# Meet Agentforce Humans + AI + Data + Action Auto-transcribed by https://aliceapp.ai on Wednesday, 18 Sep 2024. Synced media and text playback available on this page: https://aliceapp.ai/recordings/l_uxhQMP0AJ4g8RUVBptw6joIekFDLXz * Words : 5,140 * Duration : 00:32:27 * Recorded on : Unknown date * Uploaded on : 2024-09-18 17:36:07 UTC * At : Unknown location * Using : Uploaded to aliceapp.ai ## Speakers: * Speaker A - 6.61% * Speaker B - 8.07% * Speaker C - 55.33% * Speaker D - 29.98% ---------------------------- Speaker A [00:00:00] Customers across the board. We stick with our low code, no code, pro code, but there's some really amazing out of box agents that our customers can take advantage of. Let's talk a little bit about those. Speaker B [00:00:10] Yeah. So we have agent four, service agent. This is amazing for everything from answering questions about your products, deflecting cases and support, to actually taking action. So with keynote, we've seen folks like Wiley, we've talked about folks like Saks using those faqs. There's large automotive companies that are launching new cars. They want to put their whole owner manual and have the agent be able to answer questions. I was talking to a logistics customer last night and they've got 30,000 different pieces of policies that they want this agent to be able to handle answers to for their very complex business. And then there's another retail, uh, company that has all these promotions that conflict, and they want to be able to have that world class service for folks. Then for sales, it's handling all those outbound leads. Hey, I want to qualify those leads. I want to schedule the meeting so that I can get that rich relationship built by handing off to an account executive. And then of course for coach, we can have those great discussions that you can scale that coaching, maybe. I've got a big meeting tomorrow and I want to practice my pitch. My boss might want to be sleeping in with their manager. Turns out coach is pretty darn good and can give me that real time feedback before I go into a high stakes engagement. Speaker A [00:01:19] Yeah, there's a couple of things I love. There is as a consumer hearing just how much better service will be and getting the answers that I need when. Speaker B [00:01:27] I need them get answers, not links. Speaker A [00:01:29] Yeah. And as a former account executive, thinking, uh, through how you know that proactive reach out. But what I love in the way we do it is we are indicating that you are talking to an agent. So the person on the other side in that experience always knows that. Speaker C [00:01:47] Yeah. Speaker B [00:01:47] You need that trust and disclosure. That's important. Speaker A [00:01:49] Well, and then I think, let's talk about trust, because I really believe when I look at the technology, we have put trust into every single piece of these features. And so what makes Salesforce uniquely positioned to be the leader in agents? Speaker B [00:02:03] Yeah. And so just to, uh, get to the starting line, so many customers need a lot of assurances. All right, are you going to make sure that these models aren't being trained on my data check? We did that. Are you going to make sure that any of the model data isn't going to get leaked? Check we got that. Hey, how do I know if this is toxic or biased? We've got predictions on that. We track the Pii. How do we make sure that the data we're putting in is securely put in? We've got you covered there, too. How do we make it easy to set up? How can I then test the answers? We've got a huge amount of effort we've put in so that you can trust not just getting to the starting line, but then validate, as you would with any new hire, that they're doing what they're supposed to and not what they're not supposed to. Speaker A [00:02:48] Yeah, and those guardrails are super important. And I think when we look at and when customers really get comfortable, if they see the way it computes, the way it takes that action. One side of the screen here are all the topics. These are the data sources that I have selected. I look in the middle, I type in natural language instructions, and then on the right hand side, on the same screen, I can see the outputs. So I think once I saw that, I truly understood that trust layer built in. And so actually monitor and see what it's doing. So. Speaker D [00:03:20] Absolutely. Speaker A [00:03:20] Uh, wonderful. So thank you again for all of your and your team's hard work. To learn more about how your organization can use AI, click on the QR code on your screen or go to salesforce.com artificialintelligence. Use cases for now, we've got plenty more to deep dive on when it comes to AI agents. We talked about it before, but check out this theater session on meeting AI agents, your new trusted digital coworkers. Speaker D [00:03:48] Hello. Hi, everyone. Welcome to Dreamforce 24. We're almost done with day one, so congrats to you guys. Um, this session is, of course, about agent force. We're going to talk about how Salesforce is really revolutionizing AI by bringing humans with AI data. And action. 1 second. 1 second. Give a second, please. Still working out the kinks. Yeah. Speaker C [00:04:42] Uh. Speaker D [00:05:15] M. Okay, perfect. We got it. Okay, so now let's talk about everyone's favorite slide, our forward looking statements. Um, as many of you might have heard, our agent first agents are actually going to be generally available in October, so just a short month from now. But they are not generally available today. So we advise that you base any purchasing decisions on products that are generally available right now. Okay? So, first and foremost, before we get into the content, I have to say thank you. Thank you for being our channel blazers, and thank you for traveling from near and far to be with us and really share in this excitement of agents. So let me introduce myself. My name is Kamaria Wilson. I'm a product marketing manager here at Salesforce. I focus on AI, and I'm joined by my colleague Shipra, who's one of our product managers, also on the platform. Ok, so there's no secret that we are in an AI revolution. Let's talk about what this looks like. Wave one of this AI revolution was predictive AI. So this is not a new technology to most of us in this room. It's being able to anticipate what's going to happen in your business using data. And then next we saw Copilot. So it feels like it was just yesterday that generative AI took the world by storm. But in reality, it's been over a year now that we've been working with Copilot technology. And so Copilot's really revolutionized AI, uh, by bringing assistance and being able to have AI, uh, be conversational and natural and start to take those actions. And now we're in wave three with autonomous agents. And so this is a huge step change. And what I want to call out on this slide is that the time period between each of these waves is getting smaller and smaller. So now we see autonomous AI agents just about a year and a half to two years after we saw copilot technology. And the best part about these agents, that they're proactive, so they can take action without human input. So why do we need AI agents? Why now? Why do we need this technology? Well, it's no secret to anyone in this room. I'm sure people can resonate with this slide. Our teams are overstretched, and sometimes the business priorities that we have and the bandwidth that our employees has just doesn't match up. So sometimes we see employees being stalled out in their productivity because they're spending time m doing low value tasks. So, in fact, we've seen that 41% of employee time is, uh, spent doing low value tasks. I don't know about you guys, but there's definitely time spent in my day doing things that are necessary, but not necessarily the most high impact. And so at the same time that we're seeing these workforces be overstretched, we're also seeing that customer expectations are continuing to rise. Customers are aware of the impact of AI. They're aware that we have more data than ever, and they want to see higher personalization, higher touch experience. They want you to know who they are, what they like, every interaction they have with you in any channel. So this begs the question, what if you were able to scale your workforce with a team of specialists that can work around the clock to serve your customers and employees well, that's exactly what Agent Forest can do. And so agent Forest is very groundbreaking in the sense that it's fully integrated with the Salesforce platform that you all already know and love. So the CRM apps that you're using every day with all of your data connected, whether that's internal to your, to your salesforce instance or that you're pulling in through data cloud, these agents are also easy to deploy. So luckily we're building out of the box agents for you to customize. So in reality, you can spin up an agent within minutes and you can use low code tools that you're already using, like flows, prompts Apex. And you can use that to customize those agents to your liking, to your business, and to do the tasks that you want them to accomplish. And of course, it wouldn't be salesforce if this wasn't trusted and scalable. So this is all built on our trusted platform. Many of you might be familiar with the Einstein trust layer. That same technology is being leveraged with agent force. And also we have the added guardrails for your agent. So that means that you can use natural language to instruct your agent on what to do and what not to do, what actions not to take, and when to escalate to a human that can handle that more high priority work. And last but not least, we have an open ecosystem. So of course we have model partners. There's tons of large language models on the market now and we want you to have your choice of your model partner. Sometimes these LLMs are flexible and some LLMs perform better for different tasks. We want you to have that flexibility in choosing the model of your choice. And then we also have partners who are building and extending these agents for you. So if you choose to, you can partner with them and have agents and actions that they're building. Okay, so let's pause and talk about what exactly an agent is. So number one, an agent has a role. So remember when I said they're specialized? An agent is different than a copilot because it has a specific role, it has a purpose on your team. So that can be to analyze marketing campaigns or to reach out to your leads for you and that's for you to define. And then next you have the data. So of course, AI can't do much without data. So this is the knowledge that your agent has and needs to execute its role. And then next we have actions so these are the building blocks of your agent. And so those flows that you might be working on, those automations, those prompts that you're building, you can use those and invoke them with your agent. So you can have an agent automatically trigger a flow and execute that workflow for you. And then last but not least, we have the channel. So this is where the agent is getting work done. These are the applications that it's able to surface in. So that can be in slack, that can be in WhatsApp. It can also be on your own website. So that's for you to determine as well. And then last but not least, as I said, this is all trusted and secure. And so it's really important to highlight that we've been very intentional about the guardrails that you set for your agent. So, like I said, what it can and cannot do. So why are agents different than what we've seen in the past? We talked about this AI revolution. How is this different? How is it more valuable? Well, I think we've all probably interacted with a chatbot before. Those are very rigid and inflexible systems. So all of their logic is predetermined. A lot of their conversations are predetermined, and they're very fixed rules. A customer or an employee goes off script. Those chatbots largely fail out. And then with copilots, obviously, super innovative technology, because they're able to converse with your employees, and they're able to provide tailored assistance and able to make them more productive. But what a copilot cannot do is that it cannot take action without a human input. And so that's where agents come in, and agents can be triggered by a conversation, just like a co pilot, or it can be a change in data, or it can be some sort of preset automation or trigger that you define for your business. And so these agents, they understand your business with the data that you have. They're able to plan and reason. We'll talk a little bit more about the atlas reasoning engine, but that's the brain of your agent. That's how it's able to be more conversational, more capable, and more flexible. It's able to look at the actions that it can do and reason through the utterance and map that to the action. And so, last but not least, this is scalable. So the vision of these agents is that you're able to have multiple agents. You can have an agent for marketing, sales and service, and maybe one day these agents will be able to communicate with each other and work together to improve your business workflows. So I'll, um, invite my colleague Shipra to talk a bit more about our architecture. Speaker C [00:13:22] All right, thank you, Kamaria. Hello, can you guys hear me all right? All right. Thanks, Kamaria. And hello, trailblazers or Agent Blazers as ah, we are going to start calling you from now onwards. Are you guys excited about agents? I know it's evening, it's like 445, so I can see why your energy levels are somewhat low, but let's try to show you some exciting stuff here. All right, so the power of agent force is really that it brings the humans, the data, the AI, and the actionability. Together with data cloud, you can ground these agents in your enterprise data and context to make them more relevant, more personalized, more contextual. With our AI platform, um, you can not only bring one model, but you can bring any model. You can either build your own model or you can bring any, uh, model that you are invested in, all the while utilizing the integrations and security from our platform. And with c 360 apps, you can deploy these agents on any channel where your employees or your customers are and also take advantage of the automation that already exists in the platform. So anybody here has built any flows, any apexes, any automation? Yeah, some of you are nodding your head. All of that automation that you have in the platform now, you can make it actionable through agents by converting them into agents actions. And that's simple configuration. All right, so with agent four, it is very easy to augment any function, any team, to drive higher productivity, higher efficiency and top line growth. For example, with service agent, you can drive the CSEt as well as call center efficiency. With sales agents, you can increase the selling capacity of your organization and increase revenue. With marketing agent, you can create effective campaigns and drive more leads or drive customer loyalty and so on and so forth. Can think of an agent? We have an agent for that. Now, where do you build and maintain these agents? All of that magic happens in agent Builder. This is your one stop shop where you can create, customize, maintain, govern these agents. It is also a drag and drop interface where you can easily configure your actions or your agents. You can add actions to agents, then you can deploy those agents on channels. We support a number of different channels on which you can deploy these agents. If you have websites, mobile apps, SMS, WhatsApp, ah, Facebook Messenger, Apple Business chat. Right. So there's a vast plethora of channels that are available for these agents to be deployed on. You can ground these agents in knowledge that acts as the grounding source for your agent, as well as the source to answer any frequently asked questions. And then finally, you can test and preview these agents in this beautiful intuitive builder that we have this canvas. And boom, just like that, you are ready to activate your agents in minutes. It is actually that simple. All right, so you must be wondering, okay, this is all talk. You know, what is really under the hood? Is that all a black box. There is some technological innovation that has gone in, in building these agents and making them so powerful and so robust. So there are some concepts I would like to familiarize you with. Okay, topics, instructions, and actions. Those are the three concepts that are integral to the working of an agent. Number one, topics. Topics represent a user's intent, a user's goal, a job to be done. Okay, so for example, order management, repair, inquiry, frequently asked questions. All of those are user intents. When the user input comes in, the system maps, it, uh, classifies it to a topic that is configured in your agent, and that's how agent knows what this user is looking for. Okay, then we have instructions. Instructions are natural language instructions that you embed in the topics to provide additional guidance to your agent on how it should behave. What actions should it take, uh, what policies or guardrails should it apply? So let me give you an example. If you want to say you are a retailer and you want to process a order return, maybe you might have a requirement that you always want to authenticate this user before you want to process the return. So the instruction that you will provide to the agent is that, hey, always run the authenticate action before you run the order return action. So it's those kind of policies. Or maybe you may always want to exhaust all the troubleshooting, uh, tips before you create a service inquiry service request. So topic and instructions together ensure that your agent operates within the boundaries of your business specific rules and policies. And for those off scenarios, because humans are humans, we sometimes tend to get funny and creative and crazy. Uh, we also have reverse topic and off topic to handle scenarios where the users, we are off track. So that's how we make it work. And then finally we have actions. Actions. As the name indicates, this is how agent takes actions. We support a vast number of ways in which you can create configure actions for agents. You can use prompt templates in prompt builder and ground them in your enterprise data from CRM and data cloud. You can use flows and the workflow automation that you have in flows. You can make it actionable through agents. You can use Mulesoft APIs and convert them into actions for agents to give you access to external systems. Or if for your complex use cases, you can also use apex classes and convert them into agent actions. So there's lot of different ways for you to bring actionability to your agents. All ah, right. So let me now go over how does it work? How does. Let me bring it all together. Okay. So a user input comes in that acts as a trigger to get the agent working. Now that is one mechanism. Remember, these agents are autonomous, they are proactive. So there are other mechanisms to trigger the agents as well. Say a data operation takes place on your CRM data, or you have a business automation rule somewhere that says, if XYZ happens, trigger my agent to, uh, get working. So there are a number of triggers that are available. Once the trigger comes in, the agent gets all the context, all the information about that request that is coming in, and it goes to LLM to ensure that LLM understands what this user is asking for. If it does not fully understand, it can go back to the user and ask intelligent follow up questions to gather more information. Once it understands the user's intent, it maps it to a topic, right? Then it has the topic, it takes the topics instructions that I just talked about, and it takes the actions that are also part of that topic, that are required to fulfill that specific user intent, user need. And it goes to the LLM to formulate a plan. The plan would be something like, hey, first execute, uh, action a, then execute action b, et cetera, et cetera. So based on that plan, the agent starts running these actions and it generates the output. Once it has the output, it tries to evaluate if that output is meeting the user's goal or not. If it is meeting well and good, it serves that response to the user. If it doesn't yet meet the user's goal, then it can again go to the LLM to replan, rerun and reassess. So that's kind of how it operates in this, in this loop. We call it a react loop, um, to, um, generate these responses for the user. So that's kind of how the intelligence works. We'll take questions afterwards, if that's okay. Um, all right, so now let me show you all of this in action. And hopefully our screens here will still work. So let's see m, let's drop this thing down so you can see this, right. Okay. So as I said, the agents can be deployed on a number of different channels. In this case, I have built a very simple agent. I'm calling it Dreamforce agent, because I thought, we are at Dreamforce, why not build an agent that can help us get more out of our Dreamforce experience and show how agents work? So I have this Dreamforce agent that I have deployed on WhatsApp, which is one of the very widely used channels across the world. So let me start interacting with this agent and let's see, um, oh gosh, my phone. So let me ask a simple question, uh, to this agent should attend Dreamforce. It's a very simple generic question, uh, but what is going on here is more powerful. So I have configured this dreamforce agent within knowledge source. In this case, the knowledge source are the knowledge articles and using those knowledge articles, uh, it is responding to these questions. So what is going on here is a frequently asked question. Your agent is respond from the knowledge source that you have specified. Right. So we have a response here. It says, everyone is welcome and encouraged to attend Dreamforce. Looks uh, good, I mean, I think you can take a look at it. So now let's try something different. So this was about, uh, responding to frequently asked questions. Let me see, uh, what Dreamforce. Did I spell that right? Dreamforce sessions are, ah, taking place. I want to find out what sessions are happening tomorrow. Okay, so on 18th. So remember I said, and okay, there's some disturbance in the back, um, the agents can take action. So in this case, what this agent is going to do is I want this agent to go and retrieve information from CRM for me. So what I've done here is I have created a custom object called Dreamforce session in which I've created records for various Dreamforce sessions and it's able to query that custom object to provide me information. So this is to showcase the actionability of these agents. By the way, as you can see here, this is just a illustration. Um, this is not the full lineup of Dreamforce. There are many, many, many more sessions happening at Dreamforce. This is just to give you an idea of how agents work. The other key call out here is see how I mentioned what sessions are taking place on 18th. I did not say the 18 September, but because I'm talking about Dreamforce, the uh, agent is able to understand the context that this must be about Dreamforce and that's why it brought me the information for September 18. Now that's really powerful, right? So I showed you the frequently asked questions, I showed you the actionability. Now let's see, um, what is the sitting capacity of first one? So some sessions, especially uh, the keynote sessions, have reserved seating capacity or limited capacity. So I want to, uh, go ahead and try and see if I can get into one of these keynotes. So I'm asking it. Oh, sorry, I did not write it. Well, this is too. Yeah, this is understandable. What is the seating capacity of architect keynote. Sorry, I forgot my glasses. Reading glasses. All right, let's see if it brings the information. What, uh, I want to do here is go ahead and try make a reservation in that session. Okay. So it brought me this information about that session. It's telling me, uh, uh, that the seating capacity of that session is 500. Again, this is all illustrative. Please don't go by the information here. Uh, please look at the right app for your dreamforce information. Okay, so now let me go ahead and ask it to make a reservation. Can you make a reservation for me? But see how the conversation is flowing in natural language, right? Like, just as humans talk to each other, can literally talk to the agent similarly, and it can understand very clearly. Um, okay, what happened here. Something my omnichannel is probably not up. So I want to show you the transfer to a human agent because that is another key part of our, um, solution that you need to be able to transfer to a, uh, human agent safely and securely based on your business rules. I'm going to make my omnichannel online and pretend that there is a human agent online. Let me try again. Can you make a reservation in architect keynote for me? Right. This is. This is really critical, so. Oh, it's, uh. I could not show you that. Sorry. It's two screens is a bit of a problem. Uh, but let's see. So this is what I typed, and at this point in time, it is transferring me to a human agent. So now I'm pretending to be a human agent. I would accept this, and I will just say, be right back. Right back. All right, but you get the point. Right? So this is, again, showcasing how easily you can also transfer, keeping human in the loop. You can transfer to a human whenever you desire. So now let's look at my configuration and how easy and, um, simple it is to build an agent like this. All right, so this is the agent builder screen here. What you will see here is four simple topics I've created. So I have a general faq topic. This topic contains one action, which is the knowledge action. This topic is responsible for responding to those frequently asked questions. I've configured a knowledge source in this action. It is using the service knowledge to respond to those questions. Or the very first, uh, interaction that you saw that came from this particular topic, then I have Dreamforce session info. This is yet another topic. And this one is responding to those specific questions, um, specifically the ones that are querying the CRM object. I have a number of different actions here. Uh, it actually executed the query records action to bring me the information from Dreamforce session object. And then finally I have, let's go back, ah, to Dreamforce agent. And finally I have this escalation topic. Right, but the thing that you should notice here is how I can configure these very specific instructions here. So I told explicitly that, hey, every time a user asks for making a reservation, transfer them to a human agent. So this is to demonstrate that you can create these kind of business rules and policies to configure when you want to securely transfer to a human. Right. So that is my demo. So hopefully you got an idea of how you can build these powerful agents using very simple configuration and setup. And with that, let's see. All right, so to summarize all of it, you can build agent force agents. Uh, there are many, many templates that we are making available out of the box. Using these templates, you can get these agents up and running in no time, literally in minutes. You can also customize them or if you want to build them from scratch in the agent builder, you can do that as well. Two, a real powerful reasoning engine. I touched upon that a little bit. Uh, that is the autonomous reasoning engine. That is the brain of the agent. That's how it is able to make decisions and take actions and provide accurate, factual results. Uh, our agent builder, that is our no code, low code interface for you to build and maintain these agents. And then last but not the least, we are embarking on this journey to create 1 billion agents. We are supported by our wonderful partner network with some major industry stalwarts to, uh, help us on that journey. All right. With that, please continue your journey. I would highly, highly recommend if you have not checked out the, uh, launchpad, if you want to really try all of this hands on, you know, just like book an appointment, you can walk there. Uh, within ten minutes you can get the appointment, but you should definitely check that out. A lot of people, people have tried it and, uh, you'll really love this, um, more resources. So check out agentforce.com or, uh, trailhead module. Um, and with that, a big thank you, I understand. Time for dinner, but I really appreciate you staying back and, uh, joining our session here today. Hopefully you found it helpful. And please, please, uh, provide us feedback that is really, really important to us. You can win a Starbucks gift card as well. Okay. All right. Thank you very much. Have a good evening, everyone.