Cutting Through the AI Noise: Amy's Perspective on What Matters for Wealth Management

The pace of AI innovation is relentless, leaving many wealth management leaders struggling to separate the hype from what truly delivers value today—and what could fundamentally transform the industry tomorrow. In this session, industry expert Amy Young will share insights from her unique vantage point at the intersection of cutting-edge LLMs, agent-building tools, and real-world applications across financial services. You'll gain a practical framework to evaluate emerging technologies, identify where AI is already reshaping advisor workflows and client engagement, and anticipate where the next wave of disruption is headed. Walk away with the clarity and confidence to make smarter AI investments that position your firm to get—and stay—ahead.

Transcription:
Transcripts are generated using a combination of speech recognition software and human transcribers, and may contain errors. Please check the corresponding audio for the authoritative record.

Amy Young (00:08):
Good morning, everybody. Thanks for deciding to stay for this session. I want to thank Arizent for creating this space for us to focus on the impact of AI specifically for the advice industry. It's really important that we do that, I think. Oh my goodness, I'm distracted by the timing. Thank you. It's always an honor to speak at an event like this, but I am particularly pleased to be here today because today marks day two for me as a free agent. I've spent the last six years at Microsoft advising some of the largest wealth and asset management firms in the world on their AI strategies. That has truly been a privilege to have a seat at the table during such a consequential time. Over the last several months, I've noticed a growing need among business leaders for advice that is both independent and from someone who speaks their language.

(01:32):
Technology has become absolutely central to business strategy, and many business leaders don't have a deep enough understanding of the tech to have their rightful voice at the table. So going forward, I plan to offer advice to firms to help them make better AI strategy decisions by helping bridge the gap between technology and the business. What you are going to hear from me today is not a Microsoft perspective, but rather Amy's perspective on what matters for wealth management. I'm going to help you differentiate between the hype and what really creates value so that you can make more informed decisions for your firm. That's my goal. I want to start by talking about all of your goals. The number one reason that firms cite for wanting to invest in AI is because they want to help their advisors spend more time with clients.

(02:39):
There is remarkably little conversation about how to help advisors spend better time with clients.

(02:48):
To be fair, the early wave of GenAI tools really couldn't do much to help us spend better time with clients. That's mostly because LLMs are unpredictable, at least on their own. But over the past couple of years, the quality of outputs from AI-powered apps has improved substantially. That's partly because of the LLMs themselves, but two other things actually have driven much more of that impact. The first is that firms are starting to expose their proprietary data to the LLMs. The second is that we are learning to build applications that help the LLMs use that data to do useful things. That's important because while LLMs are a "black box," the applications that use them don't need to be.

(03:48):
As we're learning to build applications that can help the LLMs provide more reliable and trustworthy answers, we're starting to see the true potential of AI come into focus. I would argue it's going to help us spend better time with clients instead of just spending more time with clients. Over the next couple of days, you're going to see lots of presentations and cool demos from great product companies, but you might not have any greater clarity when you go back to the office on what your firm should actually do. My goal is to help you think beyond the capabilities of individual tools and start to think about an AI strategy that can help your firm actually support advisors to spend better time with clients. I propose to spend the next 30 minutes with you doing the following:

(04:57):
First, I want to provide a bit of a retrospective on how the technology has evolved in the three years since ChatGPT took the world by storm.

(05:08):
Then I am going to map those tech tool enhancements to the product announcements that you've all seen come from the wealth industry. I'm going to give you a preview of what the tools that we now have available are helping people work on, which will drive the next announcements you're going to see in the months ahead. Then I'll wrap up with some ideas on how to think a little differently about ROI. Now I'd like to ask you to think back to November of 2022. When ChatGPT was first launched, it was basically just two things: an LLM and a user interface. Remember how it couldn't give us information about anything that happened after October of 2021? That's because at that time, the model was ChatGPT's only data source.

(06:10):
It was basically a word predictor based on knowledge patterns that it had learned from the public internet up until that time. But in AI terms, it lacked context. There are two parts to context. The first is correctly interpreting a user's question. We all know there are many different ways to interpret a question. For example, when I ask my husband on a Saturday morning, "What do you want to do today?" He knows that what I really mean is, "What do you want to do today that you know I also want to do?" and after we've finished whatever chores I have planned for us.

(07:06):
The importance of understanding the user's intent was what drove all of the initial focus on the prompt. Remember how we thought at that point that AI was going to create all of these new jobs for something called a "prompt engineer"? Haven't heard that in a while. That's because at the time, in the absence of other context, the quality of the prompt had a huge impact on the quality of the answer. So that's what's going on at the user interface level: what's the real question? The other thing that has started to evolve is what happens with the knowledge sources. If I had asked ChatGPT at that time, "What craft beer festivals should my husband and I check out this weekend?" it would not be able to answer that question because it didn't have a current knowledge source or anything specific to my location.

(08:09):
Without the context that would help it better understand questions or provide trusted knowledge sources, ChatGPT was basically good at two things: text generation (but only on very generic topics) and summarization. Again, this was only on text that you fed into the chat window, which at the time was very small. It couldn't tell you anything about your business. It certainly couldn't do math. And oh, by the way, everything that you were feeding into that context window, OpenAI was using to train the next generation of its models.

(08:55):
It looked like at that point that this whole GenAI thing was going to be a "nothing burger" for the wealth industry, but there were innovators who saw potential in the text generation and summarization and started to focus on the context problem and ways that we could bring better context to the LLM so it could do more useful things. In May of 2023, Microsoft introduced the concept of the Copilot Stack. This wasn't a product, but it was an early set of tools and techniques to help advisors safely and securely bring context to the models. The foundation of the stack was the ability to use the LLMs in a firm's secure cloud tenant. That was obviously critical to giving enterprises the confidence that their data wasn't being used to train the next generation of the model.

(09:59):
The stack also provided tools to start understanding the user's intent better. The first of these was called a Metaprompt. A Metaprompt is basically a set of standing instructions for what an LLM should and should not do in the context of a given app. A metaprompt might say something like, "You're a tool that is designed to summarize meeting notes for advisors. You will only use meeting transcripts as your data source to which you have permission to access and you won't provide any opinions or judgments on those meeting transcripts." Defining an app's scope in this way reduced hallucinations and started to address safety concerns because it provided guardrails that prevented people from trying to use these tools to do bad things. You remember the stories about getting instructions from ChatGPT about how to build a bomb, right?

(11:12):
The stack also provided ways for developers to bring trusted knowledge sources into their tools. Instead of having to count on the black box of the LLM for answers, this gave developers the power to control where the LLM went to get its answers, and that had a huge impact on accuracy. The models also improved at this time. We saw a flurry of new models being released: small language models and models specializing in all sorts of different things. The cloud providers started to create model marketplaces that made it easier for developers to find the right model and test it for their use case. The emergence of this enterprise-grade tooling resulted in the first announcements from our industry of production use cases that actually used proprietary data with language models. One of the most common use cases at that time was enterprise knowledge search.

(12:30):
This is basically helping your people find answers to questions about policies and procedures. Morgan Stanley was one of the flagship firms to announce something like this. Morningstar announced the first client-facing GenAI tool with its MO digital assistant that was grounded in Morningstar's ratings and research. Microsoft announced M365 Copilot in November of 2023, and that brought grounding to a whole new level because you could use the LLMs grounded in all of the data in your M365 tenant. We're talking about your Teams chats, transcripts, emails, documents, SharePoint sites, and of course all the stuff in Office. That went a long way to being able to generate text relevant to our business case. Also at this time, we saw Jump and Zox announce their first meeting note tools.

(13:47):
The end of 2023 was an interesting time for the industry because opinions about AI were still quite divided. You had a handful of trailblazers that were heavily investing in learning and capability building, thinking deeply about implications for their business, but then you still had most of the industry that was still skeptical, citing all of the risks of AI as their reason for not really engaging. But 2024 brought new enhancements to the stack. Developers started to experiment at the user interface level by making LLM-powered tools available within SaaS applications and in third-party platforms.

(14:42):
Microsoft was having lots of interesting conversations with partners who wanted to bring their tools into M365 Copilot starting at this time. At the prompt and response filtering layer, the integration of code interpreters powered some important new breakthroughs. Now, a code interpreter is a tool that translates natural language into code, and that was important because it allowed the LLM now to pull structured data from enterprise systems to accurately answer questions that involved things like sales and other transactions. Code interpreters also brought in the ability to reliably do math calculations of various kinds. Another important innovation at this time was AI evaluation tools. The cloud providers created these tools so that developers could evaluate the quality, safety, and performance of the applications they were building as they were building them, which reduced a lot of the risks and started to meaningfully accelerate time to market.

(16:09):
The industry used this continually evolving toolkit to launch the second wave of wealth-centric AI announcements in mid-2024. Conquest Planning announced its SAM assistant that allowed advisors to run scenarios on their financial plans with natural language instead of having to click through all of the filters and screens. Jump integrated the ability to draft emails into its platform, and Morgan Stanley launched its own meeting note and follow-up tool that it called Debrief. For the latter half of 2024, the enterprise toolkit continued to mature, and more and more firms were doing POCs and experiments to learn how these tools and architectures could help them build apps to meet their needs. Nine to 12 months later, we had a third wave of announcements. Morningstar announced a tool that streamlined the process of creating client proposals. It extracted the current portfolio holdings from a PDF and automatically put them into their proposal generation tool, eliminating a whole bunch of steps.

(17:40):
Vanguard announced its client-ready summaries on its most popular market opinions. This summer, Jump had a really interesting announcement of a two-way integration between meeting notes and planning. You could pull the plan into your meeting prep tool, and you could also use the meeting notes to automatically populate updates to the financial plan. Wealthbox did something similar but extended it through to bringing portfolio data into meeting prep. This brings us up to date on the major announcements from the industry using GenAI tools, and I want to pause here and highlight two patterns. The first is that there's generally a nine to 12-month lag between a new capability coming into the developer toolkit and that capability showing up in a feature in a product. Think of it like that iceberg metaphor where what we see above the surface is a tiny fraction of what's actually going on underneath.

(19:00):
The other pattern is that the pace of change and innovation is absolutely accelerating because people start with a narrow set of tools, they build on those tools, and you achieve a multiplier effect. It's not like one new tool helps you do one new thing. Expanding the toolkit has that multiplier effect.

(19:24):
I want to share with you some of the innovations that have happened in 2025 that can foreshadow the announcements we're going to see over the next several months. The first of these is reasoning models. While the original language models were good at summarization and text generation, reasoning models can do much more complex multi-step things because they go through an intermediate reasoning process before they spit out their final answer. I think of this like the job a newspaper editor does; they edit each individual journalist's work and then they craft their vision of how all of the things fit together into the final edition.

(20:23):
Reasoning models brought some important new innovations at the prompt and response filtering layer. A key one was "planning." Planning helps deepen understanding of the user's intent by breaking the request down into its component parts. For example, if you asked your tool to prepare a meeting briefing, a reasoning model with a planning tool can break that meeting prep task down into tasks like: summarize my recent correspondence, go pull the financial plan, give me a summary of portfolio performance, and then aggregate it all back together. You get much more precision in how the tool is executing the request.

(21:18):
The last thing is memory. Memory is how AI systems keep track of their tasks and evolve over time. LLMs on their own have no memory, and so this capability helps them do a much better job of responding to feedback from the user about how they might want to see different answers in the future. Now that we're in this era where we've pivoted from Copilot assistance to autonomous agents, what can these agents actually automate? That really depends completely on the tools that are plugged into them. Technically, any application can be a tool, but some common examples would be using web search and a calendar tool to go find webinars about important product announcements, and you could have that set to be triggered by a certain move in the market. Another example could be combining an OCR tool with a ticket creation tool, so that you could do something like analyze client complaints about statement reporting and automatically create a ticket to send to your back office.

(22:54):
The thing everybody's excited about right now is Model Context Protocol, or MCP.

(23:04):
MCP is often described as a "USB-C connector" for third-party apps. MCP basically helps developers connect their AI tools to any third-party data source or application. Let's bring this together in a concrete example. What I'm showing you here is how you might build a meeting prep agent in Microsoft Copilot Studio. Now that I don't work for Microsoft anymore, I can tell you Salesforce, OpenAI, and AWS all have similar tools, but this is the one I know best. What's really important is that business people can start to do things like this. This meeting prep agent starts with a description—basically its MetaPrompt—which says what it's supposed to do and lays out the instructions of how it will do that. There are a number of different agents referenced here.

(24:21):
First, it pulls in an Outlook calendar tool. Then it calls a correspondence review agent, a financial plan agent, and a portfolio review agent. It's breaking down this big question into component parts so it can do it more effectively. The end task is to bring all this together into a briefing memo. Let's look at what happens when it starts to do its work. Here we see the meeting prep agent invoking the correspondence review agent. This agent on its own has its own specific description of what it's supposed to do. This is important because it keeps each sub-agent focused on its own tasks, and that's key to delivering better accuracy.

(25:19):
Now we see it move on to the portfolio review agent. The tool continues through these other agents and ultimately uses a reasoning model to bring together this holistic meeting prep output. What's important to note about this is that it doesn't just cut and paste the output from agent A and agent B. Like what I described earlier with editors, it refines the output from each agent so the complete narrative has each agent using the context from the other agents, providing a much more integrated and holistic answer. While a tool like that can save advisors time and help with better conversations, what's important is that it can be a foundation for an innovation flywheel. What you see here is our meeting prep agent, and we're plugging it into a vision for a practice management agent, which would include an agent that specializes in a life event, for example, the birth of a child.

(26:40):
You can see here that the correspondence agent, the financial plan agent, and the marketing agent are reused in the "birth of a child" agent. Each new agent you build gives you a faster time to market because you're reusing capabilities.

(27:02):
These types of tools can help us address new life events that we might be overlooking because they're too much work. Picture something like a job change. There are so many decisions to be made around a job change that have huge financial implications for clients. Imagine having an agent that helps you analyze a client's job offer letter. Think about the details that are in that: base comp, bonus amount and timing, stock awards, and matching provisions of their 401(k) plan. All of these things require advice and contribute to AUM growth over time.

(28:02):
I want to spend a few minutes talking about ROI. In my three years watching wealth and asset management firms develop their AI strategies, almost without exception, the way people decide on what use cases to create is they brainstorm 100 use cases and prioritize them in a chart. The problem with that is that it looks at each use case individually, but what is really going on is that those 100 use cases are actually 10 variations on 10 core use cases. The broader problem with realizing ROI is that firms are either choosing the wrong projects or not thinking about ROI correctly, and the two are related. If you focus on what might be similar about your use cases, you'd approach your tool development quite differently because you'd focus on what can be reused, and obviously, your ROI is going to go up.

(29:31):
The flip side of this is thinking about how use cases could complement each other. Something like that meeting prep tool is actually a combination of meeting notes, a portfolio review tool, and a financial planning agent. That's not one bubble on the chart; that's actually multiple bubbles sequenced in a very deliberate way that contributes to ROI. Finally, firms really don't think enough about integration. You get far less value from having three AI capabilities separated across three different apps than you will from having one AI tool that looks across all three.

(30:28):
The real issue with ROI comes down to not recognizing the importance of learning by doing. Obviously, we need formal training in AI, but that really should be just the beginning. What you really need to start with is giving people the right tools. Giving your advisors a tool that summarizes meeting notes doesn't teach them anything about AI because the tool is locked down. Tools like M365 Copilot or ChatGPT Enterprise, because they have access to so many different data sources and can do so many different things, are much better for learning by doing. Similarly, having all your people experimenting on their own in isolation will only get you so far. The huge value of POCs is that they provide structured opportunities to experiment that bring together both business and technology stakeholders to solve problems.

(31:48):
Of course, this only works if you're clear about your learning objectives upfront and you document them at the end; often, there isn't enough discipline around that. The third thing is just time. AI is not just a tool; it's a business model. It's like the shift from commissions to fee-based advice. There are so many complexities and interdependencies around a change like that, and it doesn't all come together overnight. You end up with a much better ROI if you shift your focus from thinking about individual use cases to broader programs.

(32:36):
Here are some key takeaways from three years of experimenting with LLMs. First, the LLM is just one component of an AI-powered solution. The quality and utility of results depend much more on the data you bring and the way you manage context than it does on the LLMs themselves. Second, innovation is accelerating because firms are starting to build platforms of reusable capabilities instead of individual, isolated apps. Third, think in terms of programs instead of individual projects. For the next two days, my advice to you is to think about these three things:

(33:54):
First, ask yourself what is truly unique about the type of advice you serve to your target client? Who is that client, and how are you differentiating?

(34:00):
Secondly, where might you find data sources, and how could you use your existing data sources to build agents that help you express that differentiation in a more consistent and scalable way? Finally, think about how you can bridge the gap between IT and the business. IT is great at building tools, but they rarely understand the big picture. Business understands the goals, but their thinking might not reflect what the technology can actually do now and how quickly it is evolving. I hope my under-the-hood look at what goes into building AI apps has helped you think about things a little differently. If you would like a copy of my slides, please connect with me on LinkedIn. I can also give you a list of all the materials and sources I used, just like a good AI tool should do.

(35:04):
Thank you for your attention.