Track 2: Conversational AI: Future impact and considerations in wealth

Recent advancements in artificial intelligence, powered OpenAI and large language models present incredible opportunities for industries like wealth management. By leveraging these tools, wealth management firms and professionals can offload tedious, repetitive tasks to these solutions in favor of fostering meaningful client relationships. Dr. Sindhu Joseph, co-founder and CEO of CogniCor, a Morgan Stanley-backed provider of AI-enabled digital assistants and business automation platforms for the financial services industry, will explore the growing role of AI within wealth management, the risks and limitations of today's solutions and future impact of this technology across this space.

Transcript :

Justin Mack (00:06):

You might remember me from such events as the last discussion on AI, I had just this morning. And as we know, developments in artificial intelligence, open AI, large language models, presenting credible opportunities for our industry and so many others, but there is a lot to keep in mind when these fast developing technologies start to clash or slam into the work that we do every day and have been doing a certain way for a very, very long time. So with that, it is my honor and my pleasure to present someone who I am very excited who's going to share her insights on that topic. Dr. Sindhu Joseph, Co-founder and CEO of CogniCor, a Morgan Stanley back provider of AI enabled digital assistance and business automation platforms for the financial services industry. Dr. Joseph, holds a PhD in AI, Co-author of six patents. And quite candidly, I am simply a fan in my opportunities to talk with her about this topic, her thoughtfulness, her compassion, her care, her consideration for how we are working AI into the work that we do. And almost that again, the PhD she holds, it's almost a outsider view of how the wealth management industry is going to weave this very impactful technology into our world. I am very excited for personally to hear her insights on this because we know it changes fast. By the time I get done talking, there will be new applications for large language models that will have to be thought of as well. So please join me in welcoming Dr. Sindhu Joseph to the stage. Dr. Joseph, the floor is yours.

Dr. Sindhu Joseph (01:39):

Thank you everyone. And I am sorry for the tech glitch. I am not able to press in my slides. So we will go old school for the new developments in AI.

(01:57)

As you know, Jason probably made it sound like I am half mission half human. So I just wanted to kind of present my human side of it a little more. And given that this is advisory business, relationship-based business, I want to really connect with you. I am a mother of three beautiful children. I lived in India, born and brought up in India and spent more than a couple of, more than a decade in Barcelona, one of the most amazing cities in the world and then moved to San Francisco around five years back. So that's my journey. I started CogniCor with my husband in 2018. When I finished my PhD, I had lots of job offers from large tech firms like Google, Facebook, you name it. I was offered really nice paychecks and jobs. And my point is I took one of the probably most illiterate financial decision in my life at that time I was a dreamer and wanted to change the world and went for really financially insecure future.

(03:09)

So I am a good candidate for all of your financial advisors here to kind of change the course of my financial life. So my focus was like because I was a financially illiterate, one of the reasons why I was not really focused on creating my wealth or really managing my wealth was not just because I was a dreamer and wanted to change the world, but also because I was really intimidated by it. So I am used to the buy experience from Amazon or hiring an Uber that kind of an experience, ubiquitous, personalized experience. But when it comes to wealth, we do, it's not because we don't have money, but really engaging with that and managing that experience with wealth becomes very intimidating. So I believe the reason Coco was there is really because by creating that seamless experience, which is not there in this industry today, I believe that you know can really democratize access to wealth and creation and expansion of wealth.

(04:21)

So that's really that what I care about, I care about, I care deeply about local communities, sustainability, as well as what makes us human. It's not just the AI part, what makes us human, and that really led me to exploring AI as a child. It maybe it's hard to believe, but I was looking at how humans have intelligence, how do we take decisions the way we do? And I was fascinated by AI because I realized using machines you can really try and create intelligence and understand human intelligence better. So I fumbled up on a theory of cognitive called theory of coherence, which actually makes decisions, how we make decisions, we put in different beliefs, actions in our mind, and then when we hear something new, we try and maximize our coherence to this new information. And if we are able to maximize it, we accept that new information.

(05:28)

If we are not maximizing it, we reject that information. So that's one of the ways humans make decisions. In my PhD thesis, I tried and kind of compute made it computationally possible to create an AI system that is able to use this theory, theory of coherence, which is a completely human cognitive theory. And with that missions could make actual decisions without we really forcing them to make those decisions. So one of my experiment was for a mission to actually decide to violate traffic signals because it has to get to an AIrport faster. So these are things that humans do, but missions, it becomes really, really difficult for missions to do because we are giving the rules for missions to act. But if mission has to decide something autonomously, so that's where I believe the intelligence comes in or that's what separates human intelligence with an artificial intelligence.

(06:34)

So I was trying to create those, all of this really background on how I think about AI and how we can really coexist with an AI system and our own human intelligence. So basically the two key things that I want you to take out of this session is AI is disrupting the whole world. We all know that, and I probably don't need to talk about it more because you have heard throughout this two days that how you can use AI to maximize your productivity. So my question is how do you know wealth managers? How do this industry take advantage of that and why should we take advantage? So that is the first thing that I want to drive home, which is really talked about. I just want to summarize that. And the second part is what exactly can be done once we all agree that AI has to be adopted, how can we do that?

(07:46)

So that is the second part. So these are the two things that I want to drive home in this conversation. So from why part of it, we all heard that around 73 trillion AUM is changing hands from baby boomers to millionaires. So I am not taking the standoff, we have to support the millennial clients, but I am taking the standoff. These millennial clients are going to demand from the wealth industry and experience shift because they are so used to experiences that are different or that are seamless. So an experience shift is the first thing that we as wealth industry has to bring into the equation. So that's one. Second is we have reducing margins as financial advisors kind of every time the fee is getting reduced as. And also in the last session we talked about retiring advisors. So there are a lot of retiring populations.

(08:52)

So there is a reducing workforce and there are reducing margins. So what comes out of this is increasing efficiency. So unless we increase efficiency, it becomes really hard for us to even survive in this industry. So there are two things that I really want to emphasize. One is unless we shift the experience, we are not going to survive as an industry and unless we create scalable efficiency, we are not going to survive as an industry. So these are the two things really why we should really care to create a revolutionary experience as well as scalable efficiency in the wealth industry. So going a little deeper into why we are not able to create these efficiency and experience shift, we all know that there are millions of regulated manual tasks that is different from other industries that is handled in the wealth space. Because of that there is revenue loss, there is lot of work that falls into the hands of the advisors as well as the home office staff.

(10:03)

Finally, all of this translates to experience broken experience for clients and these are the things that we really want to change in the industry. So with AI, I believe that both the experience as well as efficiencies can be shifted in the industry. So I am just focusing on these two things. A friend and experience change, something as seamless as Alexa switch on my kitchen light. So you should be able to interact with your wealth firms wealth operations as seamlessly as that just provide a conversational command, conversational interaction and the system should be able to take over. The second is creating efficiency. So it's not just the friend and experience that is needed, but behind the scenes you need to build that efficiency throughout and create that automation. So with automation and AI together, AI is basically creating that experience shift and automation is creating that scalable efficiency.

(11:12)

So with these two things, you can actually create a revolutionized wealth industry. And I believe this is a huge opportunity that kind of fell into the lap of the wealth industry today because AI has revolutionized itself in terms of large language models and generative models, which has huge capability now. And all of these things together we have this opportunity that we have to seize. If we don't do that, the industries, and I am not talking about it, is a nice to have, but it is a must have that we should take action now. If not, there is going to be an outside disruptor that is going to come in and disrupt the industry. So that's the why part of it, why we should care. I think it's very, very clear both experience and efficiency shift. So the second part of my talk, I just wanted to talk about AI. What is AI and how can we adopt AI in the wealth industry? Any questions before that because I am not showing slides, I am not sure how much are you getting? Yeah, yeah.

Audience Member 1 (12:23):

Oh, okay. I am pretty loud. So I used to work at Morgan Stanley, so I know that you mentioned that you're partnered with them. I guess key question here is there is so many things you could do use case wise with AI from within the wealth industry, client onboarding to rebalancing portfolios. Any perspective before you transition to your next piece on how firms are looking to apply AI across a plethora of areas?

Dr. Sindhu Joseph (12:49):

I will be coming to that. So I think I'll part that question for later. So I just wanted to give a very broad overview of AI itself. I am sure most of you have some kind of understanding, but as I was initially sharing, my intent with AI, playing with AI was really to create recreate intelligence, an attempt to recreate intelligence. So the community of AI, the real intent is to create intelligence systems design and build intelligence systems that is able to communicate and understand human language and then take certain tangible actions that is communicated through this language. So one is the ability to communicate with humans and understand human language. And then once that understanding is there the ability to take certain actions which need normally human intelligence, but the missions is able to carry out those actions. So these are the two things that in a very broad sense what AI systems are kind of trying to achieve.

(14:00)

So in terms of the machine learning, the AI community as a community, we can attempt to reach this goal in many ways. One of the most successful ways of reaching AI intelligence is through machine learning. So you all probably have heard of machine learning. So it's basically a very glorified pattern matching system, which is learning from examples. You provide a lot of data to missions. For example, if you want to teach missions how to identify a cat, then you provide examples of cats. And what the mission is trying to do is understand and map certain key features, maybe four or five, it has four posts, one tail, two eyes, so those kinds of features and then map it to certain labels, in this case the label iSCAT So how do you make sure that these features are correctly mapped to this particular label of a cat with no error, absolutely no error or the minimum error.

(15:15)

So that is basically all about mission learning. You learn from different examples. So it's kind of a ground up learning, which humans are very good at. And if you ask a child how can you recognize a cat by showing a three or four examples, she would immediately do that by again, just focusing on these four or five features and then mapping to the human representation of that concept, which is a language cat. So that is what we are doing, but that is about mission learning. If you look at humans, we are also very, very capable of another kind of learning, which is the opposite kind of learning. It is called deductive reasoning. So if you ask the same child, can Jason fly for example, she would immediately say, no, Jason cannot fly. Why? Because Jason is a human and humans don't fly so Jason don't fly.

(16:18)

So it is kind of a opposite kind of learning where before we were looking at lot of ground level examples and then trying to create certain general rules in this time we have certain general rules which are like humans don't fly and then we are assigning a sample to this particular set. So that is called deductive reasoning. Missions are not so good at it and the AI community is still kind of struggling to figure out how, you know can create AI system that is able to understand these kind of deductive reasoning. So when this is all interesting in the context of if you are a buyer, a solution buyer in your firm, how do you evaluate this kind of AI systems? You should know that when you are looking to buy these systems, the kind of learning is by providing examples. So the previous generation of AI systems, you had to train all of your AI systems giving lot of examples.

(17:28)

And that's the reason you know, probably if you adopted AI in your organization, you spent enormous time training the AI system to make it learn certain domain information. What happened after that is the generative AI, which is the large language. The only difference is it is a machine learning model, which is like what I explained, learning from examples, the difference, the crucial difference in large language models or what we call the chat GPT kind of models is that it came pre-trained with around 45 terabytes of data, which is pretty much the whole of a substantial part of the internet as well as around one 75 billion parameters. Which means that you know, actually trained the mission with pretty much every domain information that you can imagine. So it came pre-trained, which solved a crucial problem in AI deployment. AI adoption in industries which is like you know, as a company did need not sit and train the mission to understand either the human language or certain task.

(18:47)

So along with what was very interesting with chat GPT was that along with its capability to understand language and generate language, which is basically a language generation model, it also had something called instruct GPT, which was ability to understand human instructions. So it's basically trained on just the human instructions, how do you understand human instructions? So combining this, if you asked it to generate a blog post, then it would understand the instruction and then complete that instruction with the generation of language. So that is what it basically did. So I am not going into more detAIls of it. So it basically made the AI adoption much more easy, much more general purpose, much more mainstream. And also the other thing that you could do was you don't need a AI expert like me to deploy this kind of systems. You know, could basically have any of your domain experts could deploy these systems because the interaction with these systems became using prompts, which are things like in very loosely kind of coupled set of structured natural language system, generate a blog post or write an email.

(20:12)

So those kinds of instructions were very natural language based instructions that you could kind of do and generate, make the system do that job. So some examples of how this was used today, you could simply say write an email representing a financial advisor to share with her clients who are nearing retirement given the market turmoil and maybe a Silicon Valley bank collapse. So you could simply state this as a prompt. I had actually wanted to show how the answer response was generated, but since I couldn't do use that, so you can actually see that the system was able to generate a very personalized email to their client that represents all of these things. So you could actually use this conversational interaction to actually make things happen. Another example is like, you know could say, summarize the key points in a Morgan Stanley research report and then it would come up with 10 or 15 kind of research points from the report.

(21:28)

Basically read all of the report and summarize the research findings. So in essence what really became is AI became the new UI. So you are all in developed industry. If you are an advisor, you are very used to a very cluttered dashboard with several menus and sub menus and things like that. So come a conversational AI, you can really have a very blank state and just a conversational interaction, which is very seamless for humans. We are very, very used to having this kind of a conversational interaction with humans to get things done. If you have an assistant, you would simply ask the assistant to get certain things done in a conversational language. It is, as opposed to our strategy today, which is finding information. If you open the dashboard, if you want to place a trade instruction, you click five menu clicks and then finally find the link to place a trade instruction.

(22:39)

You don't have to really do that. You can simply say, I want to place a trade instruction and it would do the rest for you. So AI became the new conversational UI. Likewise, if you want to change beneficiary for a client, you could simply say, I want to change beneficiary for Alex. And then it would basically take that instruction and maybe find the right form, fill out the form, submit it to your custodial platforms. So all of this behind the scenes automation is powered by a conversational interaction system. So imagine your new advisor platform with a Google front page where you have just one single place to enter a prompt and that's basically it. So that is really how the seamless interaction, seamless experience should look like and that simplicity will bring access to wealth for more people. So that's basically what I wanted to present here to your question.

(23:46)

Previously, I just wanted to highlight some of the opportunities in the wealth space in very specific areas. One is of course operations. I think we did talk a lot of about it. You spent as an advisor maybe enormous amount of time in document management, preparing client review meetings, form filling and things like that. So all of those things can be, again, imagine that these things can be driven from a conversational friend and saying that I want to get this done, I want to start an RMD distribution for a client or things like that. And then the right form comes up to enable that. Likewise, business development lead generation. Imagine in the previous conversation you were talking about finding leads. Imagine the AI system is able to recommend leads based on your client profile and then go out and find those profiles, which is probably one degree of separation from your actual client base.

(24:53)

And then recommend those leads generate every single morning three leads for you. So that those are lead generation capabilities. Another one is client engagement, which is where we can really, really expand. How can we create, learn from market research very, very quickly and then share that market research with clients as well as kind of emails, generating emails and investment insights. Those kinds of things can be automated very easily and taken advantage of by the advisor. So all of these routine tasks the advisors spent time on can be automated using AI and having a conversational front end interaction. So I just wanted to highlight in terms of from a wealth stack perspective, and I am not asking, if you look at Michael Kite's technology map, there is maybe at least 500 firms that is scattering to this industry in terms of providing platforms. So AI should not be another platform in this huge map and it should not stay separate.

(26:13)

I am not advocating for a new platform. AI should be invisible in the sense that it should merge into these platforms. So anything that a vendor is selling AI, it should be part of the tech platform that you already have. So it should be able to embed into those platforms and then integrate nicely and pull data and push data seamlessly across. Finally, I also wanted to talk about a little bit of ethical considerations. How do you make sure the client data is protected and you have accountability and transparency and no bias. So things like explainable AI, which is not just generative AI, which is very hard to dive into, but you can combine that with things like knowledge graphs, which provides an explanation into why the AI system is doing what it's doing. So those kinds of technologies will really give that traceability for you. Likewise, you know, have open AI systems, which is the chat GPT models where your data is used to trAIn that system, which means that your data is not predicted or closed models where your client data will stay within those boundaries. So you should basically look at what is the difference. So when you are using an open chat, GPT, you know should be really terrified in terms of how my client data or how my data is training that AI system. So you should go for that closed model.

(27:55)

So that's pretty much what I wanted to share. I want to leave you with that thought that AI will not replace you anytime in the future, but people using AI will definitely will. So I want you to jump into that bandwagon. Thank you.

Audience Member 2 (28:20):

I just wanted to open up for any quick questions if there were any, I know we're going into the next session, but maybe one or two.

Justin Mack (28:27):

Alright, one second. Excuse me. Pardon me. My bad.

Audience Member 2 (28:33):

Thank you. And I have a two part question. So what do you think is the most limiting thing for AI right now? Is it more on the software we need to develop new models or is it more so on the hardware we need smaller transistors or things like that? And then the second part is if you were going into your PhD now knowing how AI has developed, how would your experience have changed?

Dr. Sindhu Joseph (29:01):

Yeah, the first part is I think hardware is developing at the more slow is broken over and over again. So I don't see there is hardware limitations, but that is overcoming as fast as you can imagine. So just to give you an example, the chat GPT model was developed with billions of dollars spent in the hardware a supercomputer. The next model was developed with 200 million and the next model, which is equally powerful as Chat GPT Aklon of it was developed with $30. So that is the, and within a week. So that's kind of the power of the hardware that is getting cheaper and cheaper and the software getting faster. In terms of if I went into PhD today, I would be looking at completely different things. For me, holy grail of AI was if we can understand natural language as spoken by humans, I think to a certain extent that is achieved today, although it's a perception of understanding, it is a simulation of understanding, not real understanding as humans map it to. So that problem to certain extent has been solved. So I would go more into how can it create this kind of deductive reasoning, which is very hard. Even today. We are able to store very few information in our brain and then still able to take decisions because we have this deductive capability. Machines don't have that. I wish we are able to create those kind of systems.

Justin Mack (30:56):

All right, well that is our time. So again, another round of applause for Dr. Sindhu Joseph.