WEBVTT
1
00:00:04.160 --> 00:00:29.109
Laura Smith: Okay. Hello, everyone. Welcome to our webinar today, hosted by Xpero and Corporate Insight, AI Adoption and Financial Services. We're excited to have you join us today. Just a couple of quick notes before we begin. The session is being recorded, so you will receive a link to the recording and a follow-up email this week. And also, if you have any questions or have any issues during the webinar, feel free to message me, the host, Lori.
2
00:00:29.110 --> 00:00:31.889
Laura Smith: I will be happy to assist you.
3
00:00:31.890 --> 00:00:56.870
Laura Smith: At the end, we will have an open Q&A. So if you have any questions, feel free to drop them in the Q&A box at the bottom of your screen. If not, and you want to send a follow-up email, we'll be happy to answer any questions that way as well. So with that, I'll go ahead and pass this over to our speakers today. We have Sebastian Good, the CEO of Expiro, Julie Inglesby, principal designer at Expiro, and Justin
4
00:00:56.870 --> 00:01:01.119
Laura Smith: Tudor, Director of Market Engagement and Thought Leadership at Corporate Insight.
5
00:01:02.410 --> 00:01:15.600
Sebastian Good: Thanks, Laura. Excited to be here to talk about AI in financial services. Today, our agenda is fairly straightforward. We're going to do a little introduction of the firms and what we do in the market and how we contribute to this conversation.
6
00:01:15.610 --> 00:01:27.899
Sebastian Good: And then we're gonna introduce Corporate Insight's, framework around thinking about, the different activities, that AIs participate in in, financial services, inform, analyze, advise, and act.
7
00:01:27.900 --> 00:01:38.080
Sebastian Good: And then if there's time and interest, do some Q&A. So I'm going to start by handing it over to Justin at Corporate Insight to talk a little bit about what they do.
8
00:01:39.910 --> 00:01:45.510
Justin Suter: Thanks, Sebastian. Happy to be here. Thanks, everyone, for taking some time out of your day to listen to us.
9
00:01:45.700 --> 00:01:58.299
Justin Suter: Corporate Insight, we've been around for over 30 years, as you can see, very proud of that. And our secret sauce is that we leverage a panel of users who have accounts at the institutions
10
00:01:58.300 --> 00:02:08.930
Justin Suter: and the insurance companies, and the retirement plans, etc, that we are evaluating on a daily basis, and we're leveraging that access to go behind the login
11
00:02:08.930 --> 00:02:13.980
Justin Suter: and track how the digital customer experience is evolving over time.
12
00:02:13.980 --> 00:02:21.959
Justin Suter: So, of course, that has positioned us perfectly to be on top of how different firms are starting to roll out
13
00:02:22.110 --> 00:02:26.749
Justin Suter: consumer facing AI implications in this brand new world that we're living in.
14
00:02:27.240 --> 00:02:45.389
Justin Suter: And it's always important to keep in mind that the competitive edge that we're bringing to the table is that key lens, right? We become the customer on behalf of our clients, so that is our lens. What does a customer see when they're logging into their account or their policy or their profile, and what can they do once they're there?
15
00:02:46.400 --> 00:03:02.189
Justin Suter: Some of the challenges that CI helps clients solve, of course, always, again, thinking about this from a customer experience perspective, is visibility. A lot of firms have limited insight into what competitors are actually doing, again, behind that login, behind that authenticated wall.
16
00:03:02.380 --> 00:03:14.000
Justin Suter: So we're able to solve that through our panel, as I talked about. Other challenge that a lot of our clients struggle with is just the velocity or the pace of change, how quickly digital platforms are evolving now, and…
17
00:03:14.000 --> 00:03:25.570
Justin Suter: Of course, in concert with that is how quickly customer expectations are shifting alongside them, and because of our continuous monitoring of how the platforms are changing, we're well-positioned to comment and advise on that.
18
00:03:26.960 --> 00:03:33.549
Justin Suter: Third, we're looking, we're objective. We're a third party. Internally, things can get very
19
00:03:33.950 --> 00:03:43.849
Justin Suter: friction oriented about what's good or bad or most useful to clients. But there's no substitute for that outside unbiased perspective. And that's the role that corporate insight plays.
20
00:03:44.600 --> 00:03:56.300
Justin Suter: And a lot of firms also just don't have the capacity to be keeping track of all these different developments, as I talked about. So that's where you can rely on us to be an extension of your team.
21
00:03:56.590 --> 00:04:03.190
Justin Suter: Taking care of all of that noise, and allowing your internal teams to really focus on execution instead of monitoring.
22
00:04:03.220 --> 00:04:11.529
Justin Suter: And last but not least, within this, again, fast-changing world, how do you know what to prioritize first? That's another area where we're a great resource because
23
00:04:11.530 --> 00:04:31.040
Justin Suter: When everything looks so urgent, it can be hard to know where to focus. But our recommendations are backed with our own expertise in the space, as well as survey data with customers getting their opinions on what matters most to them, as well as in-depth interviews with customers as well. So it's all these different perspectives to help you understand where you can get the most bang for your buck.
24
00:04:31.040 --> 00:04:34.480
Justin Suter: When you are building out your developmental roadmap.
25
00:04:36.180 --> 00:04:53.410
Sebastian Good: Awesome. Thanks, Justin. And I'll just say, as someone who sees a lot of these updates, it's relentless. It's a new feature from this brokerage today, and an insight on the digital experience at an insurance company yesterday, and today it was the checkout button at Disney not working properly.
26
00:04:53.410 --> 00:04:56.760
Sebastian Good: It's really a really in-depth stuff.
27
00:04:56.800 --> 00:05:16.019
Sebastian Good: I'll talk a little bit about what we do here at Expiro. We're a 25-year-old company this fall, old enough to drive, I guess, is the joke we're supposed to make. We focus on building the front ends that power finance, and increasingly these days, as AI redefines everything about how we work, how the AI is transforming
28
00:05:16.020 --> 00:05:24.990
Sebastian Good: those front ends. We help people across the institutional all the way down to the retail side of the world.
29
00:05:24.990 --> 00:05:25.789
Sebastian Good: But…
30
00:05:25.850 --> 00:05:43.550
Sebastian Good: deliver digital experiences and products that deliver investment insight. So, we have, built, and modernized, and delivered portfolio management, trading, and investment management, systems at very large, asset managers, asset servicers. We help asset…
31
00:05:43.780 --> 00:05:58.250
Sebastian Good: managers distribute tools to their advisors and their public to help them understand their models and their funds. And we work heavily in the area of what we call digital advice, which is the next generation of how a lot of digital wealth
32
00:05:58.250 --> 00:06:05.160
Sebastian Good: retail brokerages, advisor-facing, experiences, are delivered. So,
33
00:06:05.180 --> 00:06:22.619
Sebastian Good: working together with Corporate Insight is really exciting, because they help us understand how the industry is moving as a whole, and I think we're delivering a lot of the, sort of, what's next components into these type of experiences, seeing how people think about investment across all kinds of different personas.
34
00:06:22.620 --> 00:06:44.029
Sebastian Good: The way we deliver is as a strategy advisor and as a delivery advisor. So some of the experiences you'll hear us talk about today are things we've delivered, either through a design engagement that someone like Julie can help with, with 20 years of experience in the industry, or on the technology side, actually building and delivering.
35
00:06:44.270 --> 00:06:52.979
Sebastian Good: The kind of work we do, as you saw, is everything from helping advisors make sense of investments.
36
00:06:53.040 --> 00:07:04.109
Sebastian Good: To, civilians understanding, what kind of funds they can invest in, helping investment plan advisors put together the best options for their, fund shelf.
37
00:07:04.110 --> 00:07:14.460
Sebastian Good: We're helping institutional traders at IPC plan their calls with artificial intelligence. We're transforming asset management workflows at Vanguard for portfolio managers.
38
00:07:14.480 --> 00:07:28.270
Sebastian Good: And helping investors understand, how their hedge funds work. So, really, across the board, invite you to come kind of read some of these case studies and see how we can help you. A key part of how we work is with partners, like Corporate Insight.
39
00:07:28.270 --> 00:07:36.580
Sebastian Good: But also working with a lot of technology and sort of finance ecosystem providers to provide experiences in the industry.
40
00:07:36.580 --> 00:07:46.889
Sebastian Good: And the conversation we'll have today is informed by our products that we accelerate this delivery with. So, from portfolio monitoring, to goal tracking, to investment screening, there's a lot of
41
00:07:46.890 --> 00:08:09.530
Sebastian Good: experiences that we deliver to our customers where there's a core component based on our framework that's then customized for our customers. So that's a little bit of background about what we do and how we deliver it. And so I'll turn back now to Justin to talk a little bit about this framework we're going to talk about together about how AI should participate in our user experiences.
42
00:08:11.040 --> 00:08:13.550
Justin Suter: Thanks, Sebastian. So we now we've got the,
43
00:08:13.550 --> 00:08:33.579
Justin Suter: Housekeeping out of the way and we can get into what everyone's really here for, right? The AI, you can't escape it. And so this is our backdrop for the conversation. I think you laid it up very nicely. And our most recent research is this report. We call it our AI and financial services today report. And.
44
00:08:33.590 --> 00:08:43.720
Justin Suter: It's been the foundation for how we're now approaching all of our conversations around AI with clients. We know that the AI deployment is accelerating both within our industries.
45
00:08:44.020 --> 00:08:48.589
Justin Suter: Other industries that we're not a part of, industries we don't even know exist yet.
46
00:08:49.010 --> 00:08:54.979
Justin Suter: But the biggest takeaway, or I should say one of the bigger takeaways, is that customer trust is not keeping pace with
47
00:08:55.030 --> 00:08:57.130
Justin Suter: how fast AI is being deployed.
48
00:08:57.130 --> 00:09:17.449
Justin Suter: So we tried to structure this report in a way to help firms understand how you can strategically leverage these first AI tools that you're beginning to roll out in a way to build up that consumer trust, build up that goodwill, and then hopefully get them to be engaged more with you as a firm and benefit from the tools that you're providing for them.
49
00:09:17.450 --> 00:09:22.059
Justin Suter: So some numbers on the right side of your screen to try to, you know, drive this point home.
50
00:09:22.060 --> 00:09:44.830
Justin Suter: You know, we see 65% of consumers afraid that an AI error is going to affect their money. 58% that abandon a chat with an AI bot because they thought they were talking with a real person and then they realized it was a bot. 45% think that AI is only going to be there to replace human support. So there's a real, real challenge right now.
51
00:09:44.870 --> 00:10:02.730
Justin Suter: Regardless of where a firm sits in the value chain in an industry of making sure that as you're rolling out AI, don't expect customers to be impressed with the press release or the bells and whistles. You have to make it clear of how is this going to benefit them.
52
00:10:03.340 --> 00:10:05.070
Justin Suter: And that understanding…
53
00:10:05.110 --> 00:10:18.649
Justin Suter: Without that understanding of customer expectations, firms are at a risk of creating AI experiences that are going to feel impersonal, actually reduce engagement, end up being unhelpful or inaccurate.
54
00:10:18.650 --> 00:10:30.860
Justin Suter: contribute to the anxiety that already is out there with a lot of consumers. And I think, you know, I just certainly have some anxiety myself when it comes to AI, so I'd love some experiences that can help erode that anxiety as well.
55
00:10:30.950 --> 00:10:31.609
Justin Suter: Oh, okay.
56
00:10:31.930 --> 00:10:34.879
Justin Suter: So the framework that we built out in this report.
57
00:10:35.830 --> 00:10:46.500
Justin Suter: is based on this research methodology. You can see it's 3 different primary sources, all of which I've already hinted out a little bit in talking about corporate insight.
58
00:10:46.860 --> 00:11:00.760
Justin Suter: But we have these monitors, we call them our monitors because we, as I mentioned, are monitoring all these different industries on a day-to-day basis, and that exposure has just allowed us to organically observe how these tools have been launched
59
00:11:00.760 --> 00:11:22.209
Justin Suter: and iterated on by the different firms that we're tracking. So that's the core piece of the report, but it's augmented by in-depth interviews that we conducted in January, conducted by our in-house UX research team, and that allows us to really go beyond just a surface level survey response and try to dig into the underlying emotions of
60
00:11:22.210 --> 00:11:30.399
Justin Suter: why someone will or will not engage with a tool or what they hope to get out of an experience with an AI support.
61
00:11:30.690 --> 00:11:43.939
Justin Suter: capability. And then the last piece, of course, we do have some quantitative survey data as well. You already just saw some of our big takeaway stats from that quantitative survey that ran in February. And again, we were really trying to get at, you know, how
62
00:11:43.940 --> 00:11:52.319
Justin Suter: Have you used it? Do you think that firms are using AI? And what would make you more comfortable about it? So we try our best to present those findings here.
63
00:11:52.320 --> 00:12:05.919
Justin Suter: And it all rolls into our four buckets of the different autonomy levels for what AI looks like. And it informs, analyzes, advises, and acts.
64
00:12:06.510 --> 00:12:19.690
Justin Suter: I want to be very direct about what this framework is and what it is not. It's an attempt to be a map of where the financial services industry at large is today, potentially where it's heading.
65
00:12:20.110 --> 00:12:37.549
Justin Suter: It is not in any way a leaderboard. A tool that is in the informs bucket is in no way inferior or superior to a tool that could be placed in the axe bucket, right? These are just designations on the autonomy level of these capabilities that different firms are
66
00:12:37.550 --> 00:12:46.279
Justin Suter: rolling out. And if you're reading along, you'll see a lot of that same language in that large bucket, the note on CI's framework.
67
00:12:46.280 --> 00:13:03.619
Justin Suter: But as you can see laid out on the slide, incumbent firms are largely proceeding with caution, which makes a lot of sense. Most of them are operating in those informs and analyzes buckets, because deterministic, more rule-based virtual assistants give them more control over the outputs, and they reduce the hallucination and compliance risk.
68
00:13:03.620 --> 00:13:08.619
Justin Suter: Where we have fintechs really showing where consumer-facing AI is evolving the quickest.
69
00:13:08.620 --> 00:13:27.569
Justin Suter: They're moving very quickly because they're not weighed down by some of those legacy systems or risk averse culture that's typical of a more established institutions. And we haven't seen incumbents match that aggressive deployment yet, but they're starting to catch up. We've seen some launches already. And as you'll see later on, we'll have some teasers on how this is developing.
70
00:13:30.630 --> 00:13:46.850
Justin Suter: So with that, we're going to move into our first example of the day. We'll have an example from each of these four buckets from our report, and starting in that informs designation, right? Robinhood Digest's probably on most people's radar already if they've started to poke around AI and investing.
71
00:13:47.710 --> 00:14:00.910
Justin Suter: But these digests appear as short 3- or 4-line summaries on the right side of a stock screen, and then they open into a dedicated screen with deeper write-ups on roughly 4 or so recent stories.
72
00:14:00.910 --> 00:14:20.570
Justin Suter: that are likely to move that asset's price. Each of the stories carries a timestamp and will prompt a refresh when new information lands. Users can also manually trigger a refresh themselves if they think that or they know that there's breaking news that they want reflected in these digests right away.
73
00:14:20.750 --> 00:14:27.749
Justin Suter: One thing I want to point out is how Digest handles sourcing, because it's a really good example of…
74
00:14:27.830 --> 00:14:51.910
Justin Suter: how this can be executed good and then less well side by side. So some of the Robin Hood digests will list sources with timestamps at the bottom. And if you even link right out to the full article where the source is coming from in an embedded browser, some of the other sources will only surface a headline link with no explanation. And that leaves it incumbent upon the user to try to verify the reporting behind that.
75
00:14:51.910 --> 00:14:53.290
Justin Suter: So there's actually some.
76
00:14:53.290 --> 00:15:04.300
Justin Suter: some really great and some, like I said, opportunities for improvement just within this singular feature alone, and that's pretty reflective of AI as a whole right now, where there's lots of,
77
00:15:04.560 --> 00:15:05.530
Justin Suter: Awesome, awesome.
78
00:15:05.700 --> 00:15:12.289
Justin Suter: And helpful features, but also lots of room for improvement, even in what were some of these best-in-class examples that we're calling out.
79
00:15:13.410 --> 00:15:20.959
Justin Suter: Robinhood does also handle disclosure as well. The digests are labeled AI-powered, not advice, verbatim.
80
00:15:20.960 --> 00:15:36.829
Justin Suter: across the board, both on the stock screen and across the dedicated methodology pages, so that AI usage is clearly disclosed throughout. And that's a great blueprint for firms that want to play in this INFORMS designation, specifically because of the combination between clear disclosure.
81
00:15:37.120 --> 00:15:38.600
Justin Suter: And the sourcing depth.
82
00:15:40.100 --> 00:15:58.159
Sebastian Good: I think, Justin, it's a really interesting reflection of something Julie and I talk about a lot, which is how much of the information architecture that we experience as investors is driven by data entitlements and licenses, right? Like, I can't speak to the actual data sources, you know, that Robinhood has, and what their
83
00:15:58.160 --> 00:16:11.609
Sebastian Good: license agreements are, but I'd be willing to bet in a lot of experiences like this, this is driven by the AI willingness of some of the underlying data vendors. You can show my headline, but you can't show my article, or you can use it for intelligence, but, you know, right?
84
00:16:11.610 --> 00:16:20.630
Sebastian Good: So this is a real challenge we face when we put solutions together is, yeah, what are you even allowed to do with it before you decide how smart you're going to be about doing it?
85
00:16:21.310 --> 00:16:38.499
Justin Suter: And that's a question that we can't answer sometimes, too. Sorry, Julie, but just trying to anticipate a potential Q&A question. It's like, hey, do you know what's underlying this? No, we don't, because that's our whole lens, right? Our lens is this is what the customer sees, and that's how we're able to report on some of this. But go ahead, Jul.
86
00:16:38.700 --> 00:16:46.299
Julie Ingalsbe: Yeah, I was just gonna say exactly what you said. We might not understand what that underlying piping is, but, you know, you can kind of tell that some of the UX
87
00:16:46.460 --> 00:16:49.720
Julie Ingalsbe: Shortfalls are probably driven by that, absolutely.
88
00:16:50.110 --> 00:17:08.300
Sebastian Good: Which, which users, you know, we're not gonna go explain that to them, and they may not, may not all appreciate. Yeah, so the way we're gonna do this format, is, we'll talk a little bit, from our perspective as a solution provider about some of the user experience considerations and some of the technology considerations.
89
00:17:08.300 --> 00:17:13.749
Sebastian Good: That we encounter when we're building our own components, and when we're designing solutions for our customers.
90
00:17:14.460 --> 00:17:39.460
Julie Ingalsbe: Yeah, we thought it would be great to kind of take some of the things that Justin is seeing, you know, from some of these big players in the industry and bring it back to what we're seeing in terms of our observations around the investor's journey, right? And we actually see Inform as one of the first like key segments or key buckets in that first part of that investor's journey of really trying to understand, you know, what's the market doing? What, you know, what are my stocks doing? What is my ETFs doing?
91
00:17:39.460 --> 00:18:01.820
Julie Ingalsbe: Like, what do I need to know? Like, tell me that information. Some example patterns that we have that we've executed is around digest summaries, market recaps, research highlights, portfolio news, like being able to take a large amount of data and trying to summarize it into something that's really clear for the user to understand.
92
00:18:01.820 --> 00:18:03.360
Julie Ingalsbe: Yeah.
93
00:18:03.360 --> 00:18:20.570
Julie Ingalsbe: In the upper right corner is this Compare Insights piece that we have. Instead of having a comparison tool that is just columns of data that I don't really understand, that has probably been segmented by the back end of how the data is organized.
94
00:18:20.570 --> 00:18:32.560
Julie Ingalsbe: We have what we're calling compare insights that kind of bubble up at the top of that experience, like, these are the differences, you know, these are the things that you should know between these two stocks or these two funds that you're looking at.
95
00:18:32.560 --> 00:18:45.189
Julie Ingalsbe: So those plain language summaries, kind of to one of Justin's point, making sure that we're linking back to the original content so that the user is informed that this is coming from these, you know, these particular sources.
96
00:18:45.190 --> 00:19:03.469
Julie Ingalsbe: At the end of the day, just to kind of echo what Justin is saying, we're trying to build trust, right? This is a new world for a lot of us and for our users. And whatever we can do to provide that trust so that they can feel like their money is safe, their investments are safe, and the decisions they're making are also safe.
97
00:19:05.100 --> 00:19:23.389
Sebastian Good: Yeah, and I'll speak a little bit about how technology has to work to make these experiences accurate. The first one I'll start with is you've got to start with the semantics of the underlying data. You need to understand what's important and why. And in a sense, this is like asking, hey.
98
00:19:23.390 --> 00:19:47.769
Sebastian Good: Analyst, if you were going to write up a sector performance update like you see in the middle here, or you're going to bring it into someone's portfolio, what process would you follow? If you don't have a clear understanding of that process and the data sources that would be used, your LLM, your AI is not going to do any better. These frontier models know a lot of things, but unless you find them where they know it the way you want to do it.
99
00:19:47.980 --> 00:20:11.919
Sebastian Good: You're just it's just going to be luck of the draw. So if I'm looking at how our sector is performing, I still have to think about how am I tracking that? Am I getting intraday ETFs? Am I am I looking at, you know, index providers? And then if you look below that sector performance graph, you know, what does it mean for my portfolio? I need to be able to pull portfolio holdings. And then I need to think about, am I doing a Brinson attribution? Am I just doing absolute?
100
00:20:11.920 --> 00:20:21.400
Sebastian Good: salute reporting, you know, what's my methodology here? And then, if I'm bringing in third-party, data, like an analyst report or a news article.
101
00:20:21.400 --> 00:20:30.709
Sebastian Good: Again, all of us have probably experienced the notion that you could pull up your Claude or your ChatGPT and just throw all those things in, and you might get a good result.
102
00:20:30.710 --> 00:20:43.119
Sebastian Good: hey, here's some sector performance for the day, here's the holdings my customer has, and here's, I don't know, 50 paragraphs of stuff people wrote about the market. And we've all seen how a lot of the time it looks pretty decent.
103
00:20:43.120 --> 00:21:02.469
Sebastian Good: But a lot of the time, pretty decent is not a good bar for an investment experience at a retail brokerage or an advisor. And so for us, it all starts with pulling that unstructured data apart and making it structured. If I have a news article that's saying some analysts, you know, estimates have changed.
104
00:21:02.750 --> 00:21:15.350
Sebastian Good: estimate change is a piece of semantics. If someone says semiconductors are on a rip, what does semiconductors mean? That means a sector with performance I can track. It's a family of stocks that I can attribute to. So…
105
00:21:15.380 --> 00:21:27.150
Sebastian Good: Having a clear underlying semantic model that attributes data to the providers that you subscribe to and to your internal data is absolutely critical.
106
00:21:27.360 --> 00:21:38.819
Sebastian Good: Being very explicit about how you take unstructured data like news and editorial and match it to structured data like sectors or issuers or portfolios or news themes is absolutely key.
107
00:21:38.820 --> 00:21:49.409
Sebastian Good: Because the summaries you go build, like you see on these screens, need to come from the structured data for you to have some confidence that they work. And this matters the deeper you get into
108
00:21:49.600 --> 00:21:57.979
Sebastian Good: Smaller issues, smaller sectors, more unusual clients. Most people can just say, how's Nvidia doing today? And the AI is going to do okay.
109
00:21:58.020 --> 00:22:16.870
Sebastian Good: But the more specific you get, the more unreliable that is. And then very much to Justin's point, if you don't have an information architecture that traces the citations through that flow, you're not going to be able to explain it to the user and earn their trust. And if anyone's tried to program these things, they know.
110
00:22:16.990 --> 00:22:29.589
Sebastian Good: Asking an LLM how it came up with its answer is just an opportunity for it to tell you more things that you want to hear. It's not always right. We have to track what we've put into that LLM context and see what it's used.
111
00:22:29.590 --> 00:22:42.150
Sebastian Good: in order to put those citations through together. And then finally, you're going to want to run these sort of explanations through a series of deterministic tests, and continuously audit the responses and user feedback
112
00:22:42.150 --> 00:22:55.380
Sebastian Good: And it's subtle here in the mock-up, but there's a little thumbs up and thumbs down here, or there's a, you know, let's monitor this experience. Unless you have a chance to have the users give you some feedback.
113
00:22:55.380 --> 00:23:10.260
Sebastian Good: you're going to be flying blind. Now, you can use a firm like Justin to give you some feedback, you can have your own staff look at it, and you will, and you should, but your users are your best source of feedback, so you really want to have a design that can incorporate that.
114
00:23:10.260 --> 00:23:16.809
Sebastian Good: Often, that just means looking at the shape of that feedback and improving these underlying pipelines.
115
00:23:17.380 --> 00:23:18.250
Julie Ingalsbe: Okay.
116
00:23:18.390 --> 00:23:19.450
Julie Ingalsbe: Absolutely.
117
00:23:20.320 --> 00:23:24.209
Sebastian Good: Let's talk about the next set of activities, Justin.
118
00:23:25.180 --> 00:23:33.529
Justin Suter: Yeah, so shifting to the analyzes designation, calling out the Schwab Portfolio Insights feature, it's first.
119
00:23:33.530 --> 00:23:47.919
Justin Suter: Gen AI capability for retail clients. It's currently only available to select accounts. What this tool does is it analyzes the client's actual holdings and surfaces summaries of activity based on market news and proprietary research
120
00:23:47.920 --> 00:24:07.290
Justin Suter: from the Schwab Center for Financial Research. So this is where Schwab is trying to attack exactly what we just talked about on this previous slide of, you know, sourcing, where is this coming from, where they're going to be drawing from insights that they're producing already. And that's where they're keeping the guardrails very much intact to live within that ecosystem.
121
00:24:08.100 --> 00:24:25.619
Justin Suter: This tool lives in a expandable Portfolio Insight section on the sidebar of the Secure Site homepage, so someone can't log in, they can expand this Portfolio Insight sidebar, which is exactly the screenshot that we have up on the slide for you right now.
122
00:24:26.030 --> 00:24:38.279
Justin Suter: And each insight is a timestamp blurb built out from three content type. There's an AI powered narrative support snapshot describing that day's portfolio performance against the prior day's close.
123
00:24:38.280 --> 00:24:52.189
Justin Suter: There's a curated commentary from the Schwab Research Center and other Schwab experts. And then there's a news recap covering up to five S&P 500 equity movers that most affected the client's portfolio performance that day.
124
00:24:52.400 --> 00:25:10.609
Justin Suter: And one really important point here is that the way it's designed with this slide-out is that it can be expanded in red without navigating from the data and the feature that they're talking about itself. So that's really useful, because anytime you force a user to leave a tool.
125
00:25:10.610 --> 00:25:15.620
Justin Suter: To verify something else, you run the chance of them abandoning that journey. So this
126
00:25:15.620 --> 00:25:25.910
Justin Suter: setup and this architecture is really helpful there. And the last point worth flagging about Schwab from my end is, again, how they're handling sourcing and disclosure, because again.
127
00:25:25.910 --> 00:25:38.439
Justin Suter: It sounds like we're beating a drum, dead horse maybe already, but this is so important of building that trust, right? The source material for the portfolio insights appear in a right side light box that can overlay the homepage.
128
00:25:38.480 --> 00:25:46.900
Justin Suter: And it displays the full article text along with the publication date and timestamp. So the client can see exactly where that information is coming from.
129
00:25:47.470 --> 00:25:48.300
Julie Ingalsbe: Mmhm.
130
00:25:48.680 --> 00:26:07.650
Sebastian Good: This is really important. Some solutions we've built for folks, and I'm not a Schwab customer, so I haven't looked at this particular product, but often just citing the article definitely grants credibility, but if someone actually wants to look at it, and it's a 20-page article, or even if it's a 1,000-word article.
131
00:26:07.680 --> 00:26:21.000
Sebastian Good: highlighting the piece that informed your judgment is important, right? Which speaks to the, like, being able to understand what the article is even saying. You know, why is this relevant to the conversation? So sometimes, even like we see on Google, right? Like.
132
00:26:21.190 --> 00:26:31.930
Sebastian Good: the little snippet that explains it is important. Julie, let's talk a little bit about the design patterns that we think are important when we're offering analysis to customers.
133
00:26:32.420 --> 00:26:49.169
Julie Ingalsbe: Yeah, so analyze is really that next step on the journey, right? So we going from inform to analyze, I'm kind of seeing what's happening in the markets or my portfolio and now I really wanna get an understanding of, you know, what's my situation, what do I need to know? Like what do I need to take back from this?
134
00:26:49.170 --> 00:27:13.869
Julie Ingalsbe: And so some of the trends that we're seeing around this is just really explaining, you know, portfolio exposure, like what's under the hood of those funds that I own. You know, this is something that we've had kind of in our technology stack for a while, but being able to, you know, do an X-ray on a portfolio to really understand, oh, I actually own, you know, Shopify four times. We had a client that said like.
135
00:27:13.870 --> 00:27:37.490
Julie Ingalsbe: The amount of people that own Shopify and multiple accounts across their firm is just kind of interesting. So really just being able to give that information to the user in a really clear and informative way, as well as just trends, risk, performance, fees, cash flow, all those kinds of things. But the piece that we have been playing with is actually giving that information in bite-sized pieces.
136
00:27:37.490 --> 00:28:02.470
Julie Ingalsbe: Rather than a giant PDF report or a giant dashboard or five tabs of information, we're actually going to use the AI to inform or analyze in a smart way and giving little pieces of the story in, you know, in a kind of an impact order, if you will. So, you know, we go from a portfolio monitoring screen where we're highlighting that the user.
137
00:28:02.470 --> 00:28:08.599
Julie Ingalsbe: you know, might have their risk score has increased.
138
00:28:08.600 --> 00:28:14.790
Julie Ingalsbe: And so, you know, that's quantified data that we can kind of bring to them based off of a rating system that Morningstar has.
139
00:28:14.790 --> 00:28:37.730
Julie Ingalsbe: And we can say, hey, your portfolio seems a little, it's a little bit more aggressive than it was prior. And then we can kind of inform the user along the way of why did that happen? And again, in a small bite-sized pieces. So what's increasing my risk? What are the things that are helping my portfolio? What are the things that are hurting my portfolio?
140
00:28:37.910 --> 00:29:01.980
Julie Ingalsbe: Some of the really important things from a UX/UI perspective is just really making sure that we're showing the data behind the answer. So we're backing up the information with some sort of data source or data content or highlighting that part in the article. Making sure that the assumptions are all in plain English. We're not trying to be too fancy here. That, again, makes it so that the user is losing that trust.
141
00:29:02.210 --> 00:29:26.790
Julie Ingalsbe: We want to separate facts from interpretation, really making sure that, you know, it's clear to the user the difference in the UI, like what is interpretation, what is actually true, pure data, letting them drill into charts and holdings and transactions so that they can get information. All of these little cards allow the user to tap into them to get more content so that they can get more information around that particular topic or issue.
142
00:29:26.930 --> 00:29:30.289
Julie Ingalsbe: And then just continuing the conversation, right? So…
143
00:29:30.290 --> 00:29:50.239
Julie Ingalsbe: We have AI rooted at the bottom of a lot of our interfaces. It's not hidden in the right-hand corner where someone has to go find it. It's actually a part of the display so that the user can continue that conversation, ask questions, and actually leads into our next step that we'll talk about.
144
00:29:51.500 --> 00:30:06.389
Sebastian Good: Yeah, from a tech perspective, thinking about how you deliver these experiences is interesting, and it builds on the informed semantic base. So a key insight, again, kind of repeated from before.
145
00:30:06.400 --> 00:30:30.930
Sebastian Good: This kind of analysis, like this example here, a risk analysis and breaking it down by sector or by exposure, is not something an LLM is going to analyze reliably in one shot. You could just ask, what do you think about this? And sometimes it'll get it right, sometimes it'll get it wrong. You can convince it to be reliable by operating like a longer script, like a programmer would use to write a program.
146
00:30:30.930 --> 00:30:35.580
Sebastian Good: and do a full analysis. LLMs are getting much better about running this kind of analysis.
147
00:30:35.580 --> 00:30:41.839
Sebastian Good: Someone at a hedge fund or a big shop might use it to do bespoke analysis, but
148
00:30:41.840 --> 00:30:58.020
Sebastian Good: That's not going to be cost effective in a retail or advisory environment. And so really, you want to fall back to thinking about what analysis am I actually doing here? A lot of those analyses are deterministic. They are tools you can go run, and you can run them in advance.
149
00:30:58.020 --> 00:31:04.870
Sebastian Good: And you can cache their results. You can be smart about an analytics pipeline. You can use the analytics pipeline you still have.
150
00:31:04.870 --> 00:31:26.919
Sebastian Good: in order to inform this more AI enriched experience. So what we recommend then and what we build is that you end up using the AI in conversations like this for data tone and user customization, right? So we've presented the data into the UI, into the LLM. This is all you can reason about, but
151
00:31:26.920 --> 00:31:40.710
Sebastian Good: as Julie said, keep it tight, keep it short, recommended phraseology, the correct version of French financial terms to use in Canada. These are things you can give to the LLM to produce the useful, punchy summarizations
152
00:31:40.710 --> 00:31:54.020
Sebastian Good: of the analysis that make it so useful and prompt the set of deeper dive questions, confident that the actual numbers and analysis you're doing is powered by a deterministic model.
153
00:31:54.320 --> 00:31:55.889
Sebastian Good: If you don't…
154
00:31:56.150 --> 00:32:02.069
Sebastian Good: Turn that unstructured data into structured data, and if you don't use the structured analysis to inform
155
00:32:02.340 --> 00:32:19.499
Sebastian Good: the text that you're giving your users, you're gonna have serious accuracy problems, and you'll lose the ability, for what Julie talked about, to be able to drill in. When someone says, how is energy exposure helping my portfolio? If you offer them that click.
156
00:32:19.590 --> 00:32:33.189
Sebastian Good: you better be prepared to turn that into a view of energy exposure or a successive prompt to the LLM that is scoped around that question. So again, what feels a little unstructured is actually very structured under the covers.
157
00:32:33.930 --> 00:32:37.290
Sebastian Good: Justin, let's talk about, advising people.
158
00:32:38.090 --> 00:32:47.429
Justin Suter: Yes, this is where we start to get a little exciting and scary at the same time because we're talking about AI Invest here.
159
00:32:47.500 --> 00:32:58.379
Justin Suter: It provides portfolio analysis, investment research, trade capabilities, investment help, all anchored around the virtual assistant called Amay. And I very much hope I'm saying that correctly.
160
00:32:59.620 --> 00:33:10.190
Justin Suter: And this is a very, really powerful virtual assistant. It's able to respond to virtually any user prompt that we tried it during our testing.
161
00:33:10.190 --> 00:33:25.640
Justin Suter: And it can give you lots of advice. It can handle lots of, as you were just talking about, Sebastian, it can handle lots of different types of inputs, whether you're going to try to be very structured or unstructured with what you're asking it. But, and it's a very large but, it does so with very few guardrails.
162
00:33:25.840 --> 00:33:39.890
Justin Suter: A couple big issues that stood out in our testing is that it does occasionally provide some incorrect or misleading information, which isn't, you know, all AIs do it, but now we're doing it in the context of
163
00:33:40.030 --> 00:34:02.640
Justin Suter: actual investment advice. So, of course, it carries a lot more weight alongside it. During our testing, it also periodically lapsed into Chinese characters, which we believe is a sign of the fact that it's underlying by the Alibaba Cloud HThinkGPT model underneath it. So, some bugs that at this point, of course, you know, we're performing a lot of this report.
164
00:34:02.640 --> 00:34:12.539
Justin Suter: published in early June. We performed a lot of the research across April and May. Large chance that this is already fixed by now, but this was live, and this was doing that while, we were testing it.
165
00:34:13.260 --> 00:34:21.070
Justin Suter: And it does sometimes have some spelling errors or a little bit of gibberish in the responses, and because a lot of the responses from this particular
166
00:34:21.620 --> 00:34:32.089
Justin Suter: Chat is… they're very, very long, so, like, it's a… it can be a lot of really good content, then all of a sudden, in the middle of it, there'd just be this one line, and you're like, that's… it didn'
167
00:34:33.250 --> 00:34:52.480
Justin Suter: And notably, when we asked AMA directly, hey, what are your capabilities? They give us this really, really long list. Oh, I can do this, I can do this, I can do this, I can do this. And so like, okay, well, what can't you do? I cannot place trades and I cannot guarantee returns. So it seems like outside of those two things, pretty much everything else is in scope.
168
00:34:52.489 --> 00:34:52.829
Julie Ingalsbe: Okay.
169
00:34:52.830 --> 00:35:05.379
Justin Suter: for this particular capability, and this contrasts quite substantially with what we just saw with Schwab, where they drew a very clear line around what the tool will and won't do.
170
00:35:05.420 --> 00:35:23.759
Justin Suter: And it's a great demonstration of the difference between an incumbent pushing the boundaries a little bit, but keeping very, very tight guardrails in place, versus a fintech that is just really pushing the envelope and trying to see if, hey, let's, you know, get out there first and get after it and see
171
00:35:23.980 --> 00:35:32.599
Justin Suter: how the chips fall, I guess. But this is a great example of the types of tools that we see in this advice bucket.
172
00:35:32.970 --> 00:35:40.210
Justin Suter: They stopping just short of actually placing the trade, but pretty much telling you what to go and do, and putting all that context into it.
173
00:35:41.250 --> 00:35:56.269
Sebastian Good: Yeah, I think some of what you describe here, is really reflected in the structured versus unstructured conversation. Right. I mean, we've seen even when you query authoritative, information and hand it to, an AI.
174
00:35:56.270 --> 00:36:05.470
Sebastian Good: it may choose last year's report instead of this year's report, so it gave you an earnings per share that's correct, but it's not correct, because it's not the most recent one, right? And…
175
00:36:05.470 --> 00:36:29.039
Sebastian Good: God bless these little LLMs. They love to be so helpful. They'll spit out what you want them to spit out. Then they'll add another paragraph of stuff that's just their little imagination trying to help you. So you really have to be aware of what's of what's coming out and whether that's advice or not. So with that, let's talk about what those UX patterns look like when you want to actually give advice, not just analysis.
176
00:36:29.980 --> 00:36:46.830
Julie Ingalsbe: Yeah, so at Xpero, you know, our number one kind of thought around advice is that it's a conversation. It's questions, it's feedback, it's a back and forth. It's not a, here, this is the trade you should make, right? So we want to make sure, we want to help users evaluate those choices.
177
00:36:46.830 --> 00:37:11.029
Julie Ingalsbe: We want to give guided experiences that suggest those next steps. So we have this idea of a choose your own adventure screen. It's in the middle section there called guided navigation of what you'd like to explore next with generated ideas of where this user could go on their journey, simulate changes in a portfolio or explain volatility. And those can all be generated based off of user behavior.
178
00:37:11.030 --> 00:37:23.170
Julie Ingalsbe: what the user's done, what they've asked for, what they've, you know, engaged with before, those kinds of things. To comparing options, we kind of touched on this a little bit earlier today, but…
179
00:37:23.170 --> 00:37:39.949
Julie Ingalsbe: The comparison tool that kind of bubbles up those insights and is giving not necessarily advice, but it's giving the contrasting data in a really clear and simple way, recommending learning paths or helping the user choose those strategies.
180
00:37:39.950 --> 00:37:54.579
Julie Ingalsbe: I guess just to really echo what Justin is saying, you know, from a UI perspective, we just… we really want to make sure we're setting those boundaries between education, guidance, and advice, and making sure that they're very clear to the user so that we can keep that trust.
181
00:37:54.580 --> 00:38:09.560
Julie Ingalsbe: You know, it's not just one magic answer, it's multiple, you know, and making sure that we're clearly displaying that back to the user. And then trade-offs, like this comparison tool, it's like, we're shading the trade-offs
182
00:38:09.730 --> 00:38:32.170
Julie Ingalsbe: And like a red and a green, just to say, like, you know, this might be a positive thing, this might be a negative thing, you know, that there's, this portfolio is more concentrated, or this fund is more concentrated and, has 127 holdings versus a lower turnover, you know, that suggests a steadier approach or whatever. So, the idea here is, like.
183
00:38:32.170 --> 00:38:55.920
Julie Ingalsbe: Displaying those trade-offs is not a single, and instead of showing just one single magic answer. And then obviously risk, cost, tax, make sure that we're looking across the gamut of all the data available and trying to create a story that is holistic and tells all pieces rather than just a performance story or just a rating story. We want to make sure that it's a holistic.
184
00:38:55.920 --> 00:38:58.359
Julie Ingalsbe: Kind of, conversation.
185
00:38:59.720 --> 00:39:13.069
Sebastian Good: From the tech side, what I'll say is that advice is a conversation, right? That's what you said, Julie. That means the technology for how you talk to the user needs to be mindful of how humans interact.
186
00:39:13.330 --> 00:39:29.299
Sebastian Good: The frontier models we work with when we talk with ChatGPT or Claude or Gemini have been trained on human conversations. They'll talk with you, but they're not necessarily trained down the structured road of the advice that you're prepared to give.
187
00:39:29.300 --> 00:39:47.669
Sebastian Good: And so you've got to be able to understand, I am talking with the user about her balances. I'm talking with the user about where she can find one of my banks. I'm talking with the user about screening for funds, right? These are each activities we've built structured data into. When that user says.
188
00:39:47.690 --> 00:39:53.259
Sebastian Good: oh, yeah, I do want to trade in this stock, but I don't remember how much is in my account.
189
00:39:53.270 --> 00:40:11.019
Sebastian Good: You've got to move to that conversation, make sure you're handling it clear, and then move back to the original conversation. If you just leave that whole conversation in what's called the context window for an LLM, there may or may not sort of follow the thread. So we want to make sure we're explicit about understanding that user's intent.
190
00:40:11.020 --> 00:40:34.819
Sebastian Good: And bringing them into that conversation where we can answer with structured and reliable tools. And often those tools are filling in a worksheet, as it were, right? If someone is screening for stocks, they're ultimately saying, well, I want technology stocks that still have a low P/E ratio. They may not say it in those words, but if they're looking for cheap tech stocks, I need to turn that into a sector filter.
191
00:40:34.840 --> 00:40:51.410
Sebastian Good: and a price filter, so that I can drive a conversation about which ones are available. So, I may be filling those in, and when they say, well, they don't have to be as cheap as that, they can be a little more expensive, I want to tweak that PE filter, not tweak the technology sector, not get lost, not confused. So.
192
00:40:51.410 --> 00:41:00.830
Sebastian Good: the notion from the technology side that the conversation itself is an artifact and can be managed is really important. And then the other thing I'll say is.
193
00:41:00.840 --> 00:41:09.720
Sebastian Good: Hundreds of years of data visualization and user experience didn't get thrown out the window just because we had all LLMs that could spit a paragraph at you.
194
00:41:09.720 --> 00:41:28.419
Sebastian Good: In that lower right example you can see on the screen, I was having a conversation with Expero's natural language research tool about cheap tech stocks, and Adobe and Zoom showed up. Those are tech stocks that have been a bit beat up recently, right? Well, let's say I'm interested in those. I've looked at the summaries from the Inform, and now I want some advice. Well.
195
00:41:28.420 --> 00:41:32.700
Sebastian Good: I can compare them side by side. If there's a nice comparison tool.
196
00:41:32.700 --> 00:41:52.349
Sebastian Good: Maybe I should use that nice user experience tested comparison tool to further that conversation, as opposed to hope I can write a paragraph that captures all of it. So, you know, use tools to advise, not just prose, and then be explicit. Finally, I'll say the technology needs to decide, are you only going to say
197
00:41:52.350 --> 00:42:00.050
Sebastian Good: advisory statements that are safe by construction, meaning you'll only let the LLM say certain things that come from structured output.
198
00:42:00.050 --> 00:42:05.799
Sebastian Good: That's one approach. It's more conservative. Or you're going to let the LLM kind of go where the user wants to go.
199
00:42:05.810 --> 00:42:20.940
Sebastian Good: But check the results, with successive agents, and technology to make sure that, you know, if it's offering advice, it's only advice from, say, a research report, and it's understood that it's that advisor giving the research. Or if your bank is willing to stand behind the LLM,
200
00:42:20.940 --> 00:42:33.069
Sebastian Good: Then explicitly, check, you know, what you're confident behind. So, those are the technology considerations when we take that information analysis and admit that it may start looking like advice.
201
00:42:33.380 --> 00:42:39.540
Sebastian Good: Final bucket. I'm advised, I'm informed, I've analyzed. By golly, it's time to do something.
202
00:42:39.540 --> 00:42:42.949
Justin Suter: Yeah, let's go, let's act.
203
00:42:43.160 --> 00:42:45.860
Justin Suter: We're going to talk a little bit about Composer.
204
00:42:46.180 --> 00:42:51.410
Justin Suter: And, of course, we're on the furthest end of the autonomy scale here, and the big…
205
00:42:51.860 --> 00:43:07.899
Justin Suter: big headline about Composer was, as I mentioned earlier, a lot of the research for this report conducted in April and May. We've published late May, early June. Composer was actually just purchased by SoFi just a couple weeks ago. I think the announcement came out on June 23rd, so…
206
00:43:07.900 --> 00:43:25.420
Justin Suter: More movement in there, but still worth talking about from the capability perspective of what we're seeing here. Composer, it allows clients to build, backtest, and automate different investment strategies. It calls them symphonies, so Composer, Symphony, lots of music threads coming through here.
207
00:43:25.440 --> 00:43:30.880
Justin Suter: And on the Discover page, users can input in a natural language conversation.
208
00:43:30.880 --> 00:43:47.289
Justin Suter: Both to build a symphony, or just ask general questions about their portfolio. When someone's looking for an investment strategy, the AI tends to respond with a series of questions, like your timeframe, your approach, your risk appetite, before presenting some symphony proposals.
209
00:43:48.090 --> 00:43:55.049
Justin Suter: And that's important because it does really work hard to try to get a sense for what the user wants.
210
00:43:55.050 --> 00:44:11.670
Justin Suter: Without just jumping into it, which we do see with some AI, right? They're just gonna say, like, oh, you, you want that? Here's three ways to do it. This does definitely, we, we want to call out Composer and, give them some praise for really at least trying to make sure that they're generating something that
211
00:44:11.950 --> 00:44:35.190
Justin Suter: is personalized and within the framework of what the user is looking for. And actually for more general prompts, like questions about specific security, Composer sits more in the advisors tier where it's going to be talking about, hey, here's context, here's quote data without issuing direct recommendation. But there's a more advanced path
212
00:44:35.190 --> 00:44:47.970
Justin Suter: to the composer perspective, and that's where you can get onto a new symphonies page. So that's where they have their very sophisticated interface for building symphonies manually.
213
00:44:48.730 --> 00:44:49.180
Julie Ingalsbe: Mmh.
214
00:44:49.180 --> 00:44:50.919
Justin Suter: And this is, you know.
215
00:44:50.950 --> 00:45:11.620
Justin Suter: Definitely far beyond something that I would be willing to engage with an AI with, but it's really fascinating because you can use a logic block such as the asset, the group, the weight, different filters, and at the top of the page they have a create with an AI button that launches a pop up that has three different tabs, chat, JSON, and def.
216
00:45:11.650 --> 00:45:34.509
Justin Suter: The chat tab is going to have a conversational input around preset prompts. So that's going to help lower the barrier a little bit for someone who isn't ready to write this type of logic from scratch, which would definitely be me. You can also select a preset prompt, something like, hey, quote, you know, can you build a stock-only strategy that does well in both rising and falling interest rates? We don't love that, right? So that's going to trigger.
217
00:45:34.510 --> 00:45:57.969
Justin Suter: The creation of a symphony, it's going to have specific weights for different securities, along with some nice concise text at the bottom that outlines, hey, this is how suitable I think this is for you. It does also present what are the risks of this proposed strategy. But again, if we want to get even more advanced, a user can skip all of this preset stuff and just go in and type a custom plain text prompt.
218
00:45:57.970 --> 00:46:03.299
Justin Suter: that the AI will translate into a symphony based on their objectives.
219
00:46:03.800 --> 00:46:16.189
Justin Suter: I do want to flag how the AI does provide a justification for the symphony it creates. It doesn't directly recommend it. The client still has to decide to move forward, so that even within this acting
220
00:46:16.660 --> 00:46:35.309
Justin Suter: It's a conversation. Composure does kind of keep the human in the loop, where it's like, hey, this is what I've built for you. Would you like to do it? It offers the opportunity, as we said, to back test the strategy. But then it can, you can say yes, and then it'll automate it out for you from there. So this is definitely the
221
00:46:36.040 --> 00:46:44.470
Justin Suter: you know, futuristic look at what some of this Agentic AI can look like for sure.
222
00:46:45.290 --> 00:46:48.499
Sebastian Good: I really like this example because it, in a way.
223
00:46:48.680 --> 00:47:12.300
Sebastian Good: pulls the curtain away from a lot of what I've been talking about from the technology side, which is like underneath your conversation is a structured artifact if you want to provide it safely. And here the structured artifact is basically a program that'll algo trade for you. Right. So it's obvious why it has to be structured. But this is a really great example where you're kind of seeing, oh, this is not just a conversation. There's a there's a thing happening under here, underneath here that's very explicit.
224
00:47:12.300 --> 00:47:18.759
Sebastian Good: And that pattern of structured to unstructured mapping underlies safety and accuracy throughout all these activities.
225
00:47:19.510 --> 00:47:19.830
Julie Ingalsbe: Mmh.
226
00:47:19.830 --> 00:47:25.779
Sebastian Good: So let's talk about UX considerations when we design experiences that can act for us.
227
00:47:26.270 --> 00:47:51.170
Julie Ingalsbe: Yeah, so we really just want to get those users right from making a decision to execution, but we want to do that in a really safe way. So obviously, some of that ideas that we had around rebalancing or portfolio construction, money movement, rules-based investing, as Justin was just showing, we had some ideas around kind of a trade impact scenario where
228
00:47:51.170 --> 00:48:05.320
Julie Ingalsbe: if I'm about to make a trade, I can get a trade impact summary that will tell me some things and categorize it by all types of data, from portfolio fit to risk impact to concentrated impact to sector impact.
229
00:48:05.320 --> 00:48:21.330
Julie Ingalsbe: to overall impact. It's kind of like that previous screen, but in a, like, Julie-friendly look and feel, right? With ups and down arrows. So it's taking that same idea and putting it in more of a self-directed, you know, beginner user experience.
230
00:48:21.600 --> 00:48:46.410
Julie Ingalsbe: So just really making sure that it's really clear and easy to use and making sure that you're getting that user's approval before moving forward. Another product that we've been working on really heavily at Expiro in the last few months is a portfolio builder and rebalancing tool, which really helps the user build a portfolio from day one, where we see a lot of accounts get opened and then they just get left open and no one
231
00:48:46.410 --> 00:49:11.380
Julie Ingalsbe: really knows what to do, and then money is just sitting in cash and that's not helping anyone. So how can we actually get that user to make their very first trade in a really simple and safe way? And this really walks them through that journey of, you know, identifying their risk, identifying their goal, and being able to create something really easily. But this also flexes. So if it's someone who really wants to have like a multi fund
232
00:49:11.380 --> 00:49:27.400
Julie Ingalsbe: approach or stocks and funds in their portfolio, it will allow for those kinds of things as well. And so it kind of advances as much as the user wants it to, which again is, you know, trying to create that trust so that the tool is allowing them to have the features that they need.
233
00:49:29.240 --> 00:49:46.649
Sebastian Good: From the technology side, this is where the rubber really meets the road in terms of structured versus unstructured. You may be willing to let the LLM just kind of wing it a little bit more on the advice side, right, if that fits your risk profile, sufficient disclaimers and such, but
234
00:49:46.650 --> 00:50:05.239
Sebastian Good: if someone is actually launching the missiles, as we like to say as programmers, like, a thing you can't take back, I did a trade, you really need to have a very strictly defined safety process. And so, what you see on the AI side of it is that you have to run your agents in a sandbox
235
00:50:05.240 --> 00:50:17.040
Sebastian Good: for technology reasons, you're always gonna have a sandbox that you run an agent in. But by this, I mean the interaction is, much less now, here's what the user said, what do you think? You usually…
236
00:50:17.100 --> 00:50:36.479
Sebastian Good: pull the set of possible inputs to the LLM down to a very small set of inputs, and the actions that it's allowed to take are very small, and you have an agent or a deterministic program confirm those. Meaning, if I've got an algorithmic trading event, I'm not just having an LLM write up some random network calls to make the trade.
237
00:50:36.480 --> 00:50:44.399
Sebastian Good: there's a trade ticket it's allowed to fill out, and unless it's filled out exactly right, I can't trade, and nothing else it wants to do is even possible. So there's a very
238
00:50:44.400 --> 00:51:03.260
Sebastian Good: process there. And then from a safety perspective, it might be a process, like at Robinhood, you can now do some of these automated, excuse me, at Robinhood, you can do some of these automated trades, but they're in a designated portion of your portfolio, and only a certain amount of money can be spent on them, right? They're basically saying.
239
00:51:03.260 --> 00:51:08.389
Sebastian Good: shoot your foot off if you want, but it'll only be your foot, right? There's a very defined explanation there.
240
00:51:08.390 --> 00:51:31.020
Sebastian Good: Or, there could be an explicit approval, right? Like, we've designed some workflows, as proposals where an AI may offer a trade, but it's going to pop up on your mobile phone for a confirmation. Right? Here's a second kind of set of eyes on it. This is what's about to, to happen. So this is what I mean by queue for an approval by a safety agent and have a human in the loop process. So that adds.
241
00:51:31.020 --> 00:51:44.870
Sebastian Good: that auto rebalance that Julie was talking about, or the algorithmic trading Justin was talking about, from a technology perspective, has to be implemented as a set of actions that are explicitly reviewed by a non-LLM process.
242
00:51:44.870 --> 00:52:08.060
Sebastian Good: Maybe it's sent to a human. Maybe there's a dollar ceiling. But you know, there's a there's an explicit structure you're happy with. And only then do you run them. And it should probably go without saying, but I'll say it anyway. You've got to test the hell out of this so that you understand all the possible ins and outs of what that act may be. And there, there's really no option except to only be safe by construction. If that
243
00:52:08.060 --> 00:52:09.910
Sebastian Good: Thing is only allowed to trade.
244
00:52:09.990 --> 00:52:14.070
Sebastian Good: That's all it's allowed to do, and you've got to put the right limits on it.
245
00:52:14.260 --> 00:52:24.770
Sebastian Good: So, I think it'd be good to kind of, wrap all these, together. Justin, maybe if you could, talk a little bit about how, we… we use this?
246
00:52:26.390 --> 00:52:45.119
Justin Suter: Yeah, let's put a bow on it. And this is definitely one of the money slides. So if you've had us on in the background on a second screen and you've been multitasking at your desk, this is where you should just, you know, really pay attention. So we'll wrap it up. The three points right here on the slide, right?
247
00:52:45.360 --> 00:52:58.610
Justin Suter: consumer comfort is key, and a lot of customers are not comfortable with AI, how it's being delivered today. They know they're… they've… most people have been experimenting it, most people have an expectation that the firms they work with are using it, so they know it's there.
248
00:52:58.660 --> 00:53:20.760
Justin Suter: But most people are concerned about it. They're more concerned than excited and their dominant interpretation is that it's simply being used to replace human support. So that's a big barrier to overcome. And we've talked a lot about how to do that, right? Those transparent disclosures, visible paths to a human, human review of AI, all of these can build the customer comfort with AI and get them to take advantage.
249
00:53:20.760 --> 00:53:25.169
Justin Suter: of all the positive options that we've talked about today.
250
00:53:25.170 --> 00:53:25.880
Justin Suter: Oh.
251
00:53:25.880 --> 00:53:39.210
Justin Suter: Which leads to the second point, which is that, fix what you already have before building what's next, which, you know, could be a little bit obvious, but you can't build a really fancy AI experience on top of a poor underlying foundation, right? So,
252
00:53:40.030 --> 00:53:56.879
Justin Suter: Understand what you have, understand what's working, what's not working with the tech stack that you already have deployed, and maybe just invest in fixing that before you take the first step into deploying some consumer-facing tools. But we're talking about building trust, and trust comes from us building on a strong foundation.
253
00:53:56.890 --> 00:54:19.869
Justin Suter: And hopefully that third point should make you feel better if you're still very, very early in your AI journey. And this is that the ROI here is in retention, not acquisition. A very, very small minority of customers in our survey said that they're more likely to choose a firm specifically because it offers AI-powered features, right? These are not bringing people in the door, right? Those people are already…
254
00:54:19.870 --> 00:54:34.790
Justin Suter: You know, if the cool AI is what they want, they're already playing with Composer or something that's even further, you know, in the wild west of some of the fintech AI, right? But, you know, the clients that we work with and that Xpero works with and is talking to, right, we're talking about
255
00:54:34.790 --> 00:54:47.660
Justin Suter: really improving and deepening the existing customer relationships, and a well-deployed AI tool can certainly be a big part of that, and that should be that internal measure of success, is whether the users who are already open
256
00:54:47.660 --> 00:54:54.919
Justin Suter: to a good experience are actually getting one right now from your company, and that's the core message that I would like to end on.
257
00:54:55.890 --> 00:54:57.299
Sebastian Good: Love it. Thanks, Justin.
258
00:54:57.370 --> 00:55:06.170
Sebastian Good: Yeah, it's funny, however many tools technology throws at us, user experience is still about making users happy, right? And letting them accomplish their goals, like, don't… don't lose sight of it.
259
00:55:06.170 --> 00:55:24.780
Sebastian Good: From my perspective, what I want to wrap up just in terms of making these experiences possible is, A, this is a really great framework, Justin. This is a slide we've been using in our conversations about how our framework for building AI-enabled financial services experiences works.
260
00:55:24.780 --> 00:55:34.399
Sebastian Good: We talk about understanding decision support guidance and Agentic automation, and it's inform, analyze, advise, and act, right? So just kudos. I think this is the right way to think about it.
261
00:55:34.400 --> 00:55:48.520
Sebastian Good: For us, we can't build reliable experiences without the technical infrastructure that we've talked a little bit about today, and if anyone here is interested in how that works and can help accelerate, you know, their journey, happy to kind of dive in deeper, but…
262
00:55:48.520 --> 00:56:10.410
Sebastian Good: Having clear semantic definitions and partners with your data models, being able to generate the right citations to generate trust, generating the right language not only for translation purposes but compliance purposes, being able to manage your agents correctly and testing it all is a set of self-reinforcing behaviors that's difficult to bootstrap from nothing.
263
00:56:10.730 --> 00:56:28.500
Sebastian Good: And the last thing I'll just say from a tooling perspective is without the ability, in my opinion, to guide conversations through these kind of cited and error prone/structured/unstructured directions we've talked about, it's really hard to make these any more than…
264
00:56:28.890 --> 00:56:45.560
Sebastian Good: clawed with some data jammed on top, which, you know, users will recognize. So, definitely invite anyone who's interested to contact us. Justin, obviously, this is information from a much richer report with lots more statistics and examples.
265
00:56:45.560 --> 00:57:00.570
Sebastian Good: And obviously, we've just touched a little bit on the technical aspects and U.S. aspects of our own offering and the offerings we're designing for customers, so welcome any questions. I do see that there's been one question put in by Steve Dacus here.
266
00:57:00.870 --> 00:57:25.769
Sebastian Good: So thank you. That's very kind of you to say how great this presentation is. So way to butter us up. How much of the UX workflow design is provided by the Xpero team? How are client and third party content sources managed? So from our perspective, Steve, the Xpero offering is about doing UX design and offering the technology components and the vendor relations.
267
00:57:25.770 --> 00:57:47.140
Sebastian Good: And the data mapping with your own data to design, develop, deploy, and manage an experience for your customers. I should let Justin correct me, but where Corporate Insight thrives is helping you understand how well that experience has been constructed. They have great user research folks who will poke at it and tell you where it works well and where it doesn't.
268
00:57:47.140 --> 00:58:01.070
Sebastian Good: a competitive analysis. We do a competitive analysis when we help you think about what your next product is, but it's absolutely informed by the super broad-ranging work that Justin does. Justin, your comments there?
269
00:58:01.530 --> 00:58:12.369
Justin Suter: Now, I'm a little, you know, unsettled by how well you answered that on behalf of Corporate Insider. You're gonna be… You can start moonlighting for us. Yes, I would echo it. You know, we're, we're
270
00:58:12.610 --> 00:58:22.590
Justin Suter: taking it all into context, we're able to benchmark how one experience stacks up against a different experience offered by our competitor. And then we can also go look in other industries for
271
00:58:22.660 --> 00:58:34.820
Justin Suter: Other insights on how, hey, this is being deployed slightly differently in this other context, and surface all of that for you in what we hope is a very coherent and actionable resource.
272
00:58:35.130 --> 00:58:44.039
Sebastian Good: Awesome. And then our value proposition is, let's design it right for you and deliver it in, you know, the time that matters, like, a few months rather than a year.
273
00:58:44.330 --> 00:59:08.590
Julie Ingalsbe: Yeah, I was just gonna say, at the end of the day, we're a design-driven company, and so we're happy to lead a design effort, but also augment a design team if that's how the situation folds out. And then we do user testing and things like that. We also have a lot of designers on the team with a ton of financial experience background, so we're really grateful for having that expertise as well.
274
00:59:08.600 --> 00:59:19.729
Julie Ingalsbe: And, you know, we do partner with Justin and team to kind of learn from them and what they're seeing in the market, but we do our own research as well, if the project or the need calls for it.
275
00:59:20.470 --> 00:59:30.589
Sebastian Good: Awesome. Well, thanks everyone for joining us today, either live or on Memorex, just to date myself. And hope you guys have a great Wednesday. See you around.
276
00:59:30.960 --> 00:59:31.670
Julie Ingalsbe: Bye. Thanks.
