the moment that you sort of you've given birth to this new agent, you understand, okay, what systems it has access to and not in what ways, you know, how it can interact with the data, which bits, as you say, have been actually made f fully deterministic with software. And there's there's a sense in which, okay, you can see that not only have we created an effective tool to execute this workflow, but we also know exactly where its guardrails are. Like we know that it can't stray out of the kind of guardrails that we've put down for it, and we can call this our kind of prospectus agent, you know, and then you know, there's not a sense, you know, I think a lot of us would be, and the regulator might be understandably concerned if we said, Oh, we've got this fabulous agent, it's got access to all of our systems, all of our client data, and all the fun data and everything, and we we can just query it however we want and it does whatever we want. I I think they might say, well, okay, what where are the controls? How do you stop it going wrong? And and your your platform helps you basically construct things that are tightly specified and and well governed. Correct.

SPEAKER_00 37:21

Still using the power of cloud. I mean, if you do want to call Fable 5 because you have a lot of money to spend, you can. I mean, if you like the the the API call of the LLM, the underlying LLM, it's your choice. We are not we don't have our own, you know, like LLM. We can use any type from you know like open source or as I said, like entropic, Gemini. We have our own auto model which actually leverage you know a lot of Gemini, a lot of Google actually, um, even in the infrastructure, because we we love to work with them, and this is that's it like at the level of the LLM. Everything else is exactly what you say, is the governance that you put around. And uh, if you want to use the strength of Claude, very good. Like use you know, ICU or Opus or whatever, you can select the different type of LLM. And then what is really strong is that when you create your agent as a prospectus agent, you need indeed to have tools. So to give those tools in terms of reading those documents, and already if you have an agent that would just read documents, you want already to classify the type of documents that enter, and then maybe have an approach in terms of hierarchy, as we were saying before, when you go through a prospectus, and all of that is given by tools and skills, and this guide the the approach of the agent. Then there is the governance in terms of access. And this is another it's it's exactly what you were saying, like in terms of what I give access. I believe that governance in the sense of who is responsible of this agent, because just to be very, very clear in Spark you create collaborative agent, right? So it's not you in your cloud desktop, you interact with your LLM. In in Spark, you are creating an agent that can be accessed by your colleague or by a workflow that calls that agent where the results is available to the entire company. So the management of accesses, which is Vault ultimately with the different API key, with the authorization and the permission, is a big, big topic that as well requires big focus and big attention because you don't want to give, imagine you give to this agent the ability to extract information for emails. And you say, I give access to my email because I receive emails there, and this agent is a collaborative agent, is in Spark. I give access to my emails to extract PDF from a fund administrator. Very good. I do that. The agent has access to my email. But other colleagues may have access to that agent or to that agent workflow and may end up to have access to my emails or my documents. So the management in terms of governance, and this is why we have a full layer on that where you can manage memory, context, vault, accesses, you know, and this is extremely important in terms of how you want to build those agents and how they interact between each other and between the different types of other users that may be around. So, yes, you're you're right. I I think I I see the two levels. Look, one is harnessing and grounding the power of LLM into answering our needs in the investment operation industry with a different set of capabilities that we described before. And the second is pure governance, like how I govern it. Just put, you know, like uh like you what when you create an API key, there is often an expiry date, right? Like it's not the same when you create an agent. You should have like who is responsible, who created, um, what what is the expiry, how do you judge the performance? And there are a lot of different types of approaches for that. We are a technology provider, so we just want to give the ability to you know implement your own approach and be very flexible in terms of how you can govern those those agents.

SPEAKER_01 41:22

And uh there's a sense in which also the your platform makes it easy at once to have the benefits of tight governance, but also the user can still interact with it in a way that enables them to access all the information they want via the dashboard at once. It would be, you know, if we manage a team, it would be frustrating if we were told, oh, you you can meet with your subordinates, but only one at a time. You can't you can't have them all together in one place. You'd say, well, that's just going to greatly slow down my day. I want to have everybody all together and speaking to each other without having a single subordinate that you basically entrust with doing all the different processes. Because that that wouldn't work as well. And I think the the um the governance piece is one that maps well onto the way our regulation actually works as well. I think the um the the example we gave earlier of the the Luxembourg rule that obliged managers to be proactive in their approach to nav, you know, it was a kind of prescription that you then got your start helping them do that. Um in the UK, we have quite a prescription, uh quite a principles and accountability um based framework, but it's one in which if you have well-specified and well-governed agents, they help you map their workflows onto the kind of SMCR, the senior manager's certification regime. So if I'm the manager who's responsible for this, these processes in the business, you know, if you've got kind of agents that do a whole mishmash of different things, then okay, whose agent is that? If it does these things for me, but it also works for Jeff doing a host of other things, who's really responsible for it? Who's tightly specified what it can do and controls it? Um, and you know, then that agent doesn't really map onto that accountability framework that we rely on in the UK to make sure that we achieve the right outcomes. Whereas your your platform's going to enable you to say, no, this is this is my agent. It performs this task in this way, and is therefore its operation is something I'm accountable for. You know, and and that's I think if we're going to get to the place where the regulator is is comfortable with our use of agents, they understand that it's delivering better outcomes for consumers, what will help them do that is knowing who's responsible. And your platform, I think, plays an important role in helping us get there.

SPEAKER_00 43:44

Yeah, I I I think that if we want in in the enterprise world, not only in finance, in general in the enterprise world. And when you look at the penetration and how AI is entering into the enterprise world, if you want to do more than Document handling, so unstructured data extraction, customer support, help, and I think that's it. I mean, like when you look at real, you know, enterprise application that works at scale nowadays in terms of agent. If you want to do more than that, then it has to be with the framework, not because only there is regulation and the regulator wants to understand. I think that clients want to be able to trust the output. And you as a user as well, you want to enable the the harnessing the framework that allowed to be repeatable in the process. And so I believe that that's the only way. Like it's the way of giving the tools, giving the ecosystem, the environment to enable agents to be good where they are good. Which is crazy if you think about if we were discussing data like maybe 10 years ago, we would have said completely the opposite, right? But today you want to use agents where they are good in the sense of creativity, in the sense of exploration, reasoning, thinking. And where they are not good, it's more about manipulating those types of data or do computation. I mean, I don't know if you you guys had those types of examples where you ask LM to do weighted average or stuff like that, and the results are wrong. It's like just like in terms of pure, simple mathematical operation. And that is where you need to give those agents the tools, the skills, the grounding, the harnessing to use them in this deterministic approach, plus the governance. And I think that will give enterprise in investment operation and outside the ability to really leverage on that capability.

SPEAKER_01 45:50

And um, you know, I think I think it does spell out a the vision for the future that we're heading towards, in which you know, one imagines that lots of staff who today perform some of these kind of processes that, you know, manual reconciliation, you know, working out the nav and so on, these will be done by agents who report to a human being who is then the the key judgment maker or decision maker. And uh we can see how kind of the the roles in the industry will start to shift and change as as these tools do more of these processes. What what do you think the vision is? Where where do you think we are headed as an industry now that these tools are at our disposal? You know, supposing, as we've discussed, that that the we can get the regulator comfortable with how we want to do it, we have managers who are accountable for the outcomes and make sure that these agents stay on on track and and deliver better outcomes at greater speed and and lower cost.

SPEAKER_00 46:49

Yes, I I think you will have way less need of uh people operating the nitty-gritty kind of details of the data operation. Like today you have a a lot of the burden, a lot of the the manual activity happened really on taking those Excel, taking this data, and you know, interacting with other sources, map that, copy-paste, put shit inside, doing extraction. I mean, I I see stuff that are really like where you say, like, wow, you you spend two, three days just to preparing the data when in reality this will be fully automated by a mix of agent and deterministic approach, but will be fully automated. Like the the full scale of the data ingestion, harmonization, preparation will be automated. But then on top of that, you will still have the needs of people that help in configuring, building strategize on what you make on top in terms of checks, in terms of controls, in terms of how you will position those agents, give skills to those agents, and this you need people to do that part. So I think that the the skills will be deep industry knowledge with understanding of what you can do with with agents and what you cannot do, so that you know, like that's clear, and you set the layer of uh what we were saying, like either the conversation or the anomaly detection or the contextualization based on some deterministic controls, that then you know, like you have those checks and you contextualize with other type of data where agents are very strong. And then, yeah, of course, like the the the ultimate step is human that have the final view on either an validation, if there is, you know, like uh the need of having that. But it will be with all the data prepared, all the anomaly detected, all the contextualization run, and in a way which is extremely configurable and flexible to map the different needs of the different uh uh part of the industry. So in terms of people, it's it's difficult to answer what what it will be, you know, like in in 10 years. Like I even if we are in an industry that moves slow, we are moving and and is is going in in that direction. And I I think yes, it will be more of analytical, you know, kind of reviewing and and and and strategizing in the orchestration of how I use those tools to put together the the type of the type of system that that I want. We see the same in tech firm with with engineers. It's exactly the same. We less people that do the nitty-gritty of the coding, the nitty-gritty, you know, like uh and more people that actually are there to be on this uh problem-solving mindset, architectural, you know, like uh so I think this is the kind of mindset we need everywhere in the future, and we will need everywhere in in every industry.

SPEAKER_01 50:16

Yeah. And I I, you know, off the top of my head, if I imagine the the example of the kind of newly listed Astro Y company, we had the question about, you know, how your platform could help, could help answer a question about what our exposure to that company is. And you know, if if we if we can sort of get all these processes, we can we can perfect them in some sense with with technology. That could open up new possibilities for how we serve the end investor, like the end client. It could be I I might as the as the guy you know who's just curious about, you know, did that, did Astro Y improve the performance of my portfolio, you know, in the two months since it listed, and I can write a query into my into my um my investment platform, and then that will in turn query your system that will go ahead and work that out for me, you know, with with all the information and serve me that answer. And um it seems as though that's that's you know, we we look when you look at the um the kind of appetite there is for trading crypto assets and stuff like that. People want to, they're they're monitoring it all all hours of the day. And and you know, in some sense, if we want to for traditional financial markets to stay appealing and competitive, we have to kind of move with the times a little bit. And and we can't just say, oh, you know, that it's very complicated back office systems and it's going to be very difficult to answer your query. A great world to live in would be one in which we have the systems that enable us to do that and to keep people engaged in what's happening in their in their investments.

SPEAKER_00 51:49

Clearly. That's a question about the full evolution of the industry. But yes, the the the ability of having a clean structured back office automated in terms of controls, should please everyone, should please the regulator, should please the clients, and ultimately will please the end investor because resources can be then moved into how I can better serve the the end customer. So you are totally right. I I believe that the less we spend resources and money into try to put even more people and and and controls in a way because the data is not clean, because the processes are messy, because this is where today solutions like Nextgate and Spark can help to then, as you say, focus on other value-added part of the business, which ultimately will end up to create this appeal for the new generation of investors. So I I I believe this is this is the the direction that the industry is is going.

SPEAKER_01 53:03

So, with the hope then of pleasing everyone, uh thank you very much, Davide Martucci of Nextgate Tech, for speaking with us today. I hope listeners have found it a fascinating episode. I certainly have. And until next time, thank you for listening.