I was hoping to start a little bit with your background. So you came on board to Salesforce vis-à-vis the acquisition of Airkit, which you co-founded back in 2017. And now Airkit has been integrated into the Service Cloud. So maybe for people who are less familiar with the story, you could tell us a little bit about Airkit, why Salesforce was a good home for Airkit, just so we could get a little bit of an introduction to you.
Thanks, Keith. First time at TMT, it's great to be here, and just to kind of personally introduce myself a little bit, so Airkit was my company that Salesforce acquired in 2017. However, my story with Salesforce goes back a little longer. Marc would call me a boomerang. It's my second time. My first time was a company called RelateIQ that was acquired in 2014. The reason I bring it up is that RelateIQ was all about using AI. Back then, we called it data science or machine learning, really to automate sales, to help sellers capture what they're doing with email, calendar, and solve data entry problems. This became some of the Einstein kind of IQ type stuff that we had a decade ago, 2014 plus. Airkit, I left, started Airkit. RelateIQ was about sales. Airkit was all about service.
So this is instead saying, "Hey, let's take AI now and think about long-tail customer service applications, specifically self-service, and thinking about helping customers just serve their customers 24/7 using technology to do that." And so we joined in late 2023 back into the Salesforce family here. And then as far as how the, I think you mentioned how the technology, the product, and things have actually been kind of brought in, what you're seeing is Agentforce is a mixture of a lot of the IP that came from Airkit plus homegrown. It's really important to us that we don't, including my own company or things that we're looking at, have technology that's brought in. That is what we're launching. It needs to be very deeply integrated to our platform. Otherwise, it's not really going to have the effect that we want.
So a lot of the IP from Airkit and new tech that we built is kind of the culmination of that as Agentforce. And that's also why customer service has been kind of one of our very first areas. I see a lot of our customers interested in building self-service agents serving their customers 24/7.
Got it. Got it. So I was hoping to start talking more broadly about sort of the broader Salesforce AI platform. One of the benefits and hard parts about having Marc Benioff as your CEO, Marc gets very excited about new initiatives like Agentforce. And I think the robustness of the solution took a lot of investors by surprise. And I think part of the reason why that was a surprise is people underestimate how much there is in the foundation below that, how broad the AI platform is. So when you think about the Salesforce AI platform, which you have perspective over, talk to us about sort of the extent to which AI has been embedded into the core solutions, to which there's a foundation which enabled you to quickly kind of roll out Agentforce, if you will.
So there's a lot of dimensions to that question. Let me just kind of start getting into it here. OK, so Agentforce and what we built is a mixture of effectively out-of-the-box agents that mirror kind of our lines of business or clouds, right? So service, sales, commerce, marketing, et cetera. Those are the easiest things to understand and buy. Customers typically start there. It makes sense. But it's also a platform. What we're finding is agents and really, as we say, things like digital labor in a much kind of bigger kind of market we're getting into, this technology isn't just solving a singular problem. It's really versatile. And so we think about having agents doing all kinds of long-tail type background work. We're going to make some announcements tomorrow, by the way. We're doing our developer conferences happening right now at TMT as well.
So we'll make some announcements tomorrow. You're going to see agents move in all kinds of workflows for employees, for customers, solving problems, being able to leverage data that's in Data Cloud for us. Really, unstructured data is a, I don't want to say goldmine, but it's something now that we can do so much more with all of the data, transcripts that we're having, interactions with our customers, understanding click traffic, being able to apply AI to better understand customers, to do things like put them in the right segments for better marketing campaigns and more. There is just a very, very wide kind of aperture of the application space. So it's not just kind of a line of business only. It is truly a platform. And the only way that we could do that is having it actually deeply integrated throughout all of kind of the stack.
So I appreciate what you're saying that it was robust and that we're moving very quickly as we're here. That's the pace that we want to go. We've got a bunch of other exciting announcements planned for this week and also for, of course, Dreamforce, but more things to come.
Got it. And maybe we'll kind of walk through kind of the evolution of the capabilities. When we were talking about AI within Salesforce 18 months ago, roughly 18 months ago, the focal point was Einstein GPT, which was embedded into the solutions. Can you talk to us about sort of the core capabilities that come from an Einstein GPT versus the extended capabilities of what you could do with Agentforce and help us kind of compare and contrast where we've come?
There's kind of like even like a third distinct moment. We had a little copilot phase in the middle of that too. The Einstein GPT is, I think, what every kind of company kind of wakes up when they realize the world is changing. They say, how do we get AI into our product? The same thing for business. How do we get AI into our business? How do we change our operations? The most obvious way is to look kind of like screen by screen or workflow by workflow, task by task, job by job, and say, hey, how do we kind of add it in here? It's the most immediate thing to do.
Put a prompt that helps me close out my case if I'm a customer service rep so I can move on to the next one faster in a standard way with a prompt. We see a lot of adoption of that, but it's a simple product because it's very kind of linear. It's like one kind of solution at a time. That kind of embodies, I think, a lot of the Einstein GPT kind of suite of functionality. Then we moved into this idea of assisting employees and thinking of it that way where it's kind of always the human in the loop. And this was 18 months ago or so. And this was, I think there was a lot of kind of trepidation around this idea of, can AI actually help my customers? Can I trust it? How do I make it not hallucinate?
A lot of things have changed since then. We also talked to our customers early on, and they were saying, hey, we really want to use this, but we're a little scared. That was kind of the temperature about 18 months ago or 12 months ago. I mean, maybe still is, but what's shifting is once we actually deployed the technology, and Keith, you mentioned it was pretty robust. So they actually got their hands on to an early version of now what we call Agentforce. They saw what the technology would do as far as assisting a human customer service rep with the human in the loop. They said, and they were kind of blown away. They said, wow, actually, this is pretty good. Maybe we could turn this on all the way to our customers.
And so we started seeing that appetite for, hey, maybe we can try this. And then that's when we just started going kind of heavy into it and said, great. If you're going to have this be an agent, meaning it's more autonomous, meaning it can work without a human in the loop, it can take action, it's got some tasks, some background reasoning, it needs guardrails. It needs to be able to follow policy. It needs to be able to do that in a consistent way, in a measurable way, because that's the difference between consumer tech, like let's say a ChatGPT or something like that, versus applying this technology to actually take action with a business, with an enterprise. So investing heavy there, and then that's kind of what Agentforce was launched last year.
So it was October, I believe.
Yeah, October was the GA.
Yep. Got it. So still relatively early days. I think from the investor perspective, what we focus on most is the Agentforce opportunity within Service Cloud, one of the first ones that had come out. Can you just talk to us about the scope of where you guys have agents today and where that's going to expand to over time?
Yeah, so I think service is maybe the most obvious, as you're pointing out, kind of almost like a low-hanging fruit, but it comes up a lot. Another really kind of big interesting use case is really about more on the employee side, like internally facing agents that are doing basically pre and post meeting prep. So this is obviously good for sales, field service too. If I'm going in as a seller to a meeting, or if I'm a technician showing up on site or whatever, I might want to have a pre-brief before I do that. I might be able to want to ask some questions with it, so it's a little bit of an assistant, and then also kind of post activity capture, things like that, what are the next steps, kind of updating the CRM. That's kind of another really large one.
You're going to see we're also teeing up a handful of new features. I'm trying not to say too much about what we're going to announce this week. Let's go around it. You're going to see agents be able to do things that potentially in a more proactive way, thinking about being triggered more from data events and as background workers and kind of basically analysts coming in. And that is kind of an extension of this idea of a pre-meeting brief, where it knows my calendar and is doing a little work ahead of time. But you can imagine scaling that out for all kinds of operations within a business, not just like a meeting with a seller, but all kinds of various events that may be happening with a customer and things like that as well, that's coming through the platform.
Got it. Got it. I want to drill into the Service Cloud use case. And correct me if I'm wrong, but from our perspective, we see that as, and you said low-hanging fruit. And we kind of agree that that's a use case that seems, I don't want to say relatively obvious, but it seems like very ripe for automation. It's a core cost center for a lot of customers. So the ability to do that more efficiently just resonates a lot with them. What's the capabilities in terms of Agentforce's ability to actually offload call volumes, offload support requests from a customer today?
Yeah, well, so let me give you an example in health care, actually. So we have a company called Prisma Health. Actually, Prisma Health is a good one. They're actually using Agentforce also due to pre-briefs where they're going in. They're doing chronic disease management as they're meeting with providers to onboard new sellers, but also for customer service, it's chronic disease management. So you have a lot of touches. Another customer, 1-800Accountant. They're doing like seasonal, it's tax season coming up in the U.S. So ramping up seasonal labor. They're using Agentforce right now. They're seeing a 50% resolution rate. A story that we like to tell, you probably Marc tell it. We use it internally as well, or I should say externally, but we use the product ourselves for our help.salesforce.com. You can check it out right now. This is a high volume thing.
We've got about 40,000 conversations a week. The last stats I saw were 84% deflection or resolution rate of agents helping. And then when an agent is engaged with a customer, I really love this stat, only 2%. It's right around 2% or 3% actually escalate to a human, which is important, by the way, to have. And that's what Service Cloud allows. So if an agent's talking, how do we bring a human in the loop? It's important that you have it, but it's impressive that you don't need it. So I think those are really great kind of stats. We see this also through a lot of our other customers. They're all looking at 30%-50%. It varies depending on the business and how they want to kind of dial up customer satisfaction outcomes and how they think about handing things off.
For example, carving out some things like you're going to talk about pricing. Maybe you want to hand that off to a human. It's not even about this idea of a cost center and deflection. It's almost more topical as well. I think some of the measurements that contact centers have been using for years in the past, as you really kind of scale up this much more kind of like limitless supply of labor to handle cases all the time, you're going to start seeing the kind of different measurements and people really rethinking cross-sell, upsell, CSAT, what are the most important customers, how do we actually move them and maybe into my best human agents, and then rethink it a little bit differently. We're seeing 50% plus deflection pretty consistently.
Got it. I also want to dive into, actually, I'll have you do my job for me and answer some of the questions that I get from investors a lot. I think one of the biggest questions is understanding the buy versus build decision. Companies have a lot of their data in a Snowflake or a Databricks, and Snowflake and Databricks are giving you capabilities to build out agents for yourself. Why do this within Salesforce versus kind of building it on your own? How do you guys think about that buy versus build decision for the end customer?
Great question. We just had an industry analyst report, Valoir, come back and tell us they surveyed our customers and they came back knowing that customers that were building something, and then after they had built it, switched over to Agentforce, or they did some kind of a bake-off internally, which happens a lot, 16X faster. 16X faster to market is what they told us with our customers. I was really happy when I saw this report, but it checks out because it's never been easier. Let me just give you a little insider. I'm like an engineering kind of product nerd background. Let me tell you something, a little secret, which is it's never been easier to build a cool demo. That is what this technology really unlocks. So beware, but also it's cool to build a demo. Moving that into production is much, much different.
And so when you go down this path that every Fortune 2000 company had 18 months ago or whatever kind of thing went off on AI, let's go, let's build. Let me get 10 or 15 people. We're going to go build this internally. You saw a bunch of cool demos come out. And then when you say, great, well, how do we keep it on the rails so it doesn't do things that we don't want it to do? And it's got trust, it's got reporting. And by the way, when we build it, do we have a whole bunch of the whole idea of testing these agents before they go live is a whole nother thing because they're probabilistic. So you don't have just a typical test where it passes or fails. You need to do a portfolio of tests and like a threshold.
And so we built a testing center, an automation, and really the entire what was traditionally like a software development life cycle, is what people would say. We're calling it kind of like agent development life cycle, this entire suite of effectively tooling, not to just create a cool demo, but actually field this technology in your business. Oh, and by the way, you need analytics in your reporting. How much has this team done? How much value is it doing? How do I know how to improve it? An iteration once the agent is live, solving a problem and measuring the ROI. This is the depth of investment that it takes to really do this well.
This is what we've already built and we're building, and we've got an exciting roadmap about it, just accelerating people through this process of buying in their data and organizing it, testing it, deploying it, improving it, measurement, like the whole thing. This is like our roadmap, basically. That is the difference. That's the 16X lift, and you mentioned some other companies that are maybe more data-centric. You're going to want agents to do more than just analyze data in the background. You're going to want them to ultimately reach out to employees, and there's no better place to do that where employees are already having conversations. Slack is a very, very popular destination for that. You're going to want these agents to be able to actually have business impact, to talk to your customers, to send emails, to be able to change marketing campaigns.
I mean, where the actual business is operations hits kind of this long tail of a reality with customers. That's what Salesforce has always done, is provide systems that do that. We're just going completely through our stack and saying, how do we bring agents, or really agentic reasoning, and really kind of this beginner's mind of thinking how to remake these experiences, solving the original core problems of helping businesses run better, thinking about that, and then thinking about the tools that you need at a platform level to do this well at an enterprise grade, not just for the developers, but also for the admins, for the business owners that need the KPIs and more. And it's a lot. It's a lot to build this. And it's not most of our customers' core business. And so it's the build vs. buy, right?
It's the classic thing kind of happening over again.
Got it. The second big question I get is you guys initially launched this with a per-conversation price point, $3 per conversation. Why is that the right way to price this type of solution?
Yeah, so we talked a lot about customer service as being kind of the most obvious thing. That pricing really comes from that mentality. The average cost of a contact center touch, some people may know this well, but others may not, is somewhere between $7 and $9. It varies wildly, by the way. $20-plus, you get into insurance with deep policy understanding. This is a cost center, right? It's an expensive thing for a contact center. So when you think about a 50% deflection rate and as good of CSAT results that we're seeing our customers, when we got it 80% plus for us and our help, where you're charging $2 for something that costs you $9, it's a no-brainer. Customers like it. You do it all day. So that pricing kind of comes from that.
Now, when we launched the product, we just talked about there's all these other use cases. We have an SDR, a sales development rep product, that takes leads, knows your sales plan, does research on the leads, and then does outbound emails to them, so sales teams will take maybe the top 10% of their best leads, give it to their best human sellers, and then the longer tail and give it over to the SDR. This is something that we've seen kind of this expansion in non-service use cases, and so we're thinking one of the temptations was kind of this à la carte pricing for every possible agent. And we had a big debate about this, but what we realized was we didn't want to have, let's call it 10,000 excuses.
This is way too important for companies thinking about building your agentic layer, all these different agents that talk to each other and work. We made a decision to take that pricing that was kind of coming from service and apply it initially to these other use cases. It kind of feels a little weird, like $2 a conversation for an SDR, which is really a lead that it would take. That's where we're at now. We're moving, if you kind of heard in our, I think Marc mentioned on our earnings call last week. We're moving into the concept of a universal credit with this. This was our first step one to kind of combine. Really, it's not about conversation. It's about work. The kind of work you can deploy these agents for is massive.
There's lots of fragmentation here that we're planning to help our customers, like well beyond CRM use cases. So the Universal Credits allows for that flexibility. It allows for CIOs that are making decisions around budget. There's a lot of uncertainty about the future right now. Tech is moving fast. What do I do? How many licenses do I need? What's my team going to be next year? I'm going to buy this or that. Or what's the business case on that? Are we sure we're going to do that? Or is it going to yield this? Right now, the answer is you can buy. You can invest into this idea of a Universal Credits. Whatever your model was, we're going to deploy agents first for customer service.
And then you realize that there's another and another and another use case, which we see constantly, by the way. Once you kind of get your feet wet, there's lots of use cases. In fact, it's almost too many. It's almost a problem that can be focused. But as you expand, you have these credits that you can try experiments of new things and learn. And that fungibility is really critical at this moment when things are changing so quickly.
Got it. Third question, the durability of that price point in terms of we see token pricing coming down consistently. We hear Sam Altman, Uncertain, and Satya Nadella talking to us about 10X improvements in price performance. Can you guys sustain a price point at $2 per conversation if that token price continues to draw down over time?
Absolutely. One, we got to create value. Two, I think what you're going to see, even though things like the inference cost, like token price, things like that are dropping, which is the technology. And you look at hardware. I mean, we're on this innovation curve with it. We're seeing it play out in the market. But what's also happening is you think about agentic reasoning. And for the academics, they'll call this stuff like test time scaling or inference time scaling. What this really means is deeper reasoning when the problem's happening actually yields higher performance with the agents. It's not a one-time model hit with inference. It's multiple. So there's all this flexibility that's happening. One price is dropping. The reasoning and the output's going the other way. I think that this is still kind of early in it.
We're absolutely excited about inference prices going down, specialized hardware coming. There's a lot of plans on all this stuff. It's fantastic, but it's ultimately about what are you using this technology to do? What kind of value does it create? And can we be the easiest path for our customers to unlock that value, and then we capture it.
Right. Got it. So it's not really about sort of the comparison versus the token prices. It's the comparison between if you're going to have to do this yourself, how much would it cost to drive the same amount of value as what Salesforce can provide to you sort of out of the box.
That's right. And by the way, as the concept of let's just use the word intelligence instead of tokens for a second. You can run this machine that can kind of reason and think through data to make decisions. You can apply intelligence. It drops to lower and lower amounts. There's a couple of different ways you think about it. One, but the way I think about it is the number of use cases goes way up. I mean, we talked about long tail. There are things like workflows that agents are going to be doing that we don't have necessarily jobs right now for. There's a lot of talk about just that. But there's a huge amount of value that's completely untapped because, quite frankly, we can't even fathom it to hire the people to analyze.
Imagine analyzing every customer conversation to create something like maybe specific promotions or something that are deeper insights or analyze your business about having recommendations back to executives about how you might be able to have changes that are literally the AI's recommending for you to then research and then click on and deploy back in a flywheel. Things like this are going to get unlocked as you see continuous inference price drop and intelligence get effectively more abundant. And then all the systems to do this as more regulations come out, by the way. You have more agents talking to each other across boundaries with interop and all this where we're going to get into. These are the reasons this is part of the value add of Salesforce, back to that DIY type thing. So inference cost is important use cases.
And then we're going to have a healthy margin with that. But also, we're providing so much value being able to take this technology and actually deploy it to useful things for our business and measure it. That is a lot of value right there.
Got it. So this is kind of almost a derivation of the Jevons Paradox. You're going to bring the cost down. It's going to expand out the use cases of what you're going to be able to do with this.
100%.
I think we're on number four. Number four, a big question that I get. There's a netting equation that comes from this. Service Cloud today is sold on a seat basis. You're looking to digitize labor, which is going to mean extensively the savings are going to come from. You're not going to need as many call center reps to kind of push this out. Does it net positive for Salesforce? Are you going to be able to make up on the consumptive side of the equation what you potentially lose on the seat side of the equation?
Yeah, absolutely. We're seeing it. We haven't put any numbers about this. We're seeing it today. It's also early moments. One of the things we're also doing is having what we call a flex agreement. So you can kind of move between perhaps unused seats or seats that you're unsure of what you're going to do next year with. You can start moving that into this universal credits and start understanding this kind of dynamic at a customer basis. But look, what we see is just because you're saying, here's what we do today. The most low-hanging fruit, the Einstein GPT moment of the most simple way to bring this into your company is here's what I'm doing today. Can I do it with AI and automate and save money? That's very clear math, and does it work and these things?
As soon as people do this, they go, well, what else can we do? And there's more and there's more and there's more. And so this is what we see almost universally. And also, so there's a lot more use cases that arise from it just from necessarily doing that. And I'll give you an example. Customer service. If I call in or email in and I'm talking about, let's say, a product that I need that didn't show up at my door, this is a classic thing is where is my order or WISMO in retail? And they say, hey, look, it's been two days and it might be missing. Your neighbor might have gotten it. You should wait two days before we can file a case or do a refund or something to that effect.
The idea of, from a contact center perspective, the cost center mentality is that you want to close that ticket or that case that you're measuring on with as little average handling time as possible and then just say, call us again if it doesn't. But the customer experience would be better if two days from then there's like a proactive notification and saying, hey, Keith, just wanted to check out if that package actually arrived. Just want to make sure that you got our new shoes. And actually, do you mind letting me know how you like them? And would you mind going ahead and putting it on social media if so about how great this experience was? That entire concept. But I think any consumer would say, that's a better experience.
The current mentality with labor and these costs and the cost center mentality is, that's going to throw our KPIs off. That's not what we want. But as soon as you start realizing that you can put agents into all these additional workflows, you start rethinking your business. I mean, this is what we see. So this is an early moment. The very obvious thing is A to B, like one to one, like, oh, I can do this and automate it. But then the next thing is, wait a second, I'm going to apply this to doing things that were never possible before.
Right. Got it. Right. So this last one's where I sound like a jerk. But remember, I'm just the messenger here. These are just glorified chatbots. Voice capabilities are relatively limited today. The functionality doesn't go well beyond what you could do with a chatbot. Can you talk to us about kind of where we are with voice capabilities, but more broadly, maybe some examples of where this goes significantly beyond what you would get from a chatbot a year ago, three years ago?
So let's talk about multimodality, like voice in particular. One, we're adding in image type stuff as well for multimodality, document understanding. But voice is something that we're very, very interested in doing. We kind of gave some sneak previews of this. But you can imagine that there's some announcements coming out about that. So you can talk to these agents. It's great. They can email. They can chat. They can talk. You build them a launch, you test them in the testing center. You launch them. They can do all the things. But let me give you a different example about something that I've seen with reasoning that was kind of like was a wake-up moment for me, a little shocking, which was it happens to be a service example. But we saw a phone call come in. It's a retail situation.
A phone call came in from. It was like a. I don't remember if it was UPS or, actually, it doesn't matter. FedEx, UPS, where a delivery person was basically calling the company because it's on the box and saying, "I can't deliver a package to this address because the address doesn't exist," and this is a voicemail. It's coming into the contact center. You see the call, the phone number it's calling from, voicemail transcription, and it basically said, "Here's the document number on the box and blah, blah, blah," and the document ID on the box or the reference ID or whatever it was, these aren't terms that the business uses for an order ID or any internal system. This was a call from a non-customer trying to deliver a package, reading off things that are printed on the box with no clue what to do.
And the agent then listened to the call, understood what was going on, looked up, tried the document ID and the numbers across a couple of permutations. It tried, "Is this an order ID? Is this a thing? Is this an internal ID? What is this thing?" Because there were no instructions about it. It found a match that it was, pulled the order up, confirmed that the address that they had said in their voicemail was the order and that it wasn't it, and then proposed a plan to communicate out to that customer, letting them know that they needed to call the 1-800 number that was left on the voicemail for, I think it was UPS, to basically keep their package moving on time. And this all happened in under 60 seconds.
Seeing that kind of a navigation of reasoning, and even though there was a conversation over voice, in this case, voicemail was coming inbound and then outbound kind of email draft for a plan, I was just like, this is not a chatbot. You think about so that's reasoning and crazy capabilities. Also, again, multimodality and more. But this is, yeah, it's much bigger than that.
Got it. That is a great example. I want to talk a little bit about Data Cloud.
That was your hard question, by the way? That was not hard.
I didn't say it was hard. I was just saying I was being a jerk. Why is Data Cloud so important to sort of this overall initiative? Why is this such a critical foundation for what you guys want to do in Agentforce?
Yeah. So data and AI go hand in hand. I mean, AI can reason, but reason over what? So you need to kind of connect in data. I think that's a story we all understand. But Keith, earlier you were talking about the DIY thing. And you mentioned Databricks, Snowflake. These things are like, well, I'm investing here. How about there? So the Zero Copy Alliance that we have with Data Cloud, what makes Data Cloud pretty unique is that our enterprise customers can come in and say, we've been doing all these strategies or we've acquired companies that have different data lakes. And we want to build agents and do great stuff to modernize our businesses. Do we need to put all that stuff and rebuild it in these huge multimillion-dollar initiatives and over years of work?
And the answer is actually you can bring a lot of that data in Data Cloud really quickly from these disparate places, like break the silos down or whatever you want to call it. So that's really important strategically because that allows us to deploy, to bring data together faster for our customers without having to do these massive investments from an IT perspective. And then that data is connected very deeply to Agentforce so that as it queries through all the unstructured data, we've got vector database and RAG techniques. This is what kind of people call it. The ability for the AI to find and understand unstructured information. That's what that really means. We have techniques that show basically a 50% lift in precision and recall on this stuff that we do.
This is like what we call our Atlas Reasoning Engine as part of that IP. So as soon as that data can get from everywhere into Data Cloud, we've got this great tech and this really nice coupling with agents to be able to do that. And when it accesses this information, you're going to have a lineage of exactly what data that agent touched for that customer for all the things that you're going to need to understand in the future, not only to tune it, but also to be able to say what happened when you need to have a report on some of these things. And so it's a very, very connected, both from a product perspective, from a strategy, from a speed and productivity perspective about deploying solutions. It's, yeah, I don't know, like peanut butter jelly, whatever you want to say. It's really close.
Got it. So Service Cloud is one of the core starting points, but the portfolio is much broader than that. Maybe you could walk us through kind of some of the early customer journeys you've seen. How is Agentforce going to kind of pervade across an enterprise? What pulls it across? And who's doing it? Is it Salesforce helping to pull that across? Are you bringing in systems integrators? Are the customers able to do this themselves? Is it intuitive enough that they're kind of building out these new capabilities?
Yeah. So it's a mix across segments and industries and a little bit of the techniques. But it's kind of like a little bit all above. Let me touch on that. So we have 5,000 Agentforce deals in Q4. About 3,000 of those were paid. I think that's what was publicly announced. And we also have self-service in there too. The reason I pull that out is like, who's doing the work? And we've got this kind of ability for people that really understand what they're doing. It's not that hard, but they've got the right team to come in and just solve the problem. We have large companies that have a lot of kind of stakeholders, a lot of meetings, things like that, where there's a lot of decisions about how to apply this technology. They need advisors.
This is where the system integrator and partnerships come in very well. And there's a lot of people, a lot of questions. A lot of things are happening right now with tech. How do we think about this? How should we apply? They need advisors. And then we also have our product lens to that, where we're having AI literally help you set up your AI. So as it's ingesting data, understanding your workflow, your objective, self-configuring what we call our topics and actions. This is like saying, here is a job that an agent should do, and here's how it should do it. And then once you describe those jobs, the AI literally helps you write those jobs out. And then after you've done that, it can write its own tests.
So it says, "Oh, well, here's 100 different examples of what customers could say and how I would answer them." And here's what I think the answer should be. And then it gives you a spreadsheet of it back. And it says, pass, pass, pass, pass, pass, fail. And you can read it and you go, "Wait a second. This is giving me the confidence." So there's an acceleration across all these things, whether it's partners, whether it's just the product itself, self-serve tends to skew a little bit lower segmentation-wise. And in enterprise, a lot of our customers need those SIs to kind of help out.
Got it. Yeah. So maybe just to wrap up, 5,000 transactions in Q4, 3,000 paid transactions. But a lot of those are still kind of proof of concepts. Help us kind of set our expectations in terms of, given typical enterprise sort of adoption cycles, what are we likely to see through calendar year 2025? And then how does this expand further as we go into calendar year 2026 and calendar year 2027? How should we think about the roadmap for Agentforce?
Yeah. So lots in there. I'm not going to give you any specific quantitative expectations.
Mike Spencer is OK.
Mike Spencer is OK? OK. Well, then, no. What I would say is, look, the deals are great. I love saying that high-level number. But just so you understand, I wake up every day and I look at the funnel of implementation. I see that that's the most important one. How fast can we get these customers live to solve their problems, drive consumption? That's what we're doing. The roadmap is all about that. I mentioned the agent development life cycle. This is also code for top of funnel, bottom of funnel, and then understand what it can do, what it's doing, how to improve it, actually go back for improvements and get those implemented and keep going, more use cases, et cetera. We're building product along that whole way.
I've touched on some of those things a minute ago about testing, automation, reporting, analysis, that stuff in the product side of the house, but this is where we're going, and to give you a sense of what's possible, I mean, we've seen large companies. It's not like a one-size-fits-all. We've seen large companies like Saks Fifth Avenue actually stands out. They went live with an agent, and it was under 10 days. It's on their website doing return orders, and I thought to myself, there's no way a big company like that can even have a meeting or two in 10 days, and they went live in 10 days, so can the technology do it? Absolutely. What we're really talking about is, do you have a strong CEO or top-down directive to do it? That's kind of what made that happen at Saks.
But also, kind of, do you have good data? Are you kind of set up? If you're already a Salesforce customer, you've got all that stuff connected right there. It's much easier for you to turn the switch. But even if you've got your data in all these other places, as we mentioned, you can pull it in as well pretty quickly. So what's technically possible is you can go really fast. But it's really going to be a little bit of a case by case and what the use cases are. But every single one of the customers of these 5,000 deals and growing, we want them live. And it's not just for our benefit. It's for their benefit. They want them live.
We're working to do this through all of those mechanisms we talked, SIs, product, self-service, opening it up to just be easier for developers, all kinds of things to just keep the acceleration on.
Got it. Fascinating times in Salesforce right now. Thank you so much for coming and talking to us about Agentforce and the opportunity ahead.
Thank you, Keith.
Yep. Thank you, everybody.