Skip to main content

Alpha Advantage: AI in Capital Markets with Databricks

Summary

  • Financial firms including Franklin Templeton and LSEG are building competitive advantage not from better AI models but from proprietary data and the ability to operationalize it, with Gartner forecasting global AI spend will hit $2.59 trillion in the current year.
  • The video presents architecture patterns for governance-first AI in capital markets—including multi-agent orchestration, semantic layers for domain knowledge, and Unity Catalog with Genie Spaces—to prevent agents from operating on siloed, ungoverned data.
  • Customer stories across sales and trading, wealth management, and fixed income demonstrate how real-time intelligence agents, natural language fund screening, and the Strategy Sentinel credit risk agent are moving firms from reactive analysis to proactive decision-making.

Alpha Advantage: AI in Capital Markets with Databricks

Watch: Alpha Advantage: AI in Capital Markets with Databricks
Markets move fast, but competitive advantage doesn't come from better AI models, it comes from proprietary data and the ability to operationalize it. Financial firms face three critical challenges: data fragmentation across asset classes and desks, agents operating without governance, and automation of low-value tasks. Databricks enables firms like Franklin Templeton and LSEG to build the data foundation necessary for an intelligence loop, where multiple coordinated agents handle routine analysis while humans focus on high-value decisions.
Discover how trading desks leverage real-time intelligence agents for RFQ analysis and market briefings, how wealth advisors use AI to synthesize complex fund data and client portfolios, and how fixed income managers detect policy violations and credit risks before the market. Learn architecture patterns for governance-first AI using Unity Catalog and Genie Spaces, multi-agent orchestration, semantic layers for domain knowledge, and alternative data signals for predictive analytics. These customer stories demonstrate that with proper data foundation, firms can move from reactive analysis to proactive decision-making.
🤝

Chapters

FAQs

Why is proprietary data more important than AI model quality for capital markets firms?

This video argues that competitive advantage comes from the foundation beneath the model—proprietary, unified, and governed data—not from the latest frontier model. Gartner's forecast of $2.59 trillion in AI spend is contrasted with a Bain finding that 40% of respondents see 10% or less in actual cost savings, a gap attributed to siloed and ungoverned data rather than model limitations.

How does Franklin Templeton use AI to support wealth advisors?

Franklin Templeton built the Signals tool, a governance-first research insights platform that allows wealth advisors to synthesize complex fund data and client portfolios. This video includes a demo of the Signals tool for product comparison, showing how natural language access to governed data transforms advisor workflows on the Databricks Data and AI platform.

What is the Strategy Sentinel agent and what does it do?

The Strategy Sentinel is a fixed income agent described in this video that monitors portfolios for policy violations and credit risks before the market acts on them. It is presented as an example of how agents can move fixed income managers from reactive to proactive decision-making by continuously scanning positions and flagging emerging risks such as credit events.

What are the three critical challenges facing financial firms deploying AI?

This video identifies three challenges: data fragmentation across asset classes, desks, and vendors where agents cannot reason on data they cannot access; agents operating without governance at 10x the speed and scale of previous ungoverned Python deployments; and automation being applied to low-value tasks rather than high-value decisions. FINRA's 2026 report is cited as requiring that the chain of reasoning be constructed behind every agent.

Full transcript

[00:07] All right, good morning everyone. As convention, I'm Andrea Dosa. I'm glad to be here today. So, we are going to talk about the alpha advant advantage. So, markets are moving fast and I think underneath all the volatility we're all grappling with technology, capital intensity, and what does AI mean for
[00:23] everything. So, what are we seeing at data bricks? We're seeing that the edge is being built on proprietary data and the ability to operationalize it. It's not just about the latest model, but it's actually about the foundation underneath. So, let's start with some stats. You
[00:39] know, the numbers tell a story that people are hearing about in the room. So, Gartner just forecasted that global AI spend will hit $2.59 trillion this year. And it's not a rounding error. It's the single largest corporate
[00:55] capital deployment in history. And yet, we also heard from Bane a few weeks ago that when you look at the cost savings, 40% of the respondents, they only see 10% or less is actually happening. So, part of the reason is it's not a
[01:11] technology gap, it's a foundation gap. So you can't build a production AI system on silo data that's ungoverned and not actually being able to speak to each other. So the firms pulling ahead, as I mentioned, they're not the ones
[01:26] with the biggest AI budget or with the most data. They're the ones who are working smarter to build that foundation today. And those numbers don't actually surprise me because as I'm meeting with clients, three things are happening. First, the data isn't ready. And I know
[01:44] what you're going to say. We've been modernizing for a decade. We have a data lake. We have a uh cloud strategy, but the data is still siloed by asset class, by desk, by vendor. Equities and fixed income is separate. Risk isn't
[02:00] talking to trading. And guess what? Agents can't rationalize on data that they don't have. So last year on our data bricks stage, Jamie Diamond actually said it best. Getting the data in order, that's the hard part. So second, agents are running without
[02:16] governance. And we've seen this move before. A few years ago, we were all psyched about the Python analyst. We gave them access to production systems. No governance. We didn't know what was happening. So that's happening again at 10x the speed and scale with agents. And
[02:32] FINRA in their 2026 report actually said the chain of reasoning must be constructed behind every agent. And as they like to say, the honeymoon phase is officially over. So the regulators haven't walked in yet, but
[02:49] we're all in a regulated industry, and they will. So third, firms are automating the wrong things. Everyone's using an LLM, but everyone's asking them the same questions. We're summarizing the same documents. We're not really being efficient. We're using tokens for
[03:05] single use, and then you're rewriting that document again later. That's inefficient. We need to build processes. So with that the highv value work still sits in human hands but we need to work more efficiently more you know productive so to speak. So while
[03:22] everyone's automating nobody's actually building advantage and the window to change that is not going to be open forever. So the way that we look at it at data bricks we see this as a trajectory. So we're not seeing you know just the tech
[03:38] change. We're seeing how the work actually changes. So most of us right now are in the co-pilot phase. You're using you're putting in your research. You're putting in your notes. It's genuinely useful, but the analyst still has to ask the questions and know you
[03:53] know what they want to know. So then we're going to move into the second one, the Gent era. So this is when we start to see more of that unified view with our data. This is where we start to, you know, kind of put the pieces together in a unified platform where we're going to
[04:08] get proactive signals, a risk signal or a ratings change and you're going to have some audit trail. But what we really want to do is we want to build towards the future, which is the intelligence loop. So this is where the
[04:24] humans are using and doing that alpha advantage so to speak the thesis the risk appetite the strategy but now imagine that their workflows we put multiple agents in the background to do the heavy lifting. So what does that mean? You have one agent that's
[04:40] scrolling through all those filings. Then you have another agent that's flagging those positions that need to be looked at across all your portfolios. And third you have an agent generating those emails to the PM. So before you even had your morning coffee, you already know what you're going to be
[04:55] looking at when you hit the desk. And that's the intelligence loop and that's what we're working towards. So the firms that are working towards that, it's not about a better model. It's about unlocking their data on a platform and they're building the foundation first. So if that's where the industry is
[05:12] headed, the key question is who's building that foundation today? And capital markets are running on data bricks. We're doing it today. They're leveraging this platform across the buy side, across the sell side, exchanges, data providers. They're focusing on
[05:28] building the foundations today. You can recognize some of the large names. We're across 600 different capital markets clients, 2,000 in financial services. We have over 500 different data listings and growing. And we have 250 ISV and SI partners to get you there. So,
[05:46] the foundation is there, the platform is there. And if that's who's on the journey, what do I need in a platform? So you guys have heard it in our keynote. This is what we like to joke around, our new layer cake, so to speak. But it's about
[06:02] three things. It's about agentic data number one because let's be real, your data was not built for the Agentic era. So this is what that solves. You have your lakehouse, lake base, lake flow, single foundation, no side stats.
[06:18] Second, you have that unified governance. But it's not just about governing your data. It's about giving it the context to make sure that when you start to build those agentic workflows, it understands and unlocks your proprietary data, your secret sauce. And then we end with third,
[06:34] agentic applications and work. Genie is the AI assistant that's not only going to understand your data, it's going to allow you to build code. It's going to give you complex analysis and allow you to build that multi-gentic workflow. So, the foundation is ready and now I'm
[06:52] going to show you what that looks like in two different workflows. Sales and trading and then investment research and wealth. So to get started right now when you look at a trading desk you have your RFQ
[07:07] your execution your trace history you have it in different spreadsheets you have it across many different execution venues um you know basically the data is everywhere you lose the RFQ pre and post trade you find out after the fact where
[07:22] it was on the trace tape the client's already gone but the future looks differently AI becomes comes part of the operating model. It's not just bolted on as an afterthought. So imagine today a head of desk has a
[07:40] nice little landing page, probably has a chatbot, but what if before you even have your coffee in the morning, you had agents running in the background on your data. So first you had an agent that was looking at all of your RFQ history. You
[07:57] add a second agent looking at all the trace reporting from yesterday. You have another agent looking at your win rates. And then lastly, you have an agent summarizing everything and the information that you like, which ends up with your morning briefing IG holding steady high yield win rate slipping in
[08:13] energy. That's great to start my day, but then what about throughout the day? What about my axes? You could do that, too. Basically, it's taking your same kind of data. It's looking at your history. It's looking at the different trace prints and it's giving you then your updated access in real time, not in
[08:29] a black box with cited information so you can make the decision sooner and faster. And that's not just an agent an agent that you're asking a question. That's an intelligence loop that's making you win deals, win workflow, and let's be real, make more money.
[08:47] So, let's take a look at what that looks like on the wealth side. again how it actually works today everyone has tons of PDFs you have tons of different kind of uh documentation you're overloaded with all this data
[09:03] then you have you know analyst capacity it's not expanding they're having to do all this manual work and then third AI again it's being useful because you're able to ask a question so it's kind of like a research assistant but what do they actually want what do
[09:19] you actually want in the future you want to have what we like to call the investment on intelligence agents orchestrating the full fronttoback workflow so you can do what you need to do best and get back to your client. So again, let's walk through the
[09:34] workflow. This today in the morning, an FA goes in, opens four or five different systems, they check their emails, their spreadsheets. It's a maybe 11 o'clock in the morning before they call their first client. Now imagine that the FA walks in to their home screen where they have all
[09:51] their KPIs across the top, their AUM, you know, anything that needs to be flagged, any of the drift analysis. Now imagine that they can click into the alert on a particular position like KKR and they don't need to see a document dump. They get the summary at the top.
[10:07] They no longer have to update the spreadsheet manually. We give you the delta and then we also flag in the earnings why it's deteriorating management tone. But let's take it to the next level. Not only did the agent flag that, now it goes through all your portfolios, what clients are holding
[10:24] that position and who do you actually need to contact and have that discussion about maybe changing their position or what they want to invest in? Then you get another agent that's actually doing the tailored messages the way that your clients like to be talked to. And the human in the loop comes in and that's
[10:40] where you want to be spending your highv value talent. They're the ones doing what AI can't. they are having the conversations with the clients. So this is what that looks like with an always on intelligent agents and again the intelligence loop. So this is what we're
[10:55] looking like when we look at now next and future making AI work for you in a proactive manner and not just in a reactive way. And so I'll leave you with a story. Recently I was with an alternative
[11:11] manager really smart firm director of research and he's just like we have all the data I just need the AI capabilities on top and what he was describing was data bricks and he didn't know it yet. So the demand is there everyone in the
[11:26] room already knows that they want to get the AI capabilities. The gap is knowing that you need a system that can unify all your different data that can govern it and give you the proprietary control and not just bolt it on. So with that, you saw what it's like in production. I'm going to hand it off to two of our
[11:42] clients to show you how they're doing it today. Thanks everyone. Thanks everybody for coming out today. I'm here to talk about some of the work that we do in our portfolio solutions team here uh at Franklin Templeton. I'm going to talk about how we are structuring uh scaling structured
[11:59] investment insights using data bricks. When I look at the investment universe out there, funds, ETFs, SMAs, my first impression is too much product complexity, not enough scalable decision support, which is what we were trying to solve and what data bricks helped us
[12:15] productionalize u with some of our inality um insights. All right, let's uh let's jump into this. This is a little bit about our team. Um we're a portfolio solutions team at Franklin Templeton. We started
[12:32] out mainly supporting distribution, working with financial adviserss, working with institutional investors, helping them understand how they can partner with Franklin Templeton, with our $1.6 trillion in AUM, how they can partner with us, get into the right
[12:48] investment strategies, the right vehicles that's ultimately going to benefit their client. So, what we do um is we turn complexity into investment decisions. My one piece of advice to all of you guys as you're working with AI and data, find the interesting problem
[13:04] to tackle. Don't worry about the back office. Find the thing that's going to make decisions happen sooner and faster and you'll you'll be rewarded. Um, so we work with financial professionals. We translate research into insights. Um, I spend a ton of my time on structured
[13:20] data and we build and we buy data. And I think that's what's different about the capital market space. compared to a lot of the other industries out there, there's tons and tons of capital markets, third-party data that you need to buy,
[13:37] you need to have to be able to relate your data to it. And I think the combination of those two things sets this industry apart. And if you do it right and you do it in a meaningful way, um, you're going to have great client conversations. So, um, some of the work that we've done, uh, is really
[13:54] highlighted in our portfolio analytics tool. This is our online uh tool for adviserss to use to build portfolios, analyze portfolios, review and research products, investment products. Um hopefully over time you see some of the work that we do here with AI and data
[14:12] bubble itself up into into that tool right there. But one of the great things is we do have that as a as a tool that we can build into that's available for clients. So, not only are we working with um internal salespeople, internal portfolio managers, but where our
[14:27] research um reaches a threshold of of high usage, we can get it out to clients as well. All right. So, let me talk about the work that we've done um and why uh if you meet me, I'm probably one of the least technical people here. Um I'm an investments person, CFA. I've
[14:44] always been supporting um portfolio managers, investment teams or in distribution myself. Um and the last couple years I've been managing a team of of analysts who really help clients get into those right positions. We were doing tons of work. As you know, there's
[15:01] 6,000 mutual funds in the US. There's 4,000 ETFs. I think that's over 6,000 now. Uh 30 trillion in assets. We're trying to communicate with up to, you know, 280,000 financial advisors. And when you think about building portfolios, there's 127 morning star
[15:18] categories. I don't know how many liper categories there are, but you can fill me in. Um the uh but the reality is we can't service that, you know, and we have 300 strategies. To be honest, some of the work that my colleague and I have done here, well, when we started this journey, it was trying to understand our
[15:35] own volume of product. Um but to do that we really need to to evaluate the entire universe. So 300 strategies at Franklin uh we've got 220 people externally facing salespeople just in the US that we were trying to support. Doing that
[15:51] with the help of seven analysts and seven consultants that we were managing going flat out non-stop we could only do 6,000 engagements a year. So the question is how do we scale the type of research that we're doing? um how do we scale it and how do we do it without sacrificing judgment or or quality and
[16:08] how do we do it with governance uh because everything we do is under a microscope from FINRA the SEC every client communication that we did at an ad hoc level was under a microscope the AI piece isn't going to be any different but if you apply those same exact rules
[16:24] standard procedures u you should be able to scale with the help of your compliance um all right why the old model didn't work I kind of told we couldn't keep up with volume 6,000 and growing. Um, so we tried building tools, right? We we tried a manual model. We
[16:40] built a library. We wrote static notes and we put them out there. Um, we couldn't keep up with with with even maintaining that. We're talking about just 200 products. The analysted doing it. I couldn't tell the ROI and the things the notes quickly got out of
[16:56] date. Um, how could we solve that problem? So, that didn't really work well. Static tools. Um, again, dashboards, we all know they're useful to an extent, but then over time they become less useful. Um, and they break down. Uh, and then early gen AI, you
[17:12] know, everybody is just throwing in, hey, tell me about this fund versus this fund and to chat GPT and, you know, they think they can send it to their client without thinking for two seconds. Um, that's not repeatable. You know, that's that's ad hoc work as well. Again, we're trying to solve ad hoc work. I don't
[17:28] want to see ad hoc work uh in this workflow. I want to be able to to facilitate quick client communications investment decisions for clients. So, uh the early gen AI didn't work. Uh it wasn't until tools came along like we've got with data bricks where you can plug
[17:44] some of these AIs into your data which then quickly helped us scale what we're trying to do here. So, I'm going to go through a couple demos. Sorry, I'm going real quick. We built um signals to scale our analyst quality insights uh for the US distribution team with a consistent
[17:59] process. So, we've got a couple things. We've got our we we've really built um a structured data set that helps us communicate and we do that at an ad hoc level, but then we said, let's let's open it up to doing this on a digital level. So, uh we've got our structured data on the left, which is our our
[18:16] grading process um and performance profiles. It's all quantitative uh that I run. Uh we've got our internal tool called signals which will help identify um replacements complements and monitor portfolio construction opportunities. And then when we add agent bricks on top
[18:32] of that we can really um help expose what the data is telling us and why it's useful. So I want to highlight this AI did not create our research architecture um but it made it scalable. All right. Um this is probably the least the most
[18:48] boring slide here. uh governed inputs, agents, uh evaluation and government. So, we're using Unity catalog. We've got context judges there to make sure that our data is relevant, accurate, concise, and comprehensive. Uh we are looking for things to break down. Um and really what
[19:04] we're doing is we're taking everything on the left and moving it into what are those business outputs. So, always be thinking about what are you trying to achieve from a business standpoint. Um it needs to be important. So, in our case, we're trying to help uh facilitate investment decisions fast. Um, way
[19:22] faster than you send me an email, it takes me a day to look at it, another day to write it up, another day for an analyst on the team to review it before I send it back to you. 4 days later, the adviser and the salesperson don't even remember they had that conversation about that specific community bond fund. If I can get you product comparisons,
[19:39] research notes, uh, and portfolio comparisons almost instantaneously for you to interact with, that's going to facilitate more conversations with Franklin Templeton. Even if you're not buying our product, you're going to work with our tool. You're going to meet with our salesperson over time. You're going
[19:54] to find the right product to blend into your portfolio. So, structured data gave the agents context. Governance made the output usable. Here's the demo. Uh, let's see. It starts running. Um, this is internal So, I've got two things across the top. We're going to see this is it's called signals, but it's in our
[20:10] portfolio analytics tool, the internal version. This is our recommendation engine. So, let's just say I threw in uh iShares Russell 1000 value ETF. It's telling me quickly we've got it's going too fast. Uh we've got Putinham focus large value cap ETF in that same space that really outperforms it. So, I
[20:27] quickly here show how it outperforms it. I've got my own specific view on how I want to communicate and how I want the sales team to communicate. Does it do well in a rising market, falling market, is it consistent? Um, it gives me that detail headto head. So, I can compare
[20:42] those two products instantaneously. I know it's a go no-go on that. And I can do that for an entire matrix, an entire portfolio in seconds. Um, here, this is called Funscribe. And this one, we use a more detailed agent. This is great because we, this is where we want to
[20:58] know more about a product, more about our own product. We're looking for consistency of research. um you know, I'm a salesperson. Tell me more about, you know, p val before I um uh oops before I um talk to my client. And I can um let's see if I press play it would
[21:15] run again. Um it's going to give me, you know, the full note. And that note is actually trained off of the analyst notes my team wrote. So this is an agent that we use, the no code agent, uh that we got help from data bricks to train on those 200 notes that we've been authoring for years. Um, we've got a
[21:32] backlog of those. Um, we have it specifically writing notes in a way that I think gets the point across for our salespeople to then say, "Okay, I can go talk about this product efficiently." Um, and then I do expose all the performance data that you would see. This is a fantastic fund. You can see it's like hitting all four uh quadrants
[21:49] here. Um, but the great thing is I can quickly compare and contrast one fund to another on the recommendation piece. Matrix review, they can do an entire thing. Leaderboard kind of reminds me of like Google. It ranks products based on how good they are. Um, and it's entirely
[22:04] fed by industry data and then our own structured data. So, all right. What this has uh done for us um it's it's augmented. It's not replaced the analyst. Um, within the first month when I was measuring this stuff, uh, we got
[22:20] from the salespeople. I mean, this is great. sales people were using it and they were uncovering opportunities whether it's a takeover opportunity or a complement opportunity or even just an additional consideration for a portfolio. So just within like two weeks I got $15 million of feedback from the team and this has been out for you know
[22:36] over a year and a half and I I don't have a good way to keep track of it. Um but we've moved from you know saving analysts two hours a week. We're covering hundreds of products with AI and now the sales people have these tools at their fingertips. So, I know I'm going really long. Sorry about that.
[22:52] Um, signals today, internals internal tools. Next, how do we get those governed insights into uh the hands of financial professionals through our portfolio analytics tools? So, it'll be some version of what you've seen here today. Um, and then the future, you know, how do we help do these same type of things? Portfolio construction, asset
[23:10] allocation, how do we meet the client where they are? Do they come to us with CME? Do they come to us with an outlook? um how can we then react with the data we have to help them fulfill the portfolios that they need. So structured decision uh governed data scalable
[23:25] portfolio construction insights is really what we're we're going for. Um the destination here was not just writing better commentary or content. Um it was better portfolio construction overall and the ability to govern this uh govern these insights uh and embed
[23:40] them into workflows uh is really what's going to help us communicate with financial professionals to design and enhance their portfolios. Thank you very much. Thank you. All right. So um I'm also
[23:56] actually working in funds. So there's some overlap with what Franklin Templeton just spoke about which should be interesting to see the the contrast between those two things. Um what I'm going to talk about actually is a bit more on the technical side really what it took to uh make something like this possible. You know what did we do in data bricks? How do we architect the
[24:12] system and what were the decisions that led to us having high output quality. Um it's similar to a lot of what you heard over the past couple days right context engineering having really organized genie spaces how you orchestrate things and that's what really gives you a solid foundation to work from.
[24:35] So, first a little bit of context. So, Liipper is LSG's global fund intelligence platform. Uh, it covers both mutual funds and ETFs with fund classifications, ratings, performance data, benchmarks, and a long historical time series. Um, the reason this matters for AI is right, the scale and the complexity. It's not a small demo data
[24:51] set or a toy problem. We are talking about hundreds of thousands of funds and share classes across more than 80 countries with more than 40 years of performance history. Um has a very rich analytical surface area. But it also creates a challenge. Right? If a user asks a simple question in natural
[25:07] language, there may be a lot of financial meaning hidden inside that question. And we're really focused on democratizing access to this information to wealth advisers in places where they're acting, right? they might not have a uh they might not know all the terminology they might need to engage in these conversations.
[25:24] So there's four reasons why this worked pretty well, right? So first the the data has very robust identifiers, right? In financial services if you don't have um clear identifiers resolving across fund and share classes is complicated. Um and Liipper has through Lseg's uh uh
[25:40] data has extremely um uh solid identifiers across all the markets in which we operate. Uh the second is the taxonomy. So Liipper doesn't just say equity fund, right? It has asset class, geography, style, uh strategy and a hierarchy, right? A really rich taxonomy
[25:56] to work with. Um you need really high specificity when you're dealing with LM. If you have a rough taxonomy, you will get back very rough answers. Third is the history. So we have decades of performance and that means that we also have decades of different regimes to work with. So we can look at what
[26:12] similarity means right across long stretches of time or short or short stretches of times or other areas as well. And finally the entities are modeled in a structured way in which financial professionals think about them right so we think about the benchmarks and the currencies and classifications and the funds and share classes in a way
[26:29] that a wealth manager or a portfolio manager or investment adviser right is going to think about putting those things together. It's all encoded in the system. So it wasn't all smooth sailing, right? So we really ran into three failure modes when we started. Um the first was
[26:44] the performance at scale. Um if the model had access to too many tables, um the latency and the reliability degraded. It wasn't just that you might get back a bad answer. Very often you did not get back an answer at all. Right? All right. So that's when we started to build purpose-built views
[27:00] that really optimized around the types of things we were seeing uh in the questions our customers were asking us. The second was query routing. So different questions require different data domains. So you might have a profile question or performance question or some combination classification question. Um and you need to you you
[27:18] can't handle all those in the exact same way. And that's where we built segmented genie spaces. Right? Right. So a lot of the context engineering you heard about geneontology, we had to build some of that ourselves basically early on to get the questions directly to the right places. Basically a multi- aent architecture to hit the right data
[27:33] domains. And the third was the gap between metadata and semantics. Right? So we had thought we had pretty good semantics at the beginning. What we really had was a very nice data dictionary and there's quite a lot left. Right? So we had to encode how financial professional actually interprets the data which many data dictionaries really
[27:50] are just column level information. Right? they don't have that level of granularity and so we built that semantic layer that encoded the domain knowledge from our researchers and our product managers and our engineers who'd worked in this data for you know 35 years um so I have a very simplified version
[28:05] of the architecture right the user asks a natural language question we route that question to the correct genie space using our orchestrator and then each genie space owns a specific domain such as profile or performance and underneath that the unity catalog provides us the coverage governance lineage access
[28:21] controls, com level metadata, and semantic definitions. And then the semantic layer as the financial interpretation, your taxonomy, standardized answers, joins, and the business logic, right, that binds it together that's needed to provide consistent, reliable responses.
[28:37] Internally today, we use this for our own natural language fund screening and our cross fund performance comparisons. We're also uh experimenting with generating AI uh fact sheets. All right, so I'm going to show a quick demo. Now, this demo would ordinarily take four minutes. I'm not going to make
[28:53] you sit here for four minutes. So, I've cropped the time down to one minute. Um, it's still five times faster than the 20 minutes it takes an analyst to do this operation today. Um, when you're watching the demo, just keep three things in the back of your mind, right? So, of course, the user is not writing SQL. That's table stakes at this point, right? But the question is
[29:08] getting routed the right analytical domains. You're getting a question about performance. You're getting a question about uh the profile and the expenses. And it's going to all those right places to get it. And finally, it's a generic it's it's not a generic response. to the response that we've crafted based on how users want to look at this data.
[29:26] So this thing running. Okay. So uh here I'm actually in claude and the reason I'm in claude is because we have a hosted MCP on data bricks. Right? So we're embedding it in places where users want to do work. That's claude or um openai or other other places. Right? So you can run it
[29:42] anywhere. So I'm invested in a state street S&P 500 ETF. Um I want higher performance. I would like some lower fees. Right? Everybody would like those things. uh and it's going to do an analysis to come back with one, three, and fiveyear returns and then an expense ratio that allows me to compare those
[29:58] funds to each other in the table. And everything below the table is some additional things that Claude has thought about tax advantages, you know, can you actually invest your money in uh that fund based on your brokerage account and things like that as well, right? So, as I said before, this would ordinarily take four minutes. Um still
[30:16] five times faster from the 20 minutes it's taking our customers to do it today. Um, and with AI, right, this is the slowest it's ever going to be. Um, customers are able to save, you know, hundreds of hours by doing this process and make better decisions for their customers in a controlled, governed, and repeatable way. So now I'm going to get
[30:31] a little more technical and get into the details. Um, so here is the anatomy of the query. It's the same query that we just looked at, but we're in ML flow inside of data bricks. So you can see what's actually happening under the hood. So the blue are the LLM calls. The red are the tool calls. And so the first LM call creates the plan, right? Then
[30:48] the system executes. It evaluates, executes again and it continues until it has deemed that it has sufficient information to answer the question. Right? So it actually loops several times. The orchestrator goes and sends different agents out to different places and it's evaluating on its own is it satisfied with the quality of answer
[31:04] that it's created. Right? So um there's two things I want to highlight. Right? So actually most of what we built is the orchestration. The genie spaces were incredibly easy to work with. And so we really have to worry about the problem of answer quality evaluations and the orchestration process. Second is that because the system identifies its own
[31:21] blind spots and feeds it back to us. It's helping us improve it. It's not self it's not self-improving but it says hey you know it doesn't you don't have a certain return here. I still think it's okay. What do you think? And we can then go and change the semantics or look at the answer and update the system. We also pass along that information back to
[31:36] the customer so they have confidence in the answer that they're getting. uh sort of on theme with the past few days. Uh my advice to everyone, tune the data, not the model. Um you know, the instinct is often to work on the prompts. You should not work on the prompts. I mean, you could do a little
[31:51] bit, but I would not go nuts. The real the real reality is your failure modes are data problems. They're not most likely prompt problems. Um we found very often that we need to focus more on field descriptions or improving the join logic or explaining why you have to join tables a certain way so the AI could understand that and process that in the
[32:06] future. Um, and so we got the most bang for our buck by focusing on that data layer. Um, as I always say, the frontier model companies are not improving your data, right? They're improving their models. You've improve your data, they will catch up with you, right? And it'll
[32:21] it'll work better for you in the long run. Um, all right, last slide. Uh, quick Liipper pitch. I got to, right? I run Liipper. Um, the data and ratings are great. We've got people at the conference, booth 155, people in the front row who'd be happy to talk to you about it. Um I think the maybe the final
[32:39] lesson is that when you build an AI fun screen or really any AI tool, you're building two things, right? You're building both the application, but you're also building the data layer. That same data layer that you can provide to your customers so they can integrate it just like Franklin Templeton's integrating financial data from all over the world. Thank you. So,
[32:55] I'm going to walk you through a day in the life of a fixed income desk at one of our investment management customers and an application that we built for them to sit on top of their in uh investment policy data, their investment
[33:12] strategy data, and their portfolio data and automate a set of workflows that right now before data bricks was taking them weeks to make an investment decision because when issues arose, they'd have to go into all sorts of different systems. The researchers have their own system. The quant operate in
[33:28] their own systems, complying a risk of their own systems. They're often copying data out, combining it into spreadsheets, disjointed communication, making it very difficult for them to reach a decision quickly. So with data bricks and building an application on top of the data brick stack, they have a
[33:44] fully governed view of all this data that agents can do to actually make decisions about the data. uh leveraging things like unity catalog or AI gateway and agent bricks to streamline this decision- making into something that now
[33:59] can take hours or even minutes in some situations. So let's jump in. Starts with the strategy sentinel. So the strategy sentinel is a set of agents that are supervised that have consumed and understand the firm's investment
[34:16] policies. So it understands all the rules that govern how investments have to work at that firm and it also is looking at the portfolio information and actually overnight it's it's identified two breaches that have occurred. So as an analyst I walk in the morning and right away I see okay I've got some
[34:32] issues I need to deal with. This first one is related to the fact that we have some uh portfolios where treasuries have rolled off and no one had noticed. Right? So we have some portfolios where treasuries we need to actually swap them out. Trades need to be made and so this has automatically
[34:49] identified this for me overnight and I can see that this has been going on for a while. It's starting to impact the portfolio performance because we're starting to see that according to our benchmark we're now below where we need to be. And I can see the investment policy constraints that it actually
[35:05] triggered. So it's fully explaining to me as the analyst something that I would have had to spend hours of researching in order to identify for me very quickly. And I can even run the agent cascade to say okay how did the agent actually come up with this answer. What did it actually do? And I can see it
[35:22] engaged the drift monitor monitor. It checked that against the constraints that we have in our policies because it understands our policies. It has that context. It did a root cause analysis and it actually identified, yes, there are treasuries that have rolled off that no one has noticed and we need to
[35:37] actually update the portfolio and it could even draft uh an escalation notice for me. But I want to go to the other breach it identified and the rest of our demo is going to really focus around that and that's an issue that occurred with Boeing. Boeing has just been downgraded yet again. It's
[35:54] a fallen angel now. its bonds in our portfolio are no longer investment grade and that's violating one of our clients investment grade constraints and we're going to need to do something about it. The agents also actually given us a timeline view of events that
[36:11] have occurred that have led this to happen in Boeing. And we can see all the way back in 20 uh 24, Boeing had that infamous uh um door slug issue where they had some supply chain issues and a door came off mid-flight from one of their Dreamliners led to spreads
[36:27] widening and downgrades. In uh in 25 business um environment for Boeing had deteriorated even more and they were put on credit watch by the agencies. Once again, spreads in the market started to uh widen for them. And
[36:42] now the the fallen agent angel event has happened. And once again, it's telling us these constraints that have been violated according to our policies. It's showing us the locks that we're going to have to replace. And I can run the cascade. And once again, it's going to
[36:59] help me understand why did the agents make this decision. I can validate that as a human. And I can even draft to our portfolio manager or to our credit uh research department, to our compliance department, an instant communication about this issue
[37:15] or download my compliance report from the agent to send that on to compliance. So full audibility, full traceability. But it wasn't just the strategy sentinel that found issues with Boeing. We also have a research accelerator that's
[37:30] looking at our portfolio information, right? and it actually flagged Boeing uh for a credit issue. So I can actually generate this an analyst justification. What did the agent actually find for me? And they're actually three major events that are going on that are causing
[37:45] deterioration in Boeing's credit. So the first one is they're in close they're real close to actually breaching financial covenants. their um their debt to equity ratio is getting to a point where their lenders are are saying you
[38:02] might breach and that's a big problem for them and the market is also seeing deteriorating sentiment uh about them. They're seeing um conditions worsen. So the the market spreads for Boeing are increasing and there's also negative tone coming out about the the company as well. So as I scroll down I can actually
[38:18] see the metrics the agents brought to me the metrics that it actually used to make this decision. I can see the debt to IBIDA ratio is increasing. I can see interest in the market coverage is is going down. Their margins are compression are are compressing. Their revenue growth is slowing, right? And
[38:35] all that's leading to these issues like the covenant headroom issues and things that you know order backlogs and undelivered inventory. We can see the the environment for Boeing is worsening, right? And I can also see where the agent picked up this negative tone sentiment from different earnings calls
[38:51] or AKs or 10 Q's coming out of management. So once again, we're combining three different data points to see, okay, we've got a credit risk uh a credit risk with Boeing. So I've got two independent things that are already telling me, hey, something's wrong with Boeing, right? Got my strategy sentinel.
[39:09] I've got my my research accelerator. But wouldn't it be great? These are telling me about Boeing sort of in real time. Wouldn't it be great if I could have known that ahead of time, right? Could I have predicted this? And actually, the quants built a alternative signal engine. So, this is a series of
[39:24] different models. They've got dozens of models that operate against the ground truth, the performance of different positions in the portfolios and use alternative data. could be shipping data, could be oil future data, it could be public sentiment data, it could be any of that kind of data that's indirect
[39:40] to the actual performance of the company, but could be a leading indicator for the problems that were happening that led up to those market events that caused those downgrades and caused those spreads to widen. And in fact, it can even continually evaluate all those models over time. So, it has a hero model that it's chosen and I can
[39:57] even uh click in and see the notebook for that model, right? So these are proprietary models built by the quants and the researchers at the firm, right? It's using their proprietary and intellectual property. The agents are just orchestrating that for them
[40:19] and I can see all those different models how they're performing. It'll even recommend different models for me if I want. So the agent So what the uh the quans wanted is let me get this data into my dashboard. I can go see it in ML flow but it's bringing it into them. So they're actually always evaluating the models that they're seeing. But I think
[40:34] the most interesting thing is this chart here, which is actually predicting what the market saw later on. Right? If we look at the red line, that's the credit spreads. We can see the door plug event that occurred. We can see the the uh the credit watch that occurred. And
[40:50] all the way on the right now we know is when they became a fallen angel and they've lost their investment grade quality, right? The blue line is the signal detector, right, of that champion model. And we can see leading up to each one of those events, it was predicting something was going wrong. So we could
[41:05] have known about this 40 days plus ahead and been ahead of the market. So rather than being, you know, very responsive to the market, let's be predictive and let's even be ahead of the market. And that's what the uh this alternative engine was able to do. But I can even
[41:22] take it a step further. I might decide, okay, well, I love what these models are doing, but I'd like to actually do some interrogation myself. So, here's all my different alternative uh data sources, and I'm going to choose the travel card
[41:38] spend data, and I'm going to use the social professional post because actually, when I look at um my live signal that the agent's giving me, what it's telling me is that travel sentiment and safety concerns were actually the
[41:53] things that were driving the downgrades. Right? So, this is what the model's telling me. I want to confirm that. So, I'm looking at the data sets and I'm going to select travel card. I'm going to select these different uh
[42:08] social posts and some of them are coming from professionals and I'm going to throw them into a genie room and I'm going to ask it a question to let it validate for me. Is it true? Are these things actually happening? Right? And I can obviously ask it lots of questions about this data. So, I don't have to just be beholden to the model. I can
[42:25] actually allow the model to um I can validate what the model's going through and actually work with the data myself. I might even form a new hypothesis out of this, right? Uh and so I can see yes, I mean air travel demand uh is is um
[42:40] falling and safety is still a concern and that's also that safety sentiment's like a negative indicator. It's falling for them. So So I've got actually three independent systems that are really all drawing to the same conclusion. I need to do something about Boeing in my portfolio. So, what am I going to do? Well, the last agent is our portfolio
[42:57] construction agent. And obviously, I can communicate with the portfolio manager and they have processes that they can do to rebalance portfolios. But here, the system has all of those past portfolios. It understands their strategies. It knows their policies. So, it actually can recommend some strategy changes and
[43:13] it's done exactly that. And there are three top ones that it's chosen, right? The first is, hey, let's put on a hedge. Let's let's do an oil hedge. Let's do a fuel hedge. The other is, you know what, let's swap out Boeing, but we want to make sure we stay in uh the in in the same sector, so let's get a more
[43:29] diversified group of peers or let's just upgrade to a a better grade bond. And it's recommended the middle one. We can see why because when we look at the option comparison, we can see the metrics that they're measuring portfolios against, you know,
[43:44] diversification, risk, ratings, performance against benchmarks, um, turnover, it's doing the best. So, we're going to take that and we're going to say we like that. We can scroll further down and it's actually showing us what it would do. So, it's how we would sell 2 and a half% of
[44:00] Boeing and the different basket of other aerospace providers that we would buy in its place. It's run against the stress tests that the quants and the researchers have identified. So I can see how it did during the rate shock during COVID, during the uh investment grade compression period, right? And I
[44:17] can say, you know what, I want to construct this trade. What would this trade actually look like? And I can take it to the trade construction agent and it's going to show me what the definition of the trade is. It's going to show me the constraints that came out of that trading strategy that drove that new allocation. It's going to show me
[44:33] that allocation. It's going to show me that validation that we've gone through. And it's even going to give me an explanation. You know, our decision here is to stay long in a defense, but to swap Boeing out for a more diversified set of of positions. And I could even
[44:49] download this. I can share this with the customer or I can take this to the portfolio manager. So, I can instantly take this recommendation and send it to the folks who can actually make change. And so we've gone end to end with a process that's usually takes weeks and done it
[45:06] really in a demo here for a matter of minutes but maybe it would take hours. I mean things that processes that would have taken me all morning are now taking minutes to do. And it's really why what we see is data bricks becoming the a platform of destination for investment managers and for trading desks and for
[45:23] portfolio managers because we can give them this governed universal view on top of their data and automate their decision makingaking process and really give them the scale they need to grow their businesses.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.