Future GenAI Use Cases for Financial Services - with Marco Argenti of Goldman Sachs AI in Financial Services ยท 6 Nov 2023 Source: https://aiandbanking.libsyn.com/future-genai-use-cases-for-financial-services-with-marco-argenti-of-goldman-sachs Transcript: https://podcasts.opennash.com/shows/ai-in-financial-services/future-genai-use-cases-for-financial-services-with-marco-argenti-of-goldman-sach-c7de07/ (via OpenNash Podcasts, machine-generated) Welcome, everyone, to the AI and Financial Services podcast. I'm Matthew DeMello, Senior Editor here at Emerge Technology Research. Today's guest is Marco Argenti, Chief Information Officer at Goldman Sachs. Before working at Goldman Sachs, Marco served as SVP of Developer Services at Nokia, then Vice President of Technology of Amazon Web Services (AWS) in 2013, where he started and ran several AWS businesses, including mobile, serverless computing, Internet of Things, and augmented and virtual reality. Following his appearance on our sister podcast, the AI in Business podcast, Marco surveys emerging use cases in new generative AI capabilities for financial services on today's show, and offers his vision for what the future of the industry will look like when these technologies become commonplace. Without further ado, here's our conversation. So having you on our sister program, I thought we could dive into more specific use cases for financial services, especially given everything that we talked about in our AI and business episode about generative AI. Let's just start there. You gave an interview with Business Insider. talking about generative AI use cases, I'd like to explore in more detail. Can you tell us about your experience in intelligent document processing and the potentials that this holds for financial services? So like every financial institution, we receive and process millions of documents per month. It's like obviously documents that might represent contracts, a document that might represent filings, documents that might represent interaction with clients or third parties, instructions, various forms of communication that are coming out from the firm to the external world. There could be reporting, financial reporting, but also like regulatory reporting, et cetera, et cetera. So we are like every financial institution in a very kind of knowledge-centric and document-centric business. And so the question that obviously the whole industry has been trying to tackle for many, many years is how to really bring digitization and knowledge management within a world that is dominated by a number of artifacts that are coming in and out of the organization. So the very first thing that obviously we did, like using machine learning, not necessarily generally generative AI, to do the task of translating documents in machine-readable structures or models in a way that a computer could actually understand and process and reason. And so machine learning is used to do what's called entity extraction, i.e. you're identifying what an entity looks like and an entity could be an address or an entity could be a dollar value or entity could be a type of corporation, et cetera, or a symbol, stock symbol. Those entities then gets mapped into a vocabulary that is generally extensible. And then natural language processing kind of takes those outside of an English context into something that could look like, for example, a JSON file that then a computer can reason about. And the same way as natural language understanding for, you know, when you talk to a dictation system has, you know, had decades of evolution and then, you know, getting closer and closer to like almost 100 % accuracy. The same journey has been done on the, you know, on the kind of a natural language understanding. So the question is really like, what is new now and what are the inflection points? First of all, I have to say that in order to make full advantage of the capabilities of AI and also obviously of generative AI, there is work that needs to be done in order for your data to be ready for that kind of consumption. So we've been working for years on our data platform, which we call Alloy, and internally and externally has been open sourced as Legend. And Legend is really a way to have a standardized way to describe entities, to describe how to represent those entities, like financial entities. Like I said, like a company name or a stock symbol or a type of asset or a derivative and so forth. and you model them in a way that is consistent and coherent, you're able to get the lineage of the data, you're able to put governance around that data. That makes you ready as an organization to then be able to deploy AI to reason about that data, which can be in the form of a document, it can be in the form of a database. And then once you're able to reason about that data, then you're also able to consume that data or those documents in different formats. So I think when it comes to document processing and document-related workflow and generally knowledge management, it's really important to really understand the capabilities and the limits of generative AI. Because generative AI can give you a step change with regards to the understanding of what the user actually wants and tries to express in a natural language way. But then the drawback is that sometimes AIs, because of the nature of kind of statistical completion of a sentence and so forth, they might be prone to errors. So when it comes to document management, actually errors are like your worst enemy. And so there are techniques that the industry has kind of come up with that we've been kind of, you know, like really like focusing on to make sure that the integrity and then kind of the precision and the accuracy of the information is preserved. And so when it comes to document processing, the idea of kind of, you know, like taking documents and putting them into data sources that are accessible to the AI as a way to understand your prompt better, but also a way to validate the information that is generated. That is kind of the first step. And so you're putting the documents into a vector database. That's how they're called. And it's called vector database because you're basically putting concepts in the form of vectors. And those vectors can be used to understand the similarity between concepts, which is like the distance between. And then you have retrieval augmented generation, which is very interesting, where you're basically taking the prompt. You understand what the question is, and then the AI figures out after it's been instructed to do so, what are the repositories of documents that we can take the actual document, you put it before the prompt, and then you're actually prefacing the prompt with the question, given this document, can you answer this question? And we see that the accuracy of the answer by doing this retrieval augmented generation goes up dramatically. One thing you can do also within the model itself, you can also spawn an agent through either a line chain or through function calling that actually calls out and validates an output based on the confidence that the AI has of the answer that is generated. And so if you take those techniques together, also augmented by the fact that you train your engineers, you train your business users to write the right prompts, then at that point, you can envision a future where document management and information management really becomes a very organic mix between human capabilities and machine capabilities. You see a future where manual steps or guesswork or very complex manual processes are actually going to be reduced tremendously. And at the same time, you're not sacrificing at all. In fact, you're actually increasing the quality of the output and increasing the precision of the information. And I think that's really where kind of we're all aiming to this perfect, you know, like ideal intersection where you have increases of productivity and decrease of errors. That is kind of the magic, the magic intersection. And I think this is kind of where the world is going. And those who can actually take advantage of that, I think they will see tremendous benefit on the way they run their organization, both from an efficiency perspective, but also from an accuracy and quality of the information perspective. Yes. And I think everything you're saying here just about intelligent document processing is worth putting into the larger context of everything we were talking about in our AI and business episode, our episode one in terms of tech debt and updating legacy systems. We know that large language models will have a huge place, not only in being a reason that we need to think about data governance, but also helping keep those data governance workflows and everything in check. It seems as though even kind of the larger point that you're making is that in terms of document processing, we now are encountering this moment with generative AI where we can get folks, no matter what their background in life is, to look at documents in terms of this is just a portal to data. This isn't even a page or a contract. In fact, this helps us look at paper-based information in the way of an engineer being very similar in the same way. And I mean, a software engineer to be a little bit more specific. But their responsibilities vis-a -vis these systems are not unlike that of lawyers in that, you know, the large language models can take these documents. They can take this data. They can take that IT infrastructure. They can analyze it and they can give the exact amount of context to anybody who isn't a lawyer, isn't a software engineer and say, this is what the system needs. This is what we're seeing in this data that is mission critical for our organization because the system has been trained on what are the core business goals. And it can help basically translate whatever's in that document into a data context and really just kind of free it from where it is in terms of paper. Just want to make sure I'm getting that analogy correct, just in terms of not only in terms of contracts. Go ahead. Yeah, no, remember, I mean, when we talked about, you know, there is one of the things that generative AI seems to be doing really well is kind of bridging the gap of knowledge. You mentioned like, you know, a lawyer might have decades of experience in, you know, trying to understand, you know, what are the salient points of a legal document. The same way as a developer has like maybe decades of experience to understand what are the most critical parts of code. And then once you have developers that need to work on systems that need to kind of deal with legal documents, the question is, how do you bridge that gap? And I think AI could do a lot there because of the ability to kind of act as a sort of an adapter between different levels of specificity of knowledge. And so really being trained to point at what are the key concepts within data or within a document or within any construct around knowledge that are really salient. And if you think about like this machines are all about, you know, I know it's almost like a joke, but it's all about attention. This kind of the key part of the transformer architecture. The world of knowledge is not a flat world where everything has the same importance and the same weight. It would be like saying, hey, you now have a legal document where kind of every word is the same. It's not. An experienced lawyer will know exactly what to pinpoint. And, you know, I have this sentence that I like to say that, you know, it's kind of how do you measure your return on attention? If you have a certain amount of time, where do you actually focus your attention? What's your return of attention? And I think this ability to kind of pinpoint, which is kind of really foundational on how these models and transformers are built, of understanding that not everything, not every word is created equal. Not every word has the same meaning and the same importance, but really trying to point, you know, the reader to, hey, this term and this particular expression and this particular covenant is actually fundamental to the way the deal is constructed. I think is a very interesting, you know, potentially transforming capability, which is, by the way, if you kind of flip it the other way in the world of developers, which, you know, obviously is the world that I come from. We have examples of developers really asking the AI, you know, to pinpoint what are the areas of this code that might not scare. Or, for example, the areas where, you know, there might be potentially some edge conditions that could cause errors or where it could be the bugs. And this extreme kind of ability or uncanny ability to summarize, to pinpoint, to be, you know, kind of focusing the attention to a specific part of the code is something that, you know, obviously makes the developer or the knowledge worker much more effective. Because at the end, it's really about that is really what you learn from experience. From experience, you really know how to maximize your return of attention. You become a much more efficient knowledge machine. And that, I think, this time that it takes to kind of get to that level of experience potentially could be shrunk, you know, if you have an AI companion that can actually kind of point to you, what are the things that you really need to pay attention to? Right, absolutely. And even, let's just take a step back even from that position, knowing that, you know, I think we've spent these last two episodes just establishing that, you know, generative AI and large language models will bring down the barriers of knowledge between experts and lower skilled workers and customers. Let's even take this. We've been talking about lawyers. We've been talking about software engineers, but let's even just take financial services professionals. Their expertise is something that they're trying to get across to their customers, to human conversation agents in call centers in order to facilitate those customer relationships. And there's we run into the same sort of black box problems there, right? They know what words are not created equal or what ideas are not created equal in the context of financial services. And that's usually nine times out of 10 what a customer is calling about when they call a call center. They want to know what that is or usually their problem, whatever it might be, revolving around what they don't know. So just in terms of those walls coming down and financial services leaders thinking of their own expertise. How do you think of black box challenges when it specifically comes to difficult financial concepts and how that needs to be translated for other members of the organization in order to facilitate those relationships? How should business leaders think of those relationships as these tools become increasingly more in their hands? I think you use the word relationship, which is obviously a big part of what financial companies do. It is, you know, in many cases, a business of advisory, it's a business of relationship, it's a business of really value add with regards to, you know, what could be our client's strategy with regards to either their investments or their, you know, M &A or, you know, their ability to raise capital or, for example, how you integrate an acquisition that you just made, et cetera, et cetera. And think about, you know, people that are in a relationship position, for example, bankers, that they are exposed, like everybody else, to a very large number of inputs, a very dynamic environment with regards to macro factors, with regards to microeconomic factors, with regards to information that ends up in a public filing or, for example, corporate actions. And there's so many things that are kind of hitting people that then need to kind of develop the ability to kind of really filter those and map them to who those are more relevant to. Right. Imagine a future where a banker could wake up in the morning, something just happened during the night, somewhere, maybe in another market, and an AI could actually email this person and say, hey, this just happened. It might have repercussions, for example, on people that own this particular position or thinking about doing the following investments or on this particular portfolio, investing in particular portfolios. These are the people, these are your clients, maybe that you might want to call. And these are the speaking points, back to the summarization example. And so it's a little bit, it's almost like the shift in relationship between non-targeted and targeted advertising that happened several years ago, where I used to see a lot of ads about shoes that I never buy. And now, you know, I see a lot of ads about guitars that maybe I have, or, you know, or maybe running shoes because I like to run that, you know, I might have a higher chance to buy. And then you feel that the return of your attention as a customer is actually respected and is actually kind of, you know, increasing. And I think that will happen in the world of human relationships and also like banking relationships where the customer expectations and the client's expectations in terms of relevancy of the interaction and also the ability for our people and for people that are in a relationship position to really kind of have the world in their fingertips in a way. and having, you know, like an AI that is kind of, you know, really acting as this targeting and filtering and augmentation engine, which will allow them to, you know, kind of engage with customers in an incredibly relevant way. And I think that is something that I think it will kind of change the nature of relationships. It's something that in a way we raise the bar with regards to, you know, the way you conduct business. and those who are able to leverage that, like it always happened when you raise the bar in terms of service level expectations, will be the winners. And I think this is one of the problem spaces that we're really thinking a lot about. Yeah, I really appreciate in your last answer that it's not all about making customers, putting them through the business program at Harvard. No, it's about what are they investing their attention in and how can you make good on that return on investment? And I think even this theme that you've introduced in the last few episodes of kind of that return on attention investment, that's going to be the difference in terms of how they approach privacy. Customers are going to be much more open with their private data if they're seeing a return on attention investment. So, Marco, I don't know if you can head down to the patent office or maybe call a publisher to get that copy written, that return on attention investment slogan. But I think, oh, man, if you've got a book published with that in the title in the next few years, I think it's such a good way of looking at it. Thank you. That I think, heck, I might use it as at least a title for this episode going forward before you copyright. Please do. Anyway, thank you so much, Marco. Really appreciate you being on the program this week. Thank you, Matt. It was great to be here. Thank you. And if you're interested in hearing more of Marco's perspective, don't forget to check out his appearance on our sister podcast. That's the AI in Business podcast, also on Emerge's podcast platform. In that episode, Marco is pulling apart where banking leaders should be showing the most caution when it comes to emerging AI capabilities and how financial services professionals should think of cultivating their skills for the future, given the impact of new generative AI capabilities. On behalf of Daniel and the entire team here at Emerge, thanks so much for joining us today, and we'll catch you next time on the AI and Financial Services podcast. you