Welcome. On behalf of Salesforce, welcome to Dreamforce. My name is Mike Spencer. I run Investor Relations here at Salesforce. We're super happy that everyone was able to make it out. I know for many of you, it's a long haul from the East Coast. I can appreciate that very much, and I'm really hoping that your time here, you all have been able to spend a little bit of time, and if you haven't yet, I encourage you to do so, go down and walk the floor, talk to partners, talk to customers, because really the goal for us, having you all and inviting you to Dreamforce, is really to get a better feel for what we're experiencing as a company and why we're excited about where the technology is going.
So what we're going to do today, we're first going to start with thank you. Sincerely, thank you to all of you. It's been a journey for us, obviously, as a stock. We appreciate the support from our investors. We appreciate the feedback. We get lots of it. But most importantly, we appreciate the partnership, and I really do mean that. We understand our role. We understand the need to continue to build credibility with all of you, the need to continue to execute, and hopefully, we continue to build the belief in the stock, and at the end of the day, that's what we're trying to do. Today, what we're going to do, it wouldn't be an investor dialogue without a safe harbor. I'm not going to read it.
I'll save you the pain, but obviously, we'll probably make some forward-looking statements today. It comes with all the natural caveats with that, so please take with caution. Today, what we're going to do is we're going to talk about Agentforce. It's not going to be your typical investor day. We're not going to give you forward-looking guidance of any sort. Really, the goal of the session was to give everyone here a more intimate view of where we're going from a product standpoint. Hopefully, a number of you caught the keynote yesterday, and you heard Mark, Clara, and Patrick actually talk about Agentforce and where we're going and the vision for AI. And what we really want to convey to all of you is why we're excited about technology.
And we understand the journey we've been on, and the team will get into it here in a second, but most importantly, where we're going as a company moving forward, and why we're excited about technology and what our customers are saying. So first, we're going to have Clara come up, and as our head of AI here at Salesforce, kind of walk you through the strategy and the direction where we're going. And then after Clara, Patrick will come up, and Patrick will give you a demo, a little bit more of a double click of a demo from what you saw yesterday, to really get into the nuances of what we're seeing. Most importantly, then, after we walk you through that, we're going to do a Q&A session.
I'll come up and join Clara and Patrick on stage. What we really want is an open dialogue. We want you to push us. We get the questions. My team and I get the questions all the time from a lot of you, very direct in many instances, which I appreciate. But most importantly, we understand the landscape of what's happening, the amount of AI news that's getting thrown at all of you. And so we want to help clarify a lot of it and help you understand where we're going from a technology standpoint. And in particular, Clara and Patrick spend tons of time with customers and partners, and so they can give you a lot of first-hand knowledge of conversations, pushback, challenges, and just really where the technology is at.
So that's really the goal for today, and so I really, really encourage all of you to be ready with questions. I was getting a few of them in the precursor leading up to this, so, and I told a couple of folks to save the questions and ask them when we get into the Q&A. So with that, I'm going to hand the floor over to Clara.
Great to see everybody, and I'm going to click a little deeper versus what I showed yesterday. But in case you missed the keynote, this week is all about Agentforce. It's agents and humans working together to drive customer success, and it's really building on the heritage, legacy, and proven success that Salesforce has created over the last twenty-five years around trust and security, scalability, and accuracy, making it really easy to customize. I think about our twenty million Trailblazers and how much innovation they've been able to build over the last twenty-five years, and now being able to use a lot of those same skill sets in this agent era that we're embarking on. Integrated into the Customer 360, so that it's not just a science project, but driving real outcomes like higher, like higher sales conversions and faster service resolutions. It's our unified metadata platform.
Being able to have data that describes the data that's in Salesforce. Turns out that's really important for the AI to know what data to pull in and what actions to invoke. They need those descriptions that have been a core part of our platform since the very beginning. And last but not least, it's an open ecosystem of partners. We announced our Agentforce Partner Network. The first time I worked in Salesforce, however many years ago, 18 years ago, we launched the AppExchange, and it's that ecosystem that's been such a powerful part of why Salesforce is successful, both ISV partners as well as SI partners, that fill in the gaps and bring these strategies to life for all of our customers.
And so just to recap again, this kind of evolution that the industry has been on, where we started with pre-LLM chatbots that are very rigid. And you need a programmer to actually define every branch of the logic. So if, then, else statements. And, you know, we've all used chatbots before. We've used them on the phone, we use them in email and chat, and it's a very frustrating experience. Not to mention that at these companies, you need to wait on the backlog of IT in order to build those chatbots, so it takes a long time. Fast-forward to this world, I think we've learned so much over the last eighteen months about what it really takes to deploy an LLM, an enterprise AI into production.
18 months ago, many companies thought that they could just deploy a copilot, that they could just buy an LLM instance from one of the model companies, connect it with their data, and then be done. But we all now know that that doesn't work. In addition to having a host of security issues, it's too broad, it's too general. It's, it's a mile wide and an inch deep and not able to do anything meaningful.
And that's why we're so excited now about Agentforce and the early results that we're seeing from companies like Disney and OpenTable and Wiley, and companies like Saks, who ended their pilot early because they said, "The results are too good for us not to roll this into full production, not to get it onto our actual website, so that we can start helping customers and relieving our customer support teams." That's really an exciting evolution that we've been on, and it really is more of a revolution that we're leading. So again, if you were in the keynote yesterday, you'll remember I walked through the five components that make an agent.
The really exciting thing for our customers is that they don't have to DIY from scratch these five components, 'cause each of these are really big, big, hairy technology projects if you try to build it from the ground up. If you want an overall system that's reliable and accurate and relevant, you really have to nail each one of these five, as well as integrate them, and that's a huge project, again, as the industry has learned over the last eighteen months. So when I think about the first ingredient, which is the role, these are the roles our customers already have today in Salesforce. You know, before they deploy a digital sales development representative, they've already created a profile for their current sales development representatives. Wiley wanted to scale up their customer support team with a Service Agentforce.
Their existing customer support team is already defined in Salesforce. Their tasks, what actions they're allowed to access, what their metrics are, what their success dashboards. Now they're able to use those same tasks, metrics, and dashboards for their Agentforce. It's a real head start when you're embarking on this journey that de-risks it because you're not guessing. It's actually what your people are doing today. The second piece of this, very important, is data. Agents need data to access, knowledge to access, in order for them to perform their role. Again, our customers have their data in Salesforce already. They have structured data, like database, rows and columns, and custom objects. They also now have unstructured data that we're able to support in the form of Slack conversations, call transcripts, knowledge articles, videos, audio recordings, images, right?
It's multimodal in our Data Cloud. Data Cloud also houses metadata. I talked about that earlier. And now with our new Data Cloud Zero Copy Partner Network, that federates data from an external lakehouse and data lake, like lakehouse data lake and warehouse, such as Snowflake or BigQuery or Databricks. Now we're able to pull in that external data to activate Agentforce. It's a really amazing technology, and it doesn't involve, as you can tell from the name Zero Copy, you don't have to copy or move the data over. We actually virtualize it for security reasons and cost reasons. The third piece that agents need are they need actions. Actions are the systems and the workflows that they are authorized to access in order to perform the job.
And as I said yesterday, this is really powerful because you're driving automation when you allow the agent to access actions, but it's also high stakes. And if you don't give the agent specific actions that they're allowed to do, such as your specific way you want to run a credit check on a mortgage customer, or a specific way you want to handle returns that are high value over $200, because you want to make sure that your company doesn't lose money, then you have to have those deterministic blocks of code and logic ready for the agent to use. And that's exactly what our customers have been building into Salesforce for the last twenty-five years. I actually just learned earlier this week we're at eighty billion Salesforce flows that run every week.
So that's 80 billion times that a customer doesn't have to DIY, and 80 billion times that an LLM isn't hallucinating how to process your refund or offer a discount on your product. Right? You do not want the LLM making that up. You want it to follow the very important ways that you've defined for it. So that's the actions piece. Fourth is channels. This is where your employees and your customers interact with autonomous agents. And here also, for employees, the channels are within Sales Cloud, Service Cloud, Marketing Cloud, Commerce Cloud, as well as Slack. That's where they're already logging in. It's a natural place for them to engage with agents. For customers, it's digital channels, and we put in a lot of work.
When I was leading Service Cloud, one of the biggest things that we did was we connected all of these popular digital messaging channels, from WhatsApp to Apple Business Messenger to SMS, email. We just acquired a voice company. And so we're the only platform that offers all of these channels natively, and it's just one less thing that companies have to DIY and have to connect into this overall system. And then the last piece of this is the most important, it's the trust and security. It's everything from encryption and data residency, for where is the data housed? Think about our Hyperforce infrastructure and how much investment and how much time it took to put that in place. It'd be very difficult for a customer to rebuild that from scratch.
It includes our Trust Layer , everything from data security, data privacy, like data masking and zero retention prompts, ethical guardrails. We've developed a toxicity filter to adjust for the inherent bias in a lot of these LLMs. We have observability and monitoring so that at all times it can be supervised. We can make sure the Agentforce are performing and behaving as they're supposed to. It's actually broader than this. If you actually take a step back and look at everything that customers have invested into Salesforce, those 20 million trailblazers, every profile they've defined, every metadata schema they've defined, every flow they've defined, those actually become the guardrails for what the autonomous agent is allowed to do. If the autonomous agent ever thinks of something beyond what's been done before, it can do that.
It can suggest it, but it requires human supervision and approval before it enters kinda this trust boundary. And that's why you see so many customers. I hope you've gone through our Agent Launch Zone , and if you haven't, I highly encourage you to swing by Moscone West. That's how they're able to get up and running so fast, is because the components that they need, they're not DIY-ing from scratch a secure infrastructure or a completely new way to process returns for their retail store. It's already there. They're just connecting it up so that the agent is authorized to access that data, that role, and that particular piece of business logic.
In terms of the use cases that we're seeing, the two most popular ones that we're seeing customers deploy, this probably won't surprise you because they're our biggest clouds, are the sales development representative Agentforce and the service Agentforce. So the examples you heard yesterday on stage are the service Agentforce, resolving customer issues, answering customer questions. The sales side, we haven't talked about as much, and it's... You know, think about whether you're selling leather shoes or you're selling a complex piece of technology. You don't wanna waste your highly paid salespeople's time talking to people who are just asking very basic informational questions that, frankly, they could just Google, but, you know, wanna give them good service. So now you can deploy a sales Agentforce to patiently answer that customer's questions as much as they want.
Then when they're qualified, when they start asking about price, that's usually a good signal that they're very interested, then you kind of pass that qualified lead to the salesperson, and the salesperson is very happy to take the call. Okay, so the other part that I wanted to touch on, I talked about this yesterday, is, you know, why are we getting such high accuracy rates and relevance rates, two X better than when customers are building on OpenAI or on the competition? It's really because of our Atlas Reasoning Engine. And what Atlas is, it's a multi-agent system that is, it can generate a plan. So if the goal is to generate more leads on your website, it'll come up with a plan on how to do that.
It self-reflects, evaluates through Chain of Thought, and it loops until it has high confidence that the plan can address the goal that's been set forth. As it does so, we have advanced Retrieval-Augmented Generation , that's custom embeddings models and custom retrievers that our research team has pioneered specific to CRM data formats that pull in the necessary structured and unstructured data from Data Cloud. And then it goes in and actions according to the actions that are allowed in the Customer 360, whether that's automating a campaign in Marketing Cloud, engaging a prospect that's on your website in Sales Cloud, resolving a customer issue in Service Cloud, or in the case of OpenTable, confirming a dinner reservation for your friend's birthday, so whatever that goal is.
And then, if it ever gets stuck, this is so important, and this is why standalone bot and agent companies, it's gonna be very challenging for them because these, these agents do get stuck sometimes. There's always gonna be edge cases. But in this case, if it gets stuck, Atlas sends any low confidence scoring tasks to a team member, a human team member in sales or service, marketing or commerce, cloud or Slack. So it's very seamless. It hands off when it's unsure, which is exactly what you want it to do. And then, we also talked about yesterday, Atlas gets smarter and better the more you use it because we've pioneered a technique called reinforcement learning from customer outcomes. It's better than reinforcement learning from human feedback, where you're hiring contractors to manually label inputs and outputs.
Those contractors may not be subject matter experts, but in this case, you're actually using objective measures like, what was my conversion rate on this campaign versus a different one? What was my sales win rate? What was my sales deal cycle? What was the customer satisfaction score? And you're using that to tune the models over time so that that agent is really optimized for that company and that department. And because we're not one of those consumer companies that monetizes people's data, our customer's data is never our product, we keep that reinforcement learning outcome data proprietary in Data Cloud for each customer. So this is how it all comes together. If you think about those five components, it kind of maps perfectly, very so, just so nicely with everything that we've built over the years.
If you were starting an agent from scratch, you would, again, want to have MuleSoft integrations to other systems. You would need to have all of your channels, you'd need to have secure infrastructure. You would want to have a secure data cloud that can process structured and unstructured data. You would want to have these applications and roles defined. And so that's why I think we've been able to go to market so fast, and also why our customers are able to get to market and time to value so fast. So, I'm going to stop talking and switch over to Patrick and Gabe, who will show you what this looks like.
Okay, hey, everybody, I am Patrick Stokes. I'm not going to talk for very long because we really want to show you, but, I'm just going to tell you what we're going to do. So, we did a pretty good, I think, demo of all of this coming together, I think it was yesterday. If you haven't watched that, I encourage you to go watch it. It's pretty great. We thought about doing that for you here, but in the last, I'd say thirty-six hours, something really magical has been happening. And I'm a New Yorker and an engineer, so I don't use that word very often. But not too far away, at Moscone West, our customers are building agents right now, and it is insane to watch how quickly they are doing it.
And the reason they're able to do it so quickly is because of what Clara said: we are the only one right now that has this kind of full integrated system of all of the things that you would actually need to have an autonomous agent. There are other data platforms out there, there are other AI platforms out there, there are other app companies out there, there are other agentic platforms out there, but none of them have it all wired up together in the way that, that Clara just went through. We do, and because of that, you can deploy, build, and deploy these things to do really useful things with your business incredibly quickly. There are tens of thousands of people building agents. We've deployed, probably yesterday was twenty-four hundred. We probably have over five thousand that have been deployed as of today.
We are hoping to get over seven thousand before we leave here. So rather than showing you the whole demo that we did yesterday, what I wanted to do is bring up Gabe to show you the actual environment and a demo in that environment, which is the live environment right now that our customers are using. So I'm going to turn it over to Gabe. Also, if you follow Mark's Twitter, and you see some of the videos that have been on Mark's Twitter, you may recognize this silky smooth voice. Go ahead.
Thanks for having me today, you all. So what are we going to be doing during this demo? Are we seeing the demo? We're not seeing the demo yet. Here we go. What I'm going to do is I'm going to take this chatbot. So you heard Clara talk about this during her intro. You see these things everywhere on the web today. They have one job, which is to deflect incoming support cases, and hopefully resolve customer issues. And we just know flatly, like, if we ask these things something complicated, we all know how this ends. Like, they all fail at this job that they've given. They fail to deflect those cases, and they mostly frustrate customers in the process.
So what I'm going to do during the course of this demo is what we want all of our customers to do, which is swap those chatbots with a new Agentforce agent that will be so much more capable of understanding what customers are saying. And to do that, we're going to have to, of course, swing behind the scenes and roll into Agent Builder. This is where all of our customers are going to build and customize their autonomous AI agents, and they're going to do it in a very different way, very different than chatbots. I've built chatbots, and it's exactly what Clara described. It's just webs of if-then-else logic. And what we're going to do instead is just equip these agents with topics which are essentially the jobs they could do, and then we're going to use that reasoning engine.
I'll show it to you here. I'm going to just ask this, like, "What's the status of my order?" No dialogue trees. I'm just giving it a list of things it could do, and what's going to come back here in just a few seconds, is the agent is going to make a plan using that reasoning engine for doing this. You can see that it's classified the topic, it's executed a flow behind the scenes, and it's used that to compose a relevant response. But the best way to really understand how this works is to give this agent a new, a new ability, essentially, a new skill. Which means what I really need to do is I need to add a new topic. Now, I have built these a lot.
We have this little pithy statement we've been using, which is: If you can describe it, Agentforce can do it. I'm a coder. That doesn't really. That doesn't rattle me much because I have been trying to describe to technology for two decades how to do work. What's different here? What's different is we're able to use natural language instructions. I'm not going to make you watch while I complete all these fields, but I am going to zoom into this because this is really important. Actually, the way that I've heard this phrase, the best version of this I've heard came from our customer, Kevin Quigley at Wiley.
And he described this as basically, you have to understand who are the people in your company that can describe what good looks like, like, how that job gets done. And those people often are not coders. They're gonna use a different kind of language for describing what good looks like, and now we can just put that language in here. There are also subjective things like if the customer is mad, how do you write that in code? And these are the kind of things we can express now using natural language instructions, and that's what you're looking at here. So what I'm gonna do, I've just given it a bunch of instructions and a description of the job to be done, and I'm just gonna hit Next.
Now, the next screen, it's asking us, okay, I've described how to do the job or I described, yeah, how to do the job, but this actually equips it with the ability to do that job. These are flows, these are Apex classes, these are prompt templates. These are the business logic our customers have spent years building in Salesforce. We have the metadata that understands how their work works inside their company, so I can just equip this agent with a few of those flows that they already have in their organization to give this a new skill, and just like that, I now have equipped my agent to not only schedule orders but also schedule appointments for those orders, so I'm gonna close this, and I'm gonna type just a crazy prompt.
I'm going to ask it not only to fetch my order, but like. I'll zoom in so everyone can see this here. So I have an order arriving soon. My email address. Here's my email address. I'd like to schedule an installation for Friday after delivery, but not morning. Like, just let that wash over you. How much context is needed to understand what was the order? When was this being delivered? What is the Friday? What are the available schedule times? Now, watch this. I built this up a whole lot. Let's check and see if it works. So this is. It's going out to that reasoning engine again, and here it is, reasoning on top of that.
And we can see that like here again, it classified the topic, it found the different orders, and it composed a response based on all of this. So this is just a new action, and now it has the ability to actually answer this based on all that business logic and context inside of our company. So I'm feeling pretty good about this. I feel like I might want to activate this, but we know our customers need to do so much more testing with this stuff. And it's not just me doing this kind of ad hoc testing that I'm doing here. You want to blast this with lots of tests. So this is something that we demoed just earlier today. We're calling this Testing Center , and so I just typed...
We call these utterances, like one prompt, and you saw that it could do it, but now let's just blast it with a bunch more. So what I can do is go up here and use the large language model, in this case, to generate lots of variations of that same job. And so it's gonna go out to the large language model, create a hundred different variations, and blast that stuff across this to make sure that it works reliably, so that we feel good about actually activating that. So I feel good. I think I am gonna activate this, so let's do it. And now I need to replace that chatbot. So what I'm gonna do is head on over into Experience Builder, scroll to the very bottom of this page.
I'm gonna open up my components, and I'm gonna find somewhere towards the bottom here, this, Embedded Messaging component. And this is going to... Whoops! My, my page. I let it sit too long. I knew this might happen, so hold on. Just give it one second. I'm gonna refresh this page, and then I'll scroll back down to the bottom, and then I will drop this component, and it should default to my agent. I'm gonna publish this. It'll take just one second, and then I'm gonna head on over to that web page that you just saw. There's the old chatbot that I had once upon a time. I'm gonna click Refresh, and what we should get here is a brand-new service agent that now can reliably answer just these wild questions like this that require so much more understanding of the question being asked.
I let it sit too long, and that's my fault. So if I were to refresh this, you would see that it can now do this. The bigger point here is not that we just want to let our customers swap these, these chatbots. That is, I think, where a lot of our customers will start. But as they experience that, they're gonna learn a bigger lesson, which is what we said: If you can describe it, Agentforce can do it. And companies have so many more jobs to be done than resources available to do those jobs, and that's when they're going to roll into our new agent experience, and it's not just gonna be a service agent. We are s- we are shipping so many standard agents, and these more and more are coming online, and we will be shipping these every...
Like, we have a really ambitious workflow, roadmap for this. And our partners are going to be shipping agents of their own that our customers can discover here. But of course, our customers can also create their own agents to create these things from scratch. And just as you've seen me do before, they'll do that using natural language instructions. Just describe the job you want this agent to do, and when they click Next, what's gonna happen is something only Salesforce can do. We're gonna look for semantic similarity across their entire corpus of data, like all the data, all the metadata.
We understand they've spent years teaching Salesforce about how their business works, and because of that, we can take and just a description of the job they need to do and map that into their business data and processes and enable them to create agents that can reliably do jobs in their organization. Thanks so much. Question?
Okay. Thank you, team. That was awesome. I'm going to invite Clara and Patrick back up to join me, and we're just going to move to Q&A now, and I'm sure you all have a lot of questions. I certainly have been getting the questions, so what I'd love to do is just open the floor, and people can fire away. We have some mic runners that will bring a mic over to you. Let's go with Kash, why don't you start?
Thank you so much for doing this event. I had a question. It looks like the value proposition will click with a customer who has been trying to do this on their own, DIY. They fail, and generally, that's how applications markets are created, and the people try to do this with a BEA back in the day, fails. How close are we to the point where customers have tried to do this, they failed, and they desperately need an application software company with the best practices, and they come in, marching in, and they're like, "Okay, thank you"? How close are we, or how far are we from that tipping point? Thank you.
I think we're in that moment now, because I think every big company and every tech company has tried to do it, so they're disillusioned, and that's why I think we're seeing all this demand. And then I think everyone else has been waiting because they're not sure, because it seems so daunting to DIY. Maybe they haven't been able to afford the upfront cost of buying a private instance of a model, and so this is also very promising for them. So we're seeing a lot of demand also in our SMB segment. I just got a note from Adam Alfano, who runs that segment for us, and it's—they're relieved, right? Especially in concert with our new Foundations offering. Now it's easy. They can get every access to every part of Salesforce, including Agentforce.
Great. Brent? Second row. You got it. Oh, Alex, right behind you.
Thank you. Brent Bracelin with Piper Sandler. It's interesting around the pace of innovation at Salesforce, and I'd just love to understand how you've been able to activate the team and move so quickly. I think the R&D team published a large action model paper just in May of twenty twenty-three, and now you're actually in production with these agents. So maybe walk us through behind the scenes, how are you driving the pace of innovation, what's changed, and how you've activated some of the innovation that we're seeing now?
Sure. It's been just an amazing two-year journey.
Mm-hmm.
It's hard to sum up, but I think there's been three things that happened. You know, first is, we went through a period of time that was really rough for the technology industry. Everyone had to course correct after overhiring in the pandemic. And, but then ChatGPT happened. And when ChatGPT happened, we kind of looked at each other-
Mm.
-and we thought: Oh, my goodness, like, we were made for this moment. We've been training our own large language models since twenty eighteen. We have this prototype that we've built with Gucci that they have in production. We already have this. We were made for this moment. And so it became this rallying cry for the entire company, and so I think that was one. Two is Mark taking back over as the sole CEO. Maybe you guys have seen the founder mode meme. We're in founder mode. And I'm a founder. You know, a lot. There's a lot of, like, entrepreneurial people. Like, Patrick was at a startup before. Adam Evans, who's the product leader, he was a founder that we acquired. We're in founder mode, and we were able to...
The combination of the culture change, plus this excitement around being able to take to market what we've been working on already, it just caused everybody in the company to get super fired up and shift from shipping three releases a year to shipping every two weeks. It's, like, really, really intense. So I think that's the second piece. And then the third has been just the customer response. And it's, like, very rewarding when customers tell you that they really need something and that what they're getting from other vendors isn't working. Like, it's just very motivating.
If I could, I just wanna add a little bit to this as well. You know, a lot of what's enabled us to move so fast is that we kind of re-remembered that we're a platform company. We have some incredible applications on top of our platform, but we've always been a platform company. And if you look at the history of our company, Mark had a slide on this. We've always been able to move through these transitions very fast because of that underlying platform that's there, whether it was moving into the cloud, where we built the platform, or moving into social or mobile, and then recently, data. And it's because we take what we have. We're never really starting from scratch, and we just have to wire it up.
We have to wire it up. And when the whole generative motion happened, we kind of looked at it from a different lens, right? We looked at it from a platform lens, and what that enabled us to do is, we think, get to the right answer faster than others. So it wasn't just the pace of innovation, it was also that we decided on an approach, which was not to go out and try to beg and borrow and steal, to buy GPUs to train models. Instead, we tried to figure out how to use our platform to connect the data to the model in a safe way, and then we were able to do that, and so, boom, the innovation. And, I mean, literally, boom, the innovation kind of happened on top because we didn't waste all of our time on doing that.
I think, you know, the pace has been incredible. I think Clara and I are probably a little bit tired, and many, many hundreds and thousands of other people. But man, watching the energy today. If you haven't been over to the launch zone yet, I'm gonna, you know, mention that a few more times before we finish. You definitely should get over there, 'cause watching what customers, the speed at which they're able to deploy because of the platform is incredible.
... Yeah, I, and I'll just add one thing, and just from a tactical standpoint, we are very, very focused right now on R&D prioritization to ensure that we are putting as much muscle behind the effort as we can. Because I think everyone, what you've been hearing, everyone understands the level of urgency on it. Mark's very focused on it. Our leadership team almost ad nauseam, we spend time on it as a team, which is, I mean, speaks to the importance to the company. And so with our most valuable commodity, our headcount, you know, we are all in at every turn we can be. Mm-hmm. Yeah, go right here in the middle.
Michael Turrin , Wells Fargo Securities. Thank you very much for doing this. I'm wondering if you could just help us with what you see as optimal fit or sequencing between Data Cloud and Agentforce. Are those discrete decision points for customers? Is it a necessary or better prerequisite to start with Data Cloud, and just how that customer conversation with the interplay between those two goes?
Yeah, I mean, the answer is it depends on the customer. If the customer has, either has Data Cloud already or separately has their data ducks in a row, then they don't need Data Cloud. They can buy Agentforce, and Agentforce has Data Cloud bundled in, in a sense that it's where the outcome data and the metadata gets stored, but it's not a separate implementation or a separate purchase. It kind of is like it's, it's more of an implementation detail behind the scenes. So some customers are doing that. I'd say most customers, though, I mean, you know this industry, most companies do not have their data ducks in a row. They have a lot of silos of data. They have multiple data lakes, multiple warehouses, multiple applications. Some of them are on-premise, some of them in the cloud. And so for those companies...
That's why I think we're seeing so much demand for Data Cloud. It's just a much easier way than migrating everything into one, which you'll never actually be able to do, of virtualizing, harmonizing, unifying, and activating in a really short amount of time.
And I'll just add to that as well. From a use case, it does depend, but from a use case perspective, when you think about what our customers actually wanna start doing with Agentforce, it's going to be increasingly difficult to find a use case where you may not need to access what we call unstructured data. So voice, chats, audio, video, documents. Organizations have just petabytes of documents, right? It's the vast majority of data that's in the world is unstructured, right? CRM data, very structured, useful data, but it's the other 80% that you really need to kind of build these use cases, and you can't do that without Data Cloud. Data Cloud is absolutely vital to that with the vector database that's there.
That's what makes that RAG process actually work, and it's absolutely critical to, to really using this thing the way you might want to within your organization.
Yeah, right behind you, Ox. Yeah, Brent, right behind you. Or you go there, that's fine. Go ahead, Raymond.
Sorry, I stole that. Raimo Lenschow from Barclays. Can you talk a little bit about your strategy to kind of see the customer base? What I hear on the floor is that there's a lot of, like, people, customers get their agents already very quickly here now and then just kind of walk home with them. I'm just wondering, like, how you do it, and what's the idea behind it? I think I get it, but, like, what's the follow-through then? Thank you.
The product isn't generally available until the middle of October. So we... But we didn't wanna wait. We wanted people—'cause sometimes you hear autonomous agent, and it sounds like this crazy science fiction. So as Mark—to use Mark's words, "We wanted people to put their hands in the soil and to really experience it and get hands-on." So our amazing engineering team, they spun up this ability to create these prototype Agentforce that customers can get up and running in 10 minutes so that they can really experience it, and they're gonna keep playing with it. We're gonna give them continued access up until October, at which point we expect, and we hope, many of them convert into paying customers.
And maybe, if we could just take a moment. I've been looking for an opportunity. I'm gonna do it now. Sanj, could you come up, and maybe... I gave her a mic ahead of time, so you can see what's going on. But Sanj's been over at the Agent Launch Zone all day. She was there with when Mark came over today, and he built his first agent with who was it?
Mm-hmm.
NBCUniversal. Maybe just tell us what you're seeing over there.
The mic? Can we turn her mic on?
Is there a button? Or a different mic?
All right, thanks. Thanks, Patrick. Yeah, we were in the launch zone all day today, and I think the most impactful thing is seeing folks really building this for the first time with us there, and I think there's a couple reasons why it's been successful. The first is that we have experts at every single table having one-to-one conversations with folks. It turns out deploying AI is uniquely human, and, our customers really wanna know that we're in it with them together, and we can guide them through this new era. So that was, you know, one key unlock for them. The second is that, as Clara said, this stuff sounds like science fiction until you can really see it in action.
You know, a couple of our customers that I'm really familiar with because they send me pages of product feedback all the time, I ran into them in the zone today, and one of them, BACA Systems, who is a large-scale kind of manufacturer, they had. They're a trailblazer there. Their admin was a really early adopter of Prompt Builder and Copilot and sent pages and pages of product feedback. I mean, he's been really dialed in and engaged, and he came up to me, and he was like: "Sanjna, I couldn't break it." I was like: "What do you mean?" He said, "I tried to break the prototyping agent, but I just did, and I couldn't.
And I'm blown away." And it was such an aha moment for him because he could really build this prototype on the platform, see it in action, share it with his colleagues in a matter of really just 10, 15 minutes. So, it's really, really exciting, and as you said, Patrick, it's reminding people that we are, in fact, a platform company. It's a great moment.
... Great. Thank you. Go, Brent, yeah.
Thanks, Clara. It's Brent Thill with Jefferies. Can you speak to the halo effect that you think this will have on the rest of the portfolio? So if I start using sales, is this gonna require-- What else is this gonna require? What do you think in the next year, is this gonna pull through that maybe you haven't seen in the past?
Sure. I mean, so there's two things. First, we're seeing AI play a really big role in selling our core clouds. I can think of a number of customers who are here, who've been on stage with us, where they were doing an evaluation of CRM systems, and starting last year, they said, "Well, our number one evaluation criteria for CRM is AI." And so it's been great for us, right? Because we're leading in AI, and it's really made the difference in those CRM decisions. It's. We're also seeing it sell more of the cocktail of everything that we have to offer. And the reason is because, as you start to delegate some of these tasks to agents, we're starting to see the human reps play different roles.
We've had multiple customers where their customer service reps, because they've offloaded a lot of the basic service questions to Agentforce, now they're able to take bigger roles, like a sales role or a marketing brand storyteller role, and so they need to be able to access different data, different actions in those other clouds, and it's pulling through the purchase of those other clouds.
I think this also has a potential halo effect or kind of a creation effect on areas and categories where you may not have typically seen Salesforce play. And it's because of this platform that exists, if your data is there, if you can get your data connected and you can build on top of this, you can quickly start to see how you can use this for HR workflows, for IT workflows, for finance workflows. There's a number of these other kind of front-office systems where we haven't typically played, where I think suddenly now we have a little bit of license to go in and say, "Hey, why don't you take a look at this? This is an interesting way to start to accomplish these types of processes.
And I think that's why we're seeing the Workdays and the Googles and the Boxes of the world want to join our Agentforce partner network.
Yep.
That, that's really exciting. And then the other last part of it is we're pulling through Data Cloud. I mean, as we talked about earlier, I always joke with my peer who leads Data Cloud. I'm, like, driving so much business for him, he owes me one.
Let's go to Mr. Murphy in the back. Alex, right next to you.
Thank you. Mark Murphy with JPMorgan . Great to see a product release that has such a visceral response at Dreamforce. It's really been something to see. I wanted to ask you how you think about the cost of serving the interactions on the back end, especially relative to the $2 per conversation pricing. And so, as an example, how often... You know, when we look at what's happening on screen, how often is it calling a foundation model, right? Where that might be a little more expensive for you versus calling on kind of a small language model that might be pretty efficient. Does it cost less if the customer has adopted Data Cloud?
I was also wondering, when we see that kind of multi-step reasoning, it reminded me of the newest GPT, the o1, you know, the Strawberry model, is what we've seen out there, that feels like it could handle that kind of a request.
Mm-hmm.
You know, Friday, but after the installation, but not in the morning, you know, that type of thing. Can you give us a little window into the cost structure and what's happening on the back end?
Sure. So I can answer the first question first. I don't know if there's a way to bring up our LLM leaderboard for CRM, but if not, we can send it to you afterwards. So what we found is, unsurprisingly, there's so many new models out there right now that for most enterprise tasks, you do not need the state-of-the-art, most expensive model. It's just, it's like, kind of like bringing a bazooka to a, you know, arm wrestling match, right? It's just way too much for the task at hand. It's way... I mean, there's latency issues, performance, and cost.
And so the way that we've architected our AI stack from the beginning is it's an open architecture, which means that customers can choose what model they want to use, whether that's OpenAI, or it's ours, or it's Anthropic or Google or Amazon or Cohere or open source models. We're seeing a lot of customers do that with either Mistral or the Llama. The Llama models are excellent. So all this to say that the model landscape is constantly changing, and that for any given task, there's not one model to rule them all, because you'd be overpaying. And so we're playing that very crucial role of first collecting data on what models are best for the right task, and then second, the models that we can run in our service, we're tuning those models, including our own Einstein.
But we're not just tuning our own, we're also tuning Mistral, we're tuning Llama, and we're basically able to route each task to the lowest cost, highest performing model, and that's how we're able to create advantage over time.
There's another really important point here as well, which is for as well as any of these models perform, for as good as any of them ever get in reasoning, they still have to have your business data.
Right.
You have to be able to connect to your business data. So it's wonderful that you have a model that might be able to reason that morning is earlier than afternoon, but if it needs to schedule an appointment for a customer, you've got to, you've got to have it connected to that customer data in a meaningful way. And maybe I'll get myself in trouble for this, but I two weeks ago, three weeks ago, right before they launched the 4o model or the o1 model, ChatGPT had an announcement about their new enterprise agent builder. And when you look at their release post about it, you'll see a screenshot in there, and it's showing building a forecasting agent, a sales forecasting agent with their new model.
The screenshot shows you literally uploading a CSV of your sales data. I mean, are you gonna do that every day?
No, you have to have a data platform that is wired up and connected to those models. An individual model, the world has been obsessed for the last eighteen months with what's the best model? How many parameters does it have? The reality is, we're going to build this as a system. It's going to use many individual small models. It's going to be an optimized system that's connected to everything. You can put the fastest engine in a car, but if your car is eight thousand pounds, it's not going to accelerate very fast. You've got to have a fully optimized system, and that's really what we have here.
Yeah, and so back to the Strawberry model. We, because we have an open architecture, if a customer wants to use that, they can. You know, xLAM actually performs better in CRM benchmarks, so we're not, we don't anticipate that happening, but maybe Strawberry will improve, and customers can choose that. And so that's the bonus of having this flexible architecture.
Yeah, and I'm going to actually peel the onion one layer deeper, because we get the question very, very frequently on the $2 per conversation. So I just want to set some context to make sure everyone understands. Think of the $2 per conversation as a way for us to simplify pricing to the customer. What's happening behind the scenes is you've got basically a mix of conversations. We're all P times Q people. You get a mix of conversations. Those different flavors of conversation that are occurring come with different intensity levels. So if someone wants to reset a password, it's easier. If someone wants to rebook a flight among three cities on three different days, it's going to get really complicated really quickly.
And the intensity level will dictate what the cost profile of that looks like to us, in addition to some of the other dynamics. And so what's happening behind the scenes is we are, from, for now, at least, assuming a mixed level of those conversations, which feeds how we're analyzing the price. The $2 per conversation is a list price, right? So there's volume discounts and things like that, that go into it. More importantly, from an economic standpoint with the customer, and the dialogue we're having right now, is the economics of the model and how they think about it. Because right now, Service Cloud being the poster child of this, they're looking at it and saying: "Okay, I pay 100 agents. I'm paying them $50,000 a year.
If I can start to deflect more, there could be savings that will fall to the bottom line." That comes across in the form of interactions times a price per interaction or a cost per interaction, is how the customer is thinking about it. But it's very collaborative conversation happening with customers right now because they see the value, and so it's leading to, I would say, healthier conversations than we've had in a while from a Salesforce customer standpoint. Yeah, Rishi, right there in the middle.
All right. Wonderful. Really appreciate all the detail. Rishi Jaluria, RBC. You know, as you go back to thinking about Salesforce as being a platform company, maybe let's kind of fast-forward what Agentforce can look like in a few years. Is there an opportunity not just for all these partners, Clara, that you laid out, but maybe brand-new companies to come out and be built on Agentforce, and that to become an effective monetization opportunity, just as what happened way back in the day with the Salesforce1 platform, you know, now it's a huge business stream. Thanks.
Yes, we-
Great question.
We anticipate that, and we're also talking to founders now about that, and that's one of the big reasons why we just announced earlier this week our new $500 million additional GenAI fund, which brings our total to $1 billion. The first fund invested in a lot of model companies. Now we're seeing a lot of applications and agent companies that want to be formed, but don't want to build out Hyperforce, and don't want to build out their own Data Cloud, and don't want to build out their own Einstein Trust Layer. And so I think there's a lot of... There's vertical plays, as you know, as Patrick's mentioning, different departments, traditional workloads that we may not have operated ourselves, but that companies are going to be founded on that basis.
Great question. Yeah, Kirk.
Yeah. Thanks very much. Kirk Materne with Evercore. One of the other foundational aspects is kinda not surprising you didn't bring up. I'd be interested where you see it maybe in a year, would be just industry context. Meaning, you have company data, you understand the profile, and that brings a lot of great context. But I think to unlock in certain industries, a conversation in financial services could be very different than retail.
Mm-hmm.
When do you start layering that in? How fast can you bring it to bear? And how fast can that help companies get to a level where they see an even bigger unlock on this?
We're doing it now. Along with launching Agentforce this week, we also released over 100 industry-specific agent actions. So if you're a bank, we've pre-built a financial disputes action. If you're a hospital or healthcare provider, we have a version of after, you know, the after-visit summary that's automatically generated. And so there's a lot of similarities, but a lot of differences. When you're in the medical field, there's certain code types for the different medicines, for the different procedures. You have to get that right, and so that's what that industry cloud team has done. And every industry cloud at Salesforce, we have 15 of them, they've each shipped those out-of-the-box industry actions.
Yeah, right here in the... fella right here, or books gone.
Trip Chowdhry, Global Equities Research. I would like to have a follow-up question to Mark Murphy's. If we look at our pricing model, I mean, if we distill it down, it is all about pricing on which is based on tokens. And then if you look from that perspective, current state of art is Byte-Pair Encoding from OpenAI, which everybody kind of uses. Now, if you look at the research that is happening, is they want to reduce the size of basically tokenization so that you have less tokens. Now, if you go further from there, you will see the context length, that is how much you are trying to put in your prompts, is increasing.
Mm-hmm.
Now, you go further from there, they are being cached, so that they don't have to pay that much.
Mm-hmm.
Now, if we go back and backtrack, then if our pricing is based on tokens and the vocabulary size is going to reduce, caching is going to happen, and something better. Actually, at Stanford University, that research is really happening. They're trying to reduce the tokens by a factor of 70, like, reduce it by 70%. But if your current pricing is based as $2 per conversation, one fine day, you may be losing three-fourths of your revenues just because of technology innovation. My question is: how do we price it in such a way that you encourage the innovation, encourage people to put more longer context?
Mm-hmm.
Because there are two aspects. You ask a question, and you generate the answer. You don't want the answer to be also unlimited because of the cost. You want full freedom to the end user.
Mm-hmm.
How do you think you can manage all these things together? Excellent.
Yeah, let me start, and then you all should keep me honest. So I would first say the way to think about the pricing is not as a fixed price. I know it sounds like $2 a conversation, but really it's a usage-based structure underneath it. And so the way to think about $2 is, we're giving you a bucket of credits or tokens, to use your words. Within that, right, there is those different flavors of conversations I mentioned earlier, and those different conversations burn down against those tokens at different rates. The key to it, when we talk with customers from a value standpoint, is that really, at the end of the day, it's a value-based pricing structure, right?
And so you get to a zone where customers are saving X amount of dollars, we're gonna capture Y of that. And what it's leading to is a very healthy, actually, conversation right now, where we can do our internal modeling and get to a percentage of the savings that we think can be accretive to our business. And then there's a conversation that happens with the customer, where they can do all the math, and then you have a very open conversation about what they're willing to, quote-unquote, "share back." Now, it's a negotiated price, so I don't wanna make it sound like there's some fluid, value-based structure where we're, you know, based on what they do, we're gonna get paid, right? It's gonna be based on the actual consumption that's occurring. But I would tell you, based on the initial conversations.
Again, it's early, so the sample size is small, but the willingness to pay it for Agentforce is much higher than anything we've had thus far coming through the AI funnel. Because the value and the applicable use cases for the cost savings and the ROI that the customers are seeing is much, much higher than anything we've seen in the past.
Yeah, it's really the $2 is really more for simplicity, for customers to have a back-of-the-envelope. But of course, we'll cap the actual number of tokens and then do all types of cost-saving things on the backside.
Right.
like using small language models that we fine-tune in caching.
Right. Yep. Yeah. Terry.
Terry Tillman, Truist Securities. So earlier, I think you were talking about sales and service being very actionable in terms of the agents. Commerce and marketing have been a little bit slower growth, as of late. What about agents to help maybe, you know, kind of jumpstart the growth in those businesses and any specific use cases that seem actionable in those segments? Thank you.
Yeah, we've launched those, too. I don't know if we can get you the full list, but we have a merchant agent, a buyer agent, a merchandising agent, as well as a concierge agent-
Mm-hmm.
To help every customer get VIP service, right, from it, from that agent. And so we do. And we have—we've seen lots of interest in the marketing and commerce areas. I didn't mean to not call those out. I was just trying to focus on sales and service for the sake of time, but they are building on the same platform. And the beauty of taking a platform approach is they can take advantage of the same Atlas, Data Cloud, data security, Trust Layer, and they've also built a number of custom actions as well.
So with that, we're actually at time, and so I want to keep everyone on schedule. I wanna thank you very much, Clara and Patrick, for joining us today. If everyone... You know, you're welcome to grab another drink as you leave or grab some food. I hope everyone makes it to Dreamfest. That's why we're ending when we are. The DJ at Dreamfest starts at 6:00 P.M., and then I think the main acts come on at 7:00 P.M. or something along those lines, so again, I appreciate you all coming. Feel free to reach out to myself or anyone on my team if you've got follow-up questions, and we're happy to take them. Thank you.
Thank you.
Thank you.