How Unified Knowledge Fuels Generative AI To Improve Customer Service | Salesforce

Are you still looking to see how AI can help your service organization? Our virtual event, Smarter Service Starts with Knowledge helps service organizations get to the AI enterprise.

That's right. We are going to be unveiling all the details about our latest innovations in AI for service and we're going to walk through some great hands on demos of the product in action. We're even going to be speaking with a customer who's already using it with great success.

The data that comes in through Salesforce gives that 360 degree view to the agent. We want to empower the employees, remove all of the obstacles.

I'm actually having fun solving cases with AI auto responses. I love using the Einstein One platform. What the generative AI can now do is they can take these activities and then automatically recommend other activities that they know these guests will be interested in.

It's fulfilling. They really expect that we know everything that has happened between the brand and themselves. Every interaction we have with our customers is an opportunity to build trust and loyalty with them.

Without further ado, I think it's time to get to our first guest. So he's our executive vice president and general manager of Service Cloud. So please give a warm welcome to Kishan Chetan.

Hello everybody. How are you doing today?

Hello, Kishan. We're doing great. Thank you so much for joining us. We know you have some big news to share with us today, but before we get into that, we first want to talk to you about AI because it's very top of mind for everyone. So if you could please just tell us, where is Salesforce on this AI journey and how have we really adapted with the AI era?

Yeah, I mean, look, Salesforce pioneered AI in CRM more than 10 years ago and service was a key part of that story. Story really to help kind of drive service agents be a lot more productive. That was a big part of the investment with our predictive AI. But as the world transformed into generative AI, that's where Salesforce as a company pivoted with our Einstein One platform. And service is a key use case in that. And Einstein One platform is super crucial to driving the right AI for companies across service, sales and marketing use cases.

Now that's so true. Like you said, we've been in this game for a long time with predictive and we really had to make a strong pivot to generative. And we're seeing a lot of success with our customers even. But I want to double click into something that you said in that last portion, that Einstein One platform. So what exactly is The Einstein One platform. And why does it matter to service organizations?

Yeah, I mean, Einstein One platform underpins everything that we do. It consists of three key parts. One is really all of your applications, your customer360 application for C sales and service and marketing and many others. It consists of data cloud, which is where you bring all of your data together in Salesforce, outside of Salesforce, so that you can use that to go drive the right AI. And then it contains Einstein, which is the platform for building your predictive and generative models with our prompt builder, co pilot builder, et cetera. And on top of it all sits service cloud, sales cloud and in service, all of our GPTs and all of our predictive is built right on top of that.

Yeah, well, it sounds like the Einstein One platform is really playing that critical, crucial and big role in helping service teams mainly really deflecting more of those cases and ultimately reducing costs, which we know is very top of mind for service leaders. So while this all sounds great, we also understand from talking to customers that achieving this isn't always as easy as it sounds, especially when AI is involved. So I'd love to hear more about that from you, what your thoughts are on that.

Well, yeah, I want to add a little bit more to that, Kishan too. I was talking to a group of service leaders recently about AI and when I was talking to them, I asked everyone, okay, who wants to use AI in your business? And everyone's hands went up. But when I did the follow up question on who has implemented AI successfully, only a couple hands went up in the audience. So we noticed they're hitting a roadblock and we realized the biggest problem is their data. So what do you have to say about that piece?

Yeah, I mean, look, every AI conversation I have with a customer becomes a data conversation. Data underpins AI. I mean it's really the oil for this revolution, right? So it's really crucial. And there are two types of data. It's like the structured data that's there in Salesforce or outside. Then this unstructured data, documents and websites and all of those together are crucial for customers to get real benefits from the service. And that's why data is crucial and that's where we can really help customers.

That's key Kishan, and appreciate you sharing that, especially about the unstructured and the structured data and bringing that together. I know that that takes time. It's a great help for organizations to be able to bring those third party systems that have all of their knowledge resources together, all their data together. And which love to learn a little bit more from you about what those third party systems look like. And particularly when it comes to our latest innovation which is Unified Knowledge, how do we connect with those systems and how does that actually work?

Yeah, fantastic. Look, the reason why, and we are super excited to speak about Unified Knowledge and the reason why we did Unified Knowledge was really to bring this third party data together. Data that's the there in a file share data that's there in a cloud share in different documents in your JIRA systems and using that, using a partner zoom in in this case to bring that data together in Salesforce so that the agents can use that data and become more productive. The AI can be grounded in that data and customers can automatically discover information. So the key innovation here with Unified Knowledge that we are very excited about launching is bringing this data together from everywhere, all of these third party sources and then using that to drive better productivity, better deflection for our customers and making the agents more productive.

Yeah, absolutely. Oh man. Unified Knowledge is really going to help service organizations get all that data together. And that's super exciting. Now you were kind of touching on something there with like the things that it will enable an organization to do and I want to tap into some of those use cases so we have a lot that we can dive into. But maybe which one is your favorite? Which one do you think is going to be the game changer for service organizations?

I don't want to pick a favorite because all of them are super important. But I think, look, one of the things that I hear and in fact when we did our we worked with customers through the beta program and in the early adopters and if you looked at that one of our customers was basically took something where they were responding to their customers in service replies. It took them 15 minutes earlier, it would take them less than a minute. That's more than 90% productivity. That's the type of benefit agents are seeing when they're responding to customers. And when you go into a customer and the customer is trying to find information on a knowledge portal, not having to get an agent is type of the best type of service. And that's where it can be really powerful as well. Each of these use cases is really valuable to both the customers and the service organizations. So it'll all be great.

I love that. And yeah, with customers being able to find answers on their own quickly, that's a game changer too. So. Well, I think we're going to wrap things up with you. I mean you've Given us such good knowledge. May I say, you've given us a lot of knowledge today. You're helping every customer become that AI enterprise. So thanks for being here with us today.

Thank you. And knowledge is key.

Awesome. Knowledge is totally key. And you heard it here, folks. Unified Knowledge is officially an open beta. That's a big news.

That is so exciting.

Yes, you can get your hands on these innovations at your organization today. And those generative AI features that Kishan mentioned. Einstein Copilot, Search answers, Knowledge Answers and Bots. Those are GA as well, or they're going GA in the next month. So get ready.

This is huge. And again, so exciting to just hear all these announcements. And now that we've learned about Unified Knowledge, let's actually see it in action with a product demo from our very own Director of Product Management, Leon Kempers. He and his team have actually built Unified Knowledge. So you. He's really the perfect person to show us how to get started.

Hey, Gillian and Cristina. As you two know, we've been working on Unified Knowledge for over a year. So it's so special to finally show off this product in the studio today. And I've got my laptop right here. So let me show all of you watching how you can use Unified Knowledge to connect all of your knowledge sources and power AI driven customer service. All right, let's dive right into my favorite part of Salesforce, the setup. With Unified Knowledge, you can connect systems like Google Drive, SharePoint, Ixiosoft, and even external websites. In this example, I'm bringing in knowledge from Confluence and Google Drive. Now, for every system I connect, I can configure to which audiences I want to show the ingested content, which record type I want to use. And as you can see, I can even map the labels from the source systems to to data categories and field values. In Salesforce, you've told us how important it is to maintain a taxonomy in your source system. And we've heard you loud and clear. Now, once I've set this up, my Unified Knowledge base can be exposed across every customer touchpoint and used to ground our generative AI products for service. So now that we've set up our knowledge sources, we can use them to help our customers. Or even better, we can let our customers help themselves. In this example, I'm someone who has a few questions about garbage pickup services offered by my city, Pacifica. I go to my city's website and enter my question in the search bar. And what you'll notice right away is this generative answer here at the top now, in the past, I would have needed to scroll through a large list of articles to get my answer and then find the right section in the article. And let's be honest, who has time for that? I got places to be, but with Einstein and Unified Knowledge, I get an instant answer that is grounded in Pacifica's knowledge from Confluence and tailored to my specific issue. Because I have some additional questions, I've started a bot conversation. This generative bot is also able to provide generative answers that are grounded in your unified knowledge base. And then when I eventually need to escalate my question to a human agent, it can help me with that too. And that's the power of Unified knowledge. You don't have to spend months connecting data across your organization through complex middleware. All you have to do is enter your credentials in Salesforce and you're off to the races. Your customers can enjoy faster and better self service and you can save valuable time and money by deflecting cases, letting your agents focus on the most complex issues. Speaking of agents, I can't wait to show you how Unified Knowledge is helping them be more productive. But first I'm sending you back to Christina and Julian.

Thank you so much Leanne. That was a great first step.

It was awesome. I loved it.

Loved it.

Leanne. It was so great to see, particularly as a customer myself who loves to self serve. Great to see that Unified Knowledge is really helping change the game and help me find answers faster in the right time when I need them. Awesome first demo. We will get back to Leon in just a little bit with another demo.

I cannot wait. And as you've heard everybody, Unified Knowledge is an open beta, which means it's already making a positive impact at so many companies. And guess what? One of those companies is Sonos. So let's take a look at what their Global Vice President of Customer Success and Experience has to say about it.

My name is Dharam. Dharam Rai. I'm the Global Vice President of Customer Experience at Sonos and and Sonos, as you may know, is a brand that provides joy and happiness in terms of sound experience to customers worldwide. We have over 15 million customers and also another 10,000 or so businesses and system integrators who use our products. So we're in the B2C and B2B space as well. We are a Salesforce shop. So sales and services information is in the Salesforce system. However, the information that's required by our agents in the front lines and our support staff and in the back end, you Know, our backend sustaining engineering folks are in different locations focusing on improving customer engagement. And how do you improve customer engagement? We are looking at proactive, predictive and preventative kind of a service that we want to provide to our customers. We are looking for ways to provide self service so that customers can help themselves. And also we really focus on time to music customers. How do they get to the music quickly? And that depends on how we actually engage with our customers and provide solutions to our customers. So for all of that, unified knowledge is one thing that's going to help us.

Why?

Because our pain points are that data is in dispersed locations. And it is always a challenge for our frontline folks to actually scramble through looking for the right information at the right time for our customers. So when we're trying to provide a real time answer to our customers, the agents in the frontline are having to look at different dispersed locations and look at information that was provided through training and through knowledge and through these technical documents. So all of that had to be unified somehow because otherwise we are scrambling when we are on the call with the customers, which impacts our, you know, resolution time and subsequently just, you know, response time and hold times and all our KPIs get impacted. And of course ultimately that, like I said before, improving customer engagement is about how do we actually make it seamless for our customers so that we reduce their effort that it's required for them to use our products. All of that now can come through unified knowledge. And we've been fortunate to be using the unified knowledge in a beta program. And the success we have seen so far is great because all these resources and data points can be unified and it goes through the same system. And in a quick manner, we are able to turn it around and provide answers to our agents. It actually really creates a good ecosystem because our frontline agents, if they're happy, they're more likely to stay back with our Sonos ecosystem and they will see that they have a career path, they're learning. So it's all in all, not only that the customers benefit, but the agents benefit. And then we provide consistent, repeatable service to our customers. Which means that our agents are delivering the same kind of Sonos experience to our customers all around the globe. We have our agents in five different locations all around the world. We want to make sure that every agent that's talking to our customer provides the same level of service. So consistency, repeatability, and the things we talked about earlier in terms of proactiveness and the agents can focus on actually looking at the Unified Knowledge and providing the right kind of solutions. And all in all, it increases the productivity of the organization. I can't wait to see the benefits of this across the organizational boundaries rather than just limited to the customers. Because Unified Knowledge is what's going to help us actually put the pieces together.

Thanks for sharing the Sonos story, Dharm. That was awesome. And guess what? I think I'm ready to buy some Sonos speakers. They're amazing for sure. But I want to touch on one thing. Dharm really hit something like right on the head. The impact of Unified Knowledge and Einstein and the impact that they have on agent productivity. When we think about the agent experience. I remember Christina, we were talking to a customer last week asking them about what's the impact you've seen with your agents? Well, they actually tried to recreate or create their own version of Unified Knowledge, pulling in all of their thousands of PDFs that was taking them. I think they said well over 30 hours, if not more than that. When they implemented Unified Knowledge, that only took 30 minutes. That's a huge savings.

Such a savings.

Exactly. So we're seeing lots of success like we talked about, but I think it's time we get back into what you really want to see. A demo. So let's go back to Leon so he can show us more about what Unified Knowledge can do to help those service teams that you have.

So in that story you heard about the agent experience, let's dive into that a little deeper. We already saw how Unified Knowledge helped us to make self service more effective with AI. And that's great because now you can deflect simple cases. But it also means that the conversations that reach your agents are more complicated. They need assistance to ensure they're ready to tackle these complexities. So let's look at another example and this time see how Unified Knowledge helps your agents solve their issues. Alright, I'm going to pick up where we left off and step into the role of an agent. As soon as I accept the incoming chat request from the customer, I see a Serves reply right here on the right side of the screen that I can edit or directly send to the customer. Now, to create this reply, Einstein leverages CRM data and Unified Knowledge to auto generate a response based on the context of the chat conversation. That is what sets Salesforce apart. We bring CRM, AI and data together to ensure that responses are trustworthy, specific to the customer and in line with the company's products and processes. When the conversation ends, Einstein automatically summarizes the interaction and wraps up the case. Think of all the time your service agents can save. Einstein is shaving minutes off of every conversation. So we just saw Einstein working on a chat conversation. But what happens if a customer contacts us over email instead? Actually, we can use that same service reply technology. Einstein pulls from the Unified Knowledge base to suggest the most relevant articles. The agent can use service replies for email. And let Einstein do the hard work of drafting an email that's grounded in the case data and unified knowledge. And now, because I learned something new during this process that hasn't been captured in the knowledge base yet, I want to make sure I create a new article. Normally, drafting that could easily take anywhere up to 30 minutes. But with Einstein knowledge creation, Einstein will draft a new article for us based on the resolutions that are reached in the service interaction. As you can see right here, all I need to do is review the draft, make edits and submit it for review. And of course, that new article will be fed back into the Unified knowledge base so that the next time someone encounters the issue, we can automatically resolve it using AI. So you're seeing that agents get help to reply faster to their customers, wrap up their cases with key information, and even have knowledge content automatically generated to help customers in the future. Even just one of those would have been a huge win for agent productivity. But putting all of them together, they're going to transform the way you deliver service. And with that, it's time for a final demo. We just talked about unified knowledge and AI making your agents more effective. But what about your mobile workers in the field? The field service technicians who are installing our Internet and repairing our dishwashers. They need quick and trusted answers more than anyone while they're on site with a customer. One of the most common complaints from mobile workers is that they don't have all the resources at their fingertips to resolve on site issues quickly the first time they're on site. So let's see how Unified Knowledge and the Field service mobile app can help them with that. Alright, I let my service agent days behind me and I'm Pacifica's newest plumber. But it doesn't matter that I'm new to the job, because my job just got so much easier with the field service mobile app. When I arrive on site, I quickly pull up my phone and as you can see here, I start chatting with Einstein Copilot on the work order to understand how to proceed. Einstein processes my natural language, queries and equips me with the right information instantly. As you can see here, Einstein's response on this water meter installation is grounded in a PDF from Google Drive using Unified Knowledge. This is a game changer for technicians. I didn't need to open up a separate Google Drive application or dig through the PDF file itself. Instead, I'm directly presented with the right information from the PDF to completely tailored to the customer problem I'm here to solve. And that is how unified knowledge is going to transform the way your organization delivers service. You saw that by grounding your generative answers and replies in your organization's knowledge base, we add a layer of trust and ensure that your generative content is specific to your organization's processes. With the Salesforce platform and Unified Knowledge, all of your data powers Generative AI and Einstein Copilot to increase productivity for your agents and and most importantly, make your customers happier. So with that, I want to thank you for your time and I'm going to send you back to Julian and Christina.

Amazing. Thank you so much, Leon. Weren't those demos so fantastic?

They were great. I'm glad Leon is the expert here and he really did a great job.

He really did. Thank you, Leon. Now, as you can see, Julian and I have moved down the table a little bit because we are excited to bring our panel of unified Knowledge experts here to share some really great best practices and getting started with a great foundation. So first up, we do have Megan Myers from our service Cloud team as well as Ellen Woodcock from our Einstein product marketing team. And Megan and Alan are AI and data whisperers. So very excited to have them here. But first we're going to pass it over to Devra Stesenberg, who's also on our Service Cloud team and she is our resident knowledge expert and is going to be joining us remotely today.

So.

So Deborah, take it away.

Thanks, Christina. Hi everyone. It is wonderful to be here today. We have seen a ton of great content and learned how unified knowledge fuels generative AI. So exciting. But I wanted to take a step back and talk about the critical component that grounds all of this, which is the process of setting up and maintaining your knowledge base. I mean, you have to have your knowledge base, right? So with generative AI, the quality of your knowledge base in the content has become even more important than ever. And by properly implementing knowledge, you can set yourself up for the future success we've shown today. So let's get started. We're going to discuss the phases of setting up your knowledge base. There are three phases for the process for this process with recommendations in each one and you don't have to Complete every single one of these steps. These are just some of our recommendations, but these phases will give you an idea of how you can get started, how you can start thinking through your knowledge strategy. All right, so let's start with phase one, which is just getting started. First, you want to think about your Personas. Who are you, who are you going to have involved in creating content? Who are you going to have involved in using that content? It's really important to think that through as you're planning out the strategy for your knowledge base. Who's going to be leveraging this and where are they going to be leveraging it? Is it, you know, specifically on self service websites? Is it through AI? Is it internal audiences or external audiences? And then you want to look through and think about are you going to be using external sources of trusted knowledge? We've been talking about unified knowledge today. So do you have, you know, documentation, repositories or other sources of knowledge that you think would be really beneficial to have in your AI grounding? Then you want to have a plan to create and maintain your content. So it's not enough to just create it once you have to maintain it. The quality and content accuracy is going to depend on you having a good program in place to keep it up to date. And then if you're just getting started with knowledge or you want to just sort of get started with your AI program, one of the best ways to start is think about what are your top contact reasons and make sure you at least have content to address those areas. You can create short and, you know, modular articles so they're very focused on the specific topics or questions that you want to cover. And keep the taxonomy simple. Don't make it too complicated. Make sure you're talking in the language and you know, the way that your audience would describe your problems, describe the questions that they would be asking, and then finally define success metrics. It's important to think about what are you really trying to get out of your program when you're getting started and then how that will evolve over time. For phase two, you can then think about how to evolve. So maybe you've gotten started, you've pulled in some content sources through unified knowledge, you've created some content maybe in your salesforce knowledge base as well. You've gotten it connected up to AI, I hope, and then it's time to start thinking about what do you do to continue to evolve that so you know, one of the best recommendations I can say is to create what's in demand. You could have your audience create a lot. You could have your, your internal teams create a lot of content that isn't necessarily what your audience needs. So think about what they have been asking, what kinds of questions you're getting through self service, what kinds of questions you're getting through your cases. And then integrate external sources of trusted knowledge. We've been talking about this with unified knowledge. Maybe there's some additional sources now that makes sense to bring in. Using Knowledge Centered Service, which is also called KCS from the Consortium for Service Innovation, is a great way to have a full end to end methodology to guide you on how to build a knowledge program from the beginning and infuse quality in it throughout. So I highly recommend taking a look at that overall methodology. Salesforce is supportive of knowledge Knowledge Centered Service and maintain your quality program. We talked about this already, but it is really important to keep up to date your content. You know the age old saying garbage in, garbage out, right? So you really want to make sure that you have accurate information and have a way to keep that up to date and then refine your metrics. As discussed before, as your program evolves, you may want to measure different things. You may have different AI capabilities that you're now using. You may want to be looking at different ways that your overall, overall program is driving success for your business. And let's take a look at phase three. Now this is kind of the grow final phase, but you know, you may look at more opportunities to get value out of your knowledge. I highly recommend having one centralized knowledge base and using it for many different use cases. You know, you don't need a separate knowledge base for self service versus your internal teams. You should have one knowledge base that is using the the right content for the right use cases and that includes as you use it across different generative AI features. And then you know, finally just again looking at your metrics one more time. You know, as you bring in new experiences, you're going to want to look at how you evolve that. So by accomplishing all these phases, your organization is going to have very meaningful data built for both your customers and your agents to fit their needs and help you reach your goals. So with that I want to just turn it back over to Julian and Christina. Thank you so much and looking forward to continuing the conversation in our Q and A.

Awesome. Thank you so much Deborah. I love how you broke those down into really digestible pieces for us all to follow. So easy. And I know you're going to stick around to do a Q and A soon. So don't leave the zoom, please. We need you and audience too. If you do have some questions as a reminder, feel free to drop those into the chat as well. Deborah, you built this really nice foundation for us and we want to build on that. So Megan and Ellen are going to show us how to do something that's fairly new to most of us, and that's how to ground AI with knowledge and other data sources. So, Megan and Ellen, you are up.

Awesome. Thanks so much, Julian. So, Megan, we've talked today a lot about grounding your AI, but we haven't really taken a moment to take a step back and unpack exactly what grounding is, and particularly in the context of deploying AI for enterprises.

Yeah, absolutely. And I want to start with a framework that we should all be familiar with as consumers by now. Most of us, if not all of us, have played around with ChatGPT. Whether you're looking for support planning your next family vacation, or help writing a short poem because you're not feeling particularly creative that day, or even a simple recipe using the limited ingredients that you have in your pantry, the use cases are really seemingly limitless here. And what essentially you're engaging with is a three layer stack. And at the top layer we have the ui. This is the window where you're literally typing in your instructions in the form of a prompt, something like write me step by step instructions for how to bake cookies with flour, sugar, butter and eggs. In that same UI window, that's where you're going to receive the response from ChatGPT. And ultimately, what is powering this magical experience is a large language model, which is that middle layer there. Now, ChatGPT has their own proprietary foundation models like Turbo 3.5, but of course, there are other consumer applications out there that use different models like Anthropic Claude or Google Gemini. And what is actually making these responses possible is that they're using vast amount of data that's found on the Internet. And of course, the more data you pump into it, the smarter and more sophisticated it will get. However, if we shift this framework from the consumer to the enterprise, that's really where we see two key challenges arise. The first challenge is that these large language models lack context about your business and your customers, and that's because they don't have access to your business data. And the second challenge is that the Internet is truly a wild and crazy place. Not sure if you've been there recently, but there's a lot of inaccurate, outdated and even harmful information out there. So it's reasonable that deploying AI in the enterprise is challenging at the moment.

Yeah, Megan, you know, you're absolutely right. These models are incredibly sophisticated and especially over the last year, have gotten even more and more powerful. But I kind of like to think of them like a PhD student right out of school. Right. They're really, really, really smart, they've had a ton of training, but you've just hired them and they don't know anything about your business yet. So you have to give them that context about your business to actually put them to work and produce results for you that are actually really useful and can help drive your business. So let's talk a little bit about how to give LLMs access to your business data in a way that's safe and secure. So essentially, if we were to boil it down, there are basically two different methods for giving an LLM context about your business. The first is something called fine tuning. So let's say your LLM is a world class chef. This chef has been trained in a bunch of different cuisines, knows a ton of different recipes. But fine tuning is basically saying, all right, chef, I want you to focus on just this one type of cuisine or just this one type of recipe. I want you to get really, really specialized. So that sounds really good in theory, right? Well, the challenge is that it's really time consuming and expensive most of the time it requires hiring an entire team of data scientists. And although fine tuning might be required for certain types of really specialized industry use cases, for the vast majority of use cases our Salesforce customers are looking to deploy today, it's not really required. So that leads us to the second technique for giving a large language model context about your business, which is grounding. And we've already been talking about that a lot today. So what makes grounding so highly effective is the fact that it can access your live corpus of data. So we talked a little bit about, Devra talked a little bit earlier about knowledge articles and getting your knowledge base in order. But we can also ground on many other different types of data things like CRM, data account and contact information, any sort of external data source that we can pull in through data cloud. All of those can be data sources from which we can ground. And the amazing thing about grounding is that we're pulling from the live data sources. So think about how often contact information changes or how often you're updating your knowledge base. Right. You want to be able to access the live data, which is what makes grounding just so great and typically more of a better time to value than fine tuning.

Yeah, Ellen, I want to kind of pause there and double click on this idea of grounding. Right? We've talked a lot about it, you mentioned that. But really, what kind of data should we be grounding on? And it's pretty simple. You should be grounding your AI on information that you you trust. Information like your own knowledge base or your own business data. But we would be remiss if we didn't say or acknowledge the fact that like some of your knowledge and data doesn't always reside in Salesforce, it could reside outside of Salesforce. And that's why solutions like Unified Knowledge and Data Cloud are critical for successful AI. With unified knowledge, we're bringing together your Salesforce knowledge along with your external knowledge from sources like Confluence. And we're able to ground AI in your entire knowledge base, thus generating responses and presenting them to agents so that agents can trust and feel confident that what they're seeing is actually accurate and relevant. We also include the source article, so there's an opportunity for them to double click and learn a little bit more. And with Data Cloud, we're essentially doing the same thing, but with your business data, we're bringing together all of your Salesforce. Org, so service cloud, sales cloud, marketing cloud with structured and unstructured data like PDFs or emails or even Slack conversations, along with external data that might not reside in a CRM like IoT data or maintenance data or your ERP. And we're using all of that together, along with your unified knowledge, to ground your AI so that the responses, predictions and recommendations that are being generated are relevant, accurate and that much more useful. But again, I would be remiss if I didn't even bring up the topic of trust. We are in fact talking about sending your private business data to an LLM. So, Ellen, I'd like to hand it over to you to talk a little bit more about how we are doing that in a safe and secure way.

Yeah, Megan, I am so glad that you brought up trust. And at Salesforce, we're not new to AI. We've been developing AI for a decade now, and trust has been our number one guiding principle the entire way. So when we launched our first wave of generative AI features just last year, we did so with something called the Einstein Trust layer. Now, some of you may have heard about the Einstein Trust layer, but if you haven't, I'm going to spend a few minutes just talking a little bit about it. So the Einstein Trust layer isn't A product you have to buy, it's not just one specific feature. It's actually a collection of features, processes and other types of guardrails that we've built directly into our AI platform to instill trust for our customers when they're using generative AI. So the goal of all of these features and processes, guardrails, is a couple of different things. So number one, to promote the security and privacy of your data. Number two, to improve the accuracy and safety of your AI results over time. And then number three, to promote the responsible use of AI across the entire Salesforce ecosystem. So let's talk about data privacy and security first, because I talk to a lot of customers and this is almost always their number one concern in deploying generative AI. And for a fair reason.

Right.

Like Megan, like you said, you're basically giving company data via a prompt to a large language model, right? Well, there's a way to do that safely and securely. So I'm going to talk about a couple of different features that help you to do that. So first is something called data masking. So we talked a lot about grounding.

Right.

And grounding is basically pulling in context and data about your company into the prompt. Well, some of that data might be sensitive data like pii, like account information. So names, contact information, phone numbers, email addresses, Social Security numbers, passport numbers. It could also be pci, like credit card information. Right. And this isn't just for the U.S. right. We've also, you know, we also need to account for different types of identifiers that are specific to other regions around the world. So what we've done is we've built a number of different machine learning models that are able to detect these different types of sensitivity, sensitive data. And when they do so, we're able to replace them with a placeholder. So, for example, if you have a credit card holder's name, that might get replaced with something like cardholder one. So when you're sending the prompt with all of that data, it gets replaced with placeholders. The masked prompt then gets sent to the LLM, which uses those placeholders to generate the response, and then the response is sent back to Salesforce, at which point, once the prompt is responses safely back in the confines of the Salesforce trust boundary, we replace those placeholders back with the original points of data, so that if you're the end user, you don't even know that this is happening in the background, which is the really amazing thing. So the second thing that we're doing, which is really important that I want to call it is something called zero data retention. So this is actually a unique type of agreement that we've pioneered here at Salesforce with all of our third party large language model providers. And it's basically saying, all right, large language model, that, you know, once you've received a prompt from Salesforce and generated a response and sent it back, you are not able to retain any of that data. You must forget it immediately. So again, you know, if you're put it. Once your prompt goes from Salesforce to the large language model, the large language model generates a response and send it back to Salesforce. None of that data included in the prompt or the response is getting retained by the large language model. It's not being used to train the large language model. It's all getting forgotten and not being retained. And some folks say, well, that's just a contractual agreement. And yes, it is. But we also are able from an engineering standpoint, to go into the back end and verify that this is indeed happening, which we are doing actively. So that's just a little bit about data security and privacy. I could talk a lot more about it, but I also want to talk about accuracy and safety, because these are two other very important topics when it comes to trusted AI. So. So we've talked a lot about grounding, right? That is the best way to drive accuracy with your AI results. And as Megan mentioned, we have ways of observability, so things like citations. But I also wanted to take a couple minutes to talk about content safety, because toxicity is also a really big concern. So if you're not familiar with the concept of toxicity, it's basically saying like, okay, we've acknowledged. Megan, you said the Internet is a crazy place, right? There's a lot of harmful content out there. Things like hate speech, explicit content. The last thing you want is something like that to make its way into an AI result that gets surfaced to your employees, let alone your customers.

Right?

So we've been building a number of toxicity detection models that are built into our platform and are able to flag things like hate speech, explicit content, other things like that. So if you are for some reason or another getting some of that showing up in your AI results, we are able to detect it and you're able to see in the audit trail that that's showing up. So that's just a little bit of a snapshot into the Einstein trust layer. It's something that we're innovating on every single release with new enhancements. So I hope that this has just picked your interest. And you'll go and learn more about it later. And back over to you, Julian and Christina, Amazing.

Thank you both so much. This was great. I feel like I learn a ton every time we have a conversation.

So data is locked down.

I mean, grounding is obviously, as you can see, very complex. So we really just appreciate you guys explaining it, making it clear and simple for our audience. And speaking of our audience, now it is time to hear from all of you. So you have been sending in some of your great questions into the chat all morning. Again, if you have any questions, feel free to keep putting them in the chat. We'll do our best to answer all of them live. And with that, we really want to be able to have this live broadcast and have this live Q and A with our amazing experts here. So let's get to it. Julian, what is our first question?

Yes, so I figured out how to turn on the tablet, so that's step one. But let's see. We've seen some questions come in through the chat. So let's pop into this one. Looks like this one actually came from. It's a combo question from Srikanth and Victor as well. So. All right, this one we'll send over, I think, to Megan first. We can start with you. So I'm like, okay, let's see. Do you need Data Cloud to use Unified Knowledge? And the second part of that is how does the product fit into Data Cloud? So how does Unified Knowledge fit with Data Cloud?

Yeah, so that's a great question. And to quote my dear friend and colleague, Unified Knowledge and Data Cloud are like peanut butter and jelly. They work really well together. They taste great. So if you wanted to, as Ellen just mentioned, if you wanted to track your feedback, the usefulness of the content that is being generated or toxicity or anything like that, that's when Data Cloud would come into play. We use Data Cloud to track that sort of feedback. Unified Knowledge is bringing together your knowledge sources and Data Cloud is bringing together your data sources. And together, that's how you're going to get a really, really solid foundation for grounding your AI. So I think, in short, they work together. But for AI specifically, if you wanted to track feedback, that's when you would use Data Cloud.

Got it.

That makes sense. Thank you.

Peanut butter and jelly. Great.

Analogy.

That's wonderful.

Perfect. I love that. Then, Deborah, we'll keep it on you for a second and feel free. Also, Ellen and Megan, if you have a response to this too, feel free to chime in. But this one is from Christina in the chat. You're in the chat.

I'm in the chat.

Another Christina. But the question is, are you exploring any tools that can proactively identify issues in your grounding data, like duplication, conflicts, change terminology, things like that?

We absolutely are. One of the great features that we recently released was the ability to create knowledge using AI. So not only using AI to use your knowledge, but also to help you identify content that should be created and help you create it. You know, we always hear from our service organizations how, you know, they want to create content, maybe even in demand, or they want to create it, you know, as they learn that they need it. But it takes a lot of time. And so with knowledge creation, we can help use AI to give them a first start a draft of that to then, you know, just take a look at it and see if anything needs to be updated. It saves a ton of time. And so one of the features that we are working on with that will be being able to analyze similar articles. So that one's not there now, but it will be coming here relatively soon. And that ability to kind of identify similar articles will be really important for that, but then will be something that we can extend to other scenarios. So we'll be taking a look at, like, how could we make that beneficial across all of the experiences where people may want to look for that?

All right, another question that's coming in. Ellen, I know you're very passionate about the Einstein trust layer. So we did get a question about that. In terms of the Einstein trust layer, how does that really work with unified knowledge? How are they working together? Compatible, what it might be?

Yeah, that's a really good question. So essentially, the Einstein trust layer is built at the AI platform level. So it's part of the underlying architecture. So any of the features that we're building on top of it are all going to be working with the Einstein trust layer. So as you're building out your knowledge base, anytime that you're pulling in information from any of your knowledge articles or knowledge base into an AI prompt, the Einstein trust layer guardrails are going to be running against that AI prompt. So the things like data masking that we talked about earlier, toxicity detection, all of those things are going to be running against an AI prompt. So whatever it is that you're pulling from your knowledge base into the AI prompt is going to be working with the Einstein Trust trust layer.

That makes sense.

Thank you.

Thank you for that. Absolutely. And let's see. Looks like we have another one coming in as well. This One looks like it's from Scott. So this one, we'll read it and see who it may go to. Since Unified knowledge brings all knowledge together, will the company of the future in quotes no longer have several places to go for answers like a documentation site, a support site, a community site, et cetera. Instead, were there just one search bar for all queries? I don't know, just kind of a future product vision. So maybe. Deborah, I don't know. Did you have, could you kind of tap into that a little bit?

Yeah, happy to. I think that, yes, that could be true, but I think more likely the future will be that we can meet customers where they are. So, you know, different users want to go to different places to get the information that they need. So some really like to do mobile, you know, to be on your mobile phone and use messaging. Some hate that and they want to use some social experience. Some want to go to an official website because they feel like that's trusted. So I think, you know, even connecting into third party, you know, speakers. Right. So I think there's a lot of places that we go in our day to day to find information and really what we want to do, our vision for how we would, we would want this to take shape going forward is that you can get those answers wherever you are. So it reduces the friction of, oh, I have to go figure out, you know, where to find this information. I have to go look for this or go log in somewhere. Like we want to help meet you wherever you are. So I think that's probably the more likely future in terms of LLMs.

I know that was brought up. We had a question about what LLMs can be used with Salesforce and our AI can. And if you could touch a little bit on that, that'd be great.

Yeah. And the short answer is OpenAI, Azure, OpenAI and BYO LLM are currently supported today. But we are just like Ellen mentioned with the Einstein Trust layer. We are constantly adding and onboarding new and additional LLM providers.

One thing that I would add to that too is our philosophy is to give our customers flexibility in terms of the type of large language model that they use. So some customers have data residency requirements, some customers really just want to use their own fine tuned custom model. And so we want to where our philosophy is really to give our customers options so that they can choose the type of large language model or multiple large language models that are right for them.

Makes sense. Yeah, we give everyone the choice. Yeah, love that. All right, then we're coming. The questions are winding down, but maybe we'll take one or two more questions. So let's go over to, let's go over to Megan again. Keep it over here. So how is unified knowledge used for Genai grounding exactly? I know you touched on that a little bit, but maybe a little bit of a refresher.

Yeah, I mean, we talked about this a little bit earlier, but when you're asking the LLM a question or inputting your instructions in the form of a prompt, it actually will use whatever grounding sources you want. But in the context of knowledge, what we found during the pilot testing phase, a service reply not grounded on knowledge and a service reply grounded on knowledge was significantly different in terms of its usefulness. Once we introduced knowledge grounding, the replies and the emails that agents were presented with were infinitely more useful. So in terms of adding your Salesforce knowledge and your expertise, external knowledge sources, we find that it is critical to ensuring that you're generating replies, responses, content, anything that's going to go to your agents and customers. It makes it that much more accurate and relevant.

Cool. And unified knowledge. When you're grounding in your business's knowledge, it makes more sense. The answer is going to be what you want for your agents to use.

Bingo.

And trusted, which is great. Awesome. Well, it looks like we have just one final question for the group today, so Devra, I'm going to wrap it up with you. We have a question about visibility. Last question is how do I restrict visibility of third party knowledge content? If you want to just share a little more about that.

Yeah, for sure. As you bring that third party content in, we're able to map it to metadata that we store within Salesforce. One you can identify, is it customer facing, is it partner content, is it internal content as a customer content? Secondly, you can also tag it so we can map tags that are coming from the source system to data category tags that are within Salesforce. So that mapping helps you keep it identified, you know, helps it keep it kind of tagged and sorted essentially so that we can filter and use the right content for the right audience. When we are using AI to generate a response, and we know that that is an internal use case versus an extern internal use case, we're only going to bring the content in that's appropriate for that internal versus external audience. So if there's something that's, you know, tagged for internal use only, then we might use that only for, you know, a response to an agent themselves if they're doing some troubleshooting internally. But if it's a content that can be shared externally, then we would make that available for AI to generate a response through, like the self service portal, for example. So really, as you're coming, as you're bringing the content in, you can tag it appropriately and then we'll know how to use it in each of the AI experiences.

All right. To learn more about Service Cloud, click the link on screen. Thank you so much for watching and we'll see you next time.