AI Day: Financial Services | Salesforce

Hey, I'm Mark Sullivan, and welcome to Financial Services AI Day. Today, we're here to motivate you, educate and inspire you. But before that, we want to thank you to our customers, our partners, our trailblazers, and our communities all joining us here online from around the world. And especially to those tuning in for the very first time, thank you for joining us. When Mark and Parker founded Salesforce, they brought in a new technology model, a new business model, and a new model for giving back. Our 1, 11 model is about dedicating 1% of our equity, 1% of our time, and 1% of our products to giving back. Globally, we've been able to give back more than $621 million in grants to nonprofits and NGOs. Our employees have logged more than 8.1 million hours of volunteerism, and 54,000 nonprofits and NGOs are running for free on Salesforce. And probably the greatest thing is that we've inspired 17,000 other companies to join us in our One One One model. To that end, we are proof that you can do well and also do good. 24 years ago, we never anticipated that we'd be here and to have had this incredible growth. We just came off a strong first quarter where we delivered $8.25 billion in revenue. And we're on track to deliver nearly $35 billion in revenue this year with a 28% margin. We're a leader in philanthropy, innovation, and culture, all while being one of the fastest growing software companies in the world. We're also proud to be 100% net zero. We want to help everyone to their journey to get net zero. What drives this effort? At Salesforce, we emphasize our core values and how we are living those values each and every day in our company and in our communities. Our values guide everything that we do. Trust, customer success, innovation, equality, and sustainability. Meaning for all of you in financial services across wealth, insurance and banking. We know that trust is the foundation of any client relationship, from securing and growing their capital to protecting their assets, their homes, businesses and their lives. That's why we've built a trust framework for AI as a foundation for our AI platform to ensure you comply with the various financial services, mandates and regulations. Because AI fuels our primary vision here at Salesforce to help you connect with customers in a whole new way with a foundation of trust. And to that, let's take a look at where our AI capabilities are today. Salesforce is not new to AI. In fact, we've been pioneering AI for CRM with Salesforce Einstein for almost a decade. Einstein is our native AI that's baked into every product. It has been delivering more than a trillion predictions a week to our customers. There's no company that comes close to what we're doing in customer relationship management with AI. This is Einstein. This is AI for CRM. To spur all this innovation, we start with investing in world class talent, researchers, data scientists and Data engineers. Our AI research team has published 227 AI research papers and 300 AI patents. We've also acquired great AI innovators like RelateIQ and Prediction IO. From building Einstein to inventing completely new technologies such as auto feature engineering, auto feature selection and auto model selection, we saw where the world was heading and we knew we needed to invest deeply in large language models, or LLMs for short. And for all you developers here today, part of that investment is our development of five LLMs and two Transformer libraries. Not only are we proud to offer the world's number one CRM, but the world's number one AI CRM. We've brought AI into every app, every layer of our platform, every workflow and 12 industries, all with trust, security, privacy, scale and ethics at the core. And today, we're diving into financial services. We are already helping customers maximize their investment and accelerate time to value across banking, insurance, wealth management and corporate investment banking. How? By unlocking the power of AI across their entire CRM to create a tailored experience for customers, enabling every agent, advisor and broker to focus on what's most important to their customers, when and where they need it, transforming financial services organizations and transforming our business to lead with this innovation. But why now? Because we all know Generative AI is changing everything. It's the technology of any lifetime. It's coming at a pace that we've never seen before. Just think about this. It took ChatGPT just two months to reach 100 million users. That's faster than any other recent technology innovation of our time. This is a new innovation cycle and it's exciting. AI is redefining how businesses operate, driving new levels of growth and productivity, and creating new ways for you to interact with your customers. According to McKinsey, generative AI will have a significant impact across all industry sectors, having the potential to automate 60 to 70% of current work. Across the banking industry, for example, the technology could deliver billions in value. Banks, insurers, and wealth and asset managers can leverage AI for faster prospecting, automated underwriting, fraud detection, risk analysis, client insights and efficient service requests. Basically, every firm needs an AI strategy now. One is orchestrated safely with controlled prompts grounded with integrated real time data, while protecting customer privacy and complying with PII standards. Every leader I've spoken with is excited, but they're cautious too. They're cautious because there's a trust gap with this powerful technology. Company leaders want to embrace it. It's their number one priority. But their customers aren't so eager. Less than half of customers trust companies with their data. That's a challenge because data is the foundation of AI. When we talk about privacy, hallucinations, data control, bias, toxicity, these are terms used to describe all the things happening inside of AI language models. And we all want to move forward rapidly and have greater productivity. But we also need trust. And today we want to close that gap. We want to usher in a new era of trust. That's why we're bringing you Einstein, a trusted enterprise AI built for CRM. What makes Einstein different? It starts with the new Einstein trust layer, which allows all of our applications to harness generative AI for your most sensitive customers and business data. We talked earlier about data as the fuel for AI systems. We know it can be complicated to get your data organized to create compelling customer experiences. But that's the power of Data Cloud, our fastest growing cloud ever. Data cloud takes the challenging task of collecting and organizing your data across your entire AI CRM using our open ecosystem, leveraging your current data architecture strategy to create personalized experiences and business value. That means that we will be able to open many LLM ecosystems for our customers, allowing you to solve problems the way you want and need to. Now, to tell us more, please welcome Patrick Stokes.

Thanks so much. Hi everyone. I'm going to walk you through exactly how we do this at Salesforce. Now, to do that, we're going to get started with the question I hear from all of you all of the time, how do I trust generative AI? And perhaps more importantly, how do I achieve the productivity gains from generative AI while protecting my most important asset, my data and my customers data. Today we know how to do this because we know where all of our data is. Our data is stored in databases, it's stored in files, it's stored in spreadsheets. Now, what all of these things have in common is they have an inherent sense of location. We put data in a particular database and within that database, we choose the table and we choose the field. And if it's going to a file, we're choosing where the file is. There's always this sense of location to it, and on top of that location, we're able to install access controls or permissions. So I can say that one particular person in my organization can have access to this piece of information in this field, but this other person cannot. But in AI, it's completely different. In large language models, it's different because data isn't really stored, it's learned. And that's a very different concept. Let me give you an example. If I asked all of you what an apple is, you'd probably be able to tell me. But if I asked you to point to the location in your brain where the data about an apple is stored, you wouldn't be able to do that. And that's because you don't really store data about an apple in particular. Instead, what you're doing is, is your brain learns properties about an apple, and you're collectively taking all of those properties and using them to identify an apple. So, for example, you know that an apple is round. You know that an apple grows on trees. You know that an apple is red, but also sometimes green. It's this collection of different properties and different patterns that you bring together that define what an apple is. But because of that, there's no one location we can point to to create those access controls on top of that. So what are we going to do? How are we going to solve that problem? Well, the good news is Salesforce has been solving problems just like this since 1999, helping enterprises use their data while also protecting it. More than 20 years ago, you came to us and said, I can't possibly move my entire business into the cloud. I need to protect my data. And the only way to do that is to keep everything on premises. We said, no, we can help you. We introduced multi tenancy in our core sharing model, and we showed you how we can put all of the permissions and access controls on your data to protect it. Not long after, in 2016, when we made our first foray into AI with Einstein, we started building predictive models. We showed you how we can train those models without ever blending customer data across customers. Because your data is your product, it's not ours. And today we're doing the exact same thing with generative AI. But how are we doing it? Well, it all starts with a prompt. Now, a prompt is just a question. It's the question that we're going to ask the LLM or ask the AI. And you've probably all done this before. If you've opened up ChatGPT, you've had this experience, you ask it a question. Well, the quality of the response, you're going to get is directly correlated to the quality of the question. So let's pretend I'm an investment manager and I'm inviting my client Lauren to discuss some of our investment services. If I'm a salesperson, this might be something that I have to do dozens or maybe a couple hundred times a day. It's going to take a lot of time. So I'd like to be able to improve that workflow by getting an email generated for me. So I'm going to ask a large language model for some help. Now, I'm going to get a generation or a response back a generated email. And this email is two things at the exact same time. First, it's pretty darn amazing what a large language model can return from very little. A simple prompt. I'd like you to write me an email. And it kind of understands some context from that. And I'm able to get this pretty impressive email. But the second thing that this email is is it's entirely unusable. It's unusable because there's very little context here. It doesn't understand my business, it doesn't understand my products or my customer. This kind of looks like those LinkedIn recruiting messages we all get. Just not very usable. Not something I'm likely to reply to. So how do we fix this? Well, we fix it by adding more context by writing a better prompt. The good news is Salesforce has a ton of context about your business across your CRM, across sales service, commerce, marketing, and in data cloud. We know how your business runs, we know your products, we know your customers, we know your workflow and all of your metadata and all of that engagement data is streaming in today. Data Cloud is processing over 30 trillion transactions per month. So what if we could connect all of that data to the large language model? Now, I know what you're thinking. You're probably thinking, great, I need to train the model to understand my business so I can get a better response. But that's the trick. You don't actually need to train the model. You just need to go back to that prompt. So what if instead of a simple prompt, like I'm an investment manager, what if we started adding more context? So let's add some context about me. I'm Patrick Stokes. I'm an investment manager at Cumulus Bank. Let's add a little bit of information about my business, about my customer, Lauren Bailey. She's been a customer for seven years. Let's add some real time data. Now, I'm sure you've all seen how large Language models are only trained on data that's up to a year old. But we can add real time data in this. Now, if we add all of this into our prompt, let's take a look at what we're going to get. This time we get a much, much better response. But we have two problems, two things that we still need to solve. First, we have PII data in here. So we need to mask that. We don't want that personally identifiable data about Lauren going back into the model. We don't want that to be learned, so we're going to mask it. Now, let's take a look at our generated email. I can take this and immediately start using it. If I could do this a couple dozen or a couple hundred times a day, this would dramatically improve my workflow as a salesperson. But remember, we had two problems. You see, I've got a bunch of information about my business here and I want to protect this. I need to know where it's going. And that is precisely why I'm so excited to announce the launch of the Einstein Trust layer. The Einstein Trust layer creates separation. It separates all of your CRM data from the large language model. It enables you to securely blend all of the context found in your CRM and data cloud into a prompt in order to get a generative response from in a safe way. And it does that using a number of methods like secure data retrieval, dynamic grounding, masking, which we just saw, toxicity detection, auditing, and zero retention. So let's take a look at exactly how that works. Follow the flow of data that you see on the screen and you'll notice we start in our CRM apps. Now, you may have noticed before that the prompt we wrote was pretty long. In fact, it was longer than the email itself. Now, you don't want your users writing prompts to take longer to write than the email that we want to get generated. Well, the good news is you don't have to. Salesforce creates those prompts for you. That's the magic of how this works. And so it starts from our CRM applications. Imagine we were in Sales Cloud looking at an opportunity and we wanted to generate an email by just clicking a button. That button is going to take the prompt and it's going to securely retrieve data. It's going to retrieve Data from your CRM from Data Cloud or from external systems via MuleSoft, and it securely blends all of that data in a process called dynamic grounding. Then we're going to mask all of the PII data just like you saw before. And finally, we're going to send the fully compiled prompt out to our large language model via our secure gateway. Now, hold on to that secure gateway thought for a few moments because we're going to come back to it. Okay, so we're going to hit a large language model and we're going to get a response. Our prompt has now turned from a prompt to a full generation, and we're going to start our path back to the applications. The first thing we're going to do is we're going to check the response for toxicity and then we're going to create an audit trail. So we're going to audit it, take some metadata about what the prompt was, who the user was, what the context was, and we're going to store all of that in an audit trail that we can go back and look at later. Then finally, we're going to take that generated response and hand it back to our application. So the user experience isn't writing a prompt, it's clicking a button. This all happens transparently and the user simply gets a perfectly usable email that's ready to go. Now let's think back to the secure gateway I mentioned before. We need that. So we can build all of this in an open way. Because we may need to use different large language models depending on the use case, since many customers have different data residency requirements for their data. So there's three ways to do that. The first is with our shared trust architecture, which is a way to access large language models across the Internet. We pioneered this with our incredible partner OpenAI, where we're hitting their secure gateway without ever having any of your customer data stored outside of Salesforce. But you can also host models inside of Salesforce. So if you want to bring models directly into Salesforce and host them in our private vpc, you can do that as well. And finally, if you're already making investments in your own models, you can host those in your own infrastructure via our byom, or bring your own model capability and connect them via Amazon, SageMaker and Vertex AI. So, three different deployment strategies. Now let's dive into a demo and show you how to build these prompts, how Salesforce is building these prompts, and how you all can build these prompts as well. To show us how, please welcome Elizabeth Markey. Over to you, Liz.

Thanks, Patrick. Patrick just showed us the power of prompts to generate highly relevant content. But creating these prompts is time consuming. Now, rather than having users write those prompts, what if the prompts were created for them? What if every admin were a prompt engineer? Well, now they can be. With the new Prompt Studio. From here, you can select the type of prompt you want to create and the place where you would like your prompt to appear. These prompt templates come with a lot of foundational work already complete, so they're easy to use. We have email to help you generate personalized emails to your customers, Field population to help you generate data fields like Descriptions. And the sidebar gives you suggested actions like getting important leads. But don't forget, we're building this to be extensible so that you can use these prompts from anywhere, like in Flow and Apex. All right, let's get started. With an email, we give the template a name and add a description. This takes us right into Prompt Studio, where you can create a prompt to generate emails to fit the specific needs for your organization and industry. Here we can set the language, the style, and even the LLM provider that we would like to use. With the Einstein Trust layer, it's easy to create the perfect prompt and you can focus your time on customizing. Let's do it. What makes the Prompt Builder special is that it has access to all of your metadata and allows you to ground the prompts directly with your CRM data. Here we have user and contact, and it integrates with Data Cloud. This allows you to personalize the offer delivery method to your customer's preferred channel. Once you're ready to test your prompt, you can select a record from the dropdown and then generate a response. First, you get to preview the prompt with all those fields populated from Andrea's record. This is how we verify the data that is used to create this prompt. And when you click generate, boom. Just like that, the Einstein Trust layer has generated a response. You can see that this response is personalized with the contact and user info. The Einstein Trust layer has given this a toxicity rating of harmless. This is the powerful capability that enables you to become a pro at prompts in no time. Now that it's ready, let's activate this prompt. It's that easy to build out these experiences with prompts to and with that, back to you, Patrick.

Thanks, Elizabeth. It is amazing to see the power of a prompt. So you might be thinking, where do the prompts go? Well, where they go is into our applications. And that's because Einstein drives productivity across your entire company. With productivity for any workflow, for creating prospecting emails, for sales reps in SalesGPT or knowledge responses, for service reps in Service GPT or for creating landing pages for marketers and marketing GPT for writing product descriptions, for commerce pros and Commerce GPT for helping everyone explore data in Tableau GPT and for summarizing across your entire company in Slack GPT. So let's dive into another demo and show you exactly how all of this comes to life and how these prompts are used across the entire platform to show us how. Please welcome Sanjana Parulaker. Over to you Sanjna.

Thank you Patrick hi, I'm Sanjna and every day I talk to customers who are excited about generative AI. But you know what makes them totally fall out of their seat? Generative AI in the workflows they use and rely on. I'll take you through three of my favorites Sales, Service and Productivity Sales reps spend day in and day out communicating with customers and prospects over email. But crafting personalized outreach emails is time consuming and tedious. The Einstein Trust layer is built into the sales user experience so every rep can get assistance in writing these emails and automatically bring in the right context from their CRM data. The prompt template surfaces as a simple button for a sales rep to write an outreach email. It's so simple for the sales rep and so powerful behind the scenes. Each prompt passes through the Trust layer so you can ensure your data is safe and no data is ever retained outside of Salesforce. Now this is especially important when dealing with sensitive data like strict privacy regulations or medical information or credit card numbers. Sales reps also spend their days on the phone with customers. And you know what's even more tedious? Writing call summaries. The Einstein Trust layer is in this daily workflow too, and after a call is done and transcribed, it can be automatically summarized into the key purpose, action items, and the moments that were discussed. So sales reps can now spend more time talking to customers closing business, and worry a whole lot less about managing all those mundane tasks. So you've seen what a game changer the Einstein Trust layer is for sales. What about service? Let's take a look. Every service rep cares about improving their resolution time, and every minute spent searching for knowledge articles, product recommendations, and writing summaries is a minute that's taken from helping another customer. The Einstein Trust layer is built directly into this flow of work for every single service rep. So as they have conversations with customers, replies are then recommended to them that are personalized to their end customer. And as that conversation goes on, the service rep also gets assistance with the next best offers for the customer so they don't need to search for what's new in their product catalog and the rep is always in control so they can accept, edit or decline these recommendations. Now, Next Best Action has been a hit with customers for the last several years, but pairing generative replies with predictive product recommendations saves reps so much time and makes for happier customers. Now, once the case is closed, reps also need to summarize their cases, and the Einstein Trust layer can help here as well with the click of a button. This is a simple step, but an important one with this new knowledge article now in our knowledge base, similar cases in the future can be automatically deflected, saving our reps even more time. So service agents can now spend more time helping customers, whether that's a patient or an account holder, and worry a whole lot less about managing the outcome. So we've seen what a game changer the Einstein Trust layer is for sales and for service. But what if we need to leverage the collective expertise of an entire organization to drive better collaboration and productivity? That's where slack comes in. This is a case warming channel where we've brought in a variety of people across the organization to solve a particular case. And this channel is equipped with rich detail from service cloud about the case itself, the account, and the source of the issue. But it also has an automated workflow to update this case in other systems like Jira. Great service isn't just about finding knowledge, it's about creating it. So as the chat goes on, various folks can jump in to provide important context that should be documented and shared to easily handle common questions spanning everything from loyalty programs for shoppers to public health program eligibility for residents. So Einstein is at the ready and with a click of a button can summarize the case form and then post it as a knowledge article for others to benefit from in the future. Now that's how the Einstein Trust layer is changing the game for sales, service and productivity. Back to you Patrick.

Thanks Sanjna. It's awesome to see how all of these use cases are coming to life. Now there's one last concept I want to leave you with. These LLMs and AI in general, how do they get better? How do they learn? And this is where Salesforce is different. You see, most LLMs learn from usage. When you get those responses back from ChatGPT, it might ask you was this useful? And you give a thumbs up or a thumbs down. But with Salesforce we do things a little bit differently. Salesforce is where you entrust not only your customer data, but also your customer outcomes. And Einstein learns not just from usage but but from your outcomes like your deals closing in sales cloud, your service cases being resolved, your marketing email open rates, and having your company conversations in Slack. So for usage, when a sales email is generated or a service agent response is suggested, did they use it or did they edit it? But for outcomes, did the generated sales email close the deal? Did the marketing message reach the open rate goal? Did the service response close the case? Signals like these are unique to every company. And this ensures that every customer will develop the best models for every use case that's specific to their industry, their organization and their task. Now, as you can see, at Salesforce, we have tons of innovation happening and we're excited to bring you even more of it in the coming months.

Thanks, Patrick. Okay, who's ready to take all this innovation to their businesses? Because we've hit the hyperspace button on trusted Generative AI. Generative AI is changing fast and we're helping our customers meet the moment. With 16 AI first releases out right now, we've mobilized the entire company around AI to get this incredible technology into the hands of our customers. With all this innovation, we need everyone to be AI ready. And that's where Trailhead comes in. It's our free online learning platform that lets anyone anywhere skill up on Salesforce. We have 35 AI badges on the way, plus a generative AI certification launching at Dreamforce next month, which means any trailblazer anywhere can learn to skill up for the future. This is just the beginning of where we're going together with Salesforce and Einstein. From all of us here, thanks for tuning in.