AgentForce Decoded: Inside Salesforce's $80B AI Revolution

Speaker A: Welcome everybody. This is Ryan and I am back with a very special guest. Today I have A.J. kumar. A.J. is the head of AI product growth at Salesforce, specifically responsible for looking at the deployment of enterprise scale adoption of agents across complex business environments, typically with what the top 1000, 2000 clients is heavily deep into agent four. So I'm like super, super pumped to have AJ on the show because I've been trying to find an amazing guest from Salesforce to have and then he kind of popped up, man. So welcome aj Happy to have you on the show.

Speaker B: Thanks Ryan. Nice to be, um, with you and sort of seen some of your other guest shows. So really excited to be here.

Speaker A: Love it, man. Yeah, I think that's how we got Connected was on LinkedIn. I think you said something about the podcast. You're like, yeah, I love some of the guests you have. I'm like, you would be an amazing guest. You gotta come on. Anyways, um, that was one of the things where I was like, okay, we gotta let this rip. Do you want to give like a real quick backdrop on yourself and kind of like what you're focused on now? I know I gave a brief overview and then I want to start getting into like Agent Force because like I said, there's a lot of talk, a lot of momentum around it and like one of the reasons why I was super excited for you is because like you're in the details, you see what's working really well and so I'm really excited to pick your brain on that and share it with the world.

Speaker B: Awesome. Absolutely. Thank you. So, um, with Salesforce for about eight years now, I've done a bunch of roles from distribution to customers product and now uh, more into the AI product particularly. And what's really exciting is as we are rolling out our uh, new feature capabilities of AgentForce. I'm actually a part of the Agent Force scale up team and the product organization which allows me to help see where the gaps are really and what do we do to really drive it up and really the innovation of the product that is steading beyond the product roadmap and how our customers use our products. And prior to that I've done a, uh, startup gig and before that I did uh, a decade at Microsoft building products again and uh, doing advisory stuff. So really excited to be a part of this scale up journey as we go into this new wave of AI with all the things around agents in particular.

Speaker A: Well, I am too, man, and I love how you worked your way up the organization over a period of like close to nine years right from being more on like the pre sales support all the way now as like a key cog. Not a key cog. Cog sounds like a bad word but like key, key person of influence involved with the agent for scaling and so like why don't we kind of kick things off there. I would love to hear what you've seen as the most surprising way clients used agent force that you would have never have expected.

Speaker B: Yeah, 100%. Um, I think, I think particularly we are, where we are deeply excited about is the art of possible because when we said we will launch these agents and probably this is a story that goes behind the scene. Our CEO Mark Benioff was so excited about the vision that there's a point in time until last time last year's Dreamforce which where he wanted to rename the company as Agentforce from Salesforce. The board actually voted against it I'm sure and uh, that's how serious is the focus and the area of investment. So think of Agentforce product team. Right now it's about 70, 80 people. It's like a startup under the large 80,000 employee company. Um and we are trying to pull the whole company forward in this wave and I'm so glad to be part of that 85 member team. And we, what we are trying to do is do a lot around our natural uh scope of the product. And coming back to your question, what really are some of the surprising use cases when we started launching this we typically would launch from how agents could be used by a ah, sales customer or a service use case or a marketing use case et cetera. But customers today have gone beyond the traditional use cases of sales service marketing where typically would have been asking questions hey can you guide me in the sales process et cetera to actually go and take an action. One of those examples I can probably take is an example of uh, Royal bank of Canada. Royal bank of Canada is uh largely uh one of the national banks in Canada but they also manage advisory for high net worth individuals who are typically always almost all the time not based in Canada but spread across the globe. The CEO of Royal bank of Canada reached out to uh US saying hey, what if there is an advisor who's actually able to communicate 247 an example is such an I sitting in Singapore but a Canadian citizen, let's say just before midnight they have an idea that hey I want to invest about a million dollars in a uh, PE fund and you do not want that idea to flip overnight when they go to Bed. So what if that advisor is available to them on call just not to make and finish that recommendation but just to keep them hot enough until the human in the loop comes back in the morning. So that's what we have done with Royal bank of Canada's financial advisor agent. So that was a very interesting use case. Ah, similar is with Goodyear, the tire company we've got got an agent that actually is being deployed in that public scope which if you're a car manufacturer you could typically understand different scopes of the tires that are there in place and what then becomes a natural extension of their procurement cycle. One other example which is probably very close to my heart is based m in London at the Heathrow. At Heathrow we've deployed an agent on WhatsApp which is typically started off as a customer service agent but today uh, you could ask questions like hey, I'm transiting from this terminal to that terminal and I have some time to kill, I need to do some shopping for my 3 year old. How much time would it actually take me to finish the process? So it pulls data from different data points and is able to give you a very specific answer. So these kind of questions cannot be hard coded. But what we are deeply excited is about the vision forward. Like because we end up working with a lot of other airlines as well. Singapore Airlines, British Airways, American Airways, you name it. Probably 90% of the airlines industry we have an agent deployed right now. Some of those airlines now want to talk to the Heathrow agent and they want to have a use case. Like if a flight is getting delayed and the passenger before they actually arrive into Heathrow, the agent should have sorted out two or three different flight options and a human in the loop before a human gets involved to make that decision. Those flight paths should already be charted out. And this is again intra agent communication between multiple airlines coming into the airport and then going back. So you can imagine the complex network of communication channels. But that's how typically we can solve a lot of those intraocacies.

Speaker A: Love that man. Really, really great use cases. I can imagine. Yeah, if someone's going to make a nine figure PE deal at uh, midnight they might have questions or thoughts that go through it. So there's a uh, good, I could see the utility for that. Um, and I love the Heathrow example too. I mean super pract, so tons of examples like that. Like what would you say is like the most impactful use case you've seen on the sales side and then on the marketing side I would love to Hear that? Because like, before we get into the like the walkthrough of like how it works and what it looks like, because that's what a lot of people are trying to crack the code on that I work with and.

Speaker B: Mhm.

Speaker A: So I think it'd be great to hear what you see just knocking the COVID off the ball in those two areas.

Speaker B: Yeah. One which is probably I'm a computer science, uh, engineer, uh, by education. And one of those books that I used to read back in the day when I was in uni was the Wiley Book Publishers, the computer science publishing books that Wiley publishers. So Wiley came to us about six to eight months ago saying, hey, we do have a lot of these college students reaching out to us during the seasonality of the university, uh, tranches that they come particularly fetching for code samples. You know, some of these code samples are written by the book authors and then it's stored somewhere in the public cloud on a particular folder. But as the older version of the book gets older and older and the newer version comes in, the older repositories have not maintained well. But still these students would have paid a full premium or money to actually buy those books. So 90% of their queries are about code samples. Hey, this code sample is not working. But while he doesn't actually own that code, it has to go back to the author of that book. That author would have moved on, he would have been somewhere else. Wiley used to take about six weeks to resolve a ticket which involves a complex. Hey, this code doesn't work. I tried, I used this, but I'm still getting this error. So they will not bring the author in front of the student. They will do the handover and they manage all the communication. So you can imagine the time to resolve a very complex. My code sample is not working. I paid a hundred dollars for this book. That sort of a ticket, average ticket size, uh, time duration was about six weeks. We've cut it short to less than six hours. And that is us cutting short a lot of handover processes which were manual by putting an agent that can almost tackle some of these lookups of information without a human actually getting involved. A human will typically be bought in when there is a real complexity of like, hey, this book is outdated, we don't support that book anymore, something like that, or hey, let me get you a discount code. You bought an older book, why don't we give you a new book? So we've almost got them about 270% increment in number of customers Coming back with repeated success. And I think that's a massive success we have seen on a customer service use case. But now that is transitioning into a more cross sell story. So they're able to upsell to those same students or candidates over a period of time. But again, you understand Wiley is a very seasonal business. They do not have this spurts every time during all parts of the year. But whenever there is a university window opens, new students join in, candidates switch from one semester to the other. There are these humps of growth that they see and we are very glad to be there with them during those

Speaker A: spurts of growth with the school year. Yeah, so what are they cross selling then? Are they cross selling additional books then when they go through that customer service process, is that kind of what it's enabled?

Speaker B: There's newer topics of books and I think particularly the previous author would have written something some on older topics. I probably there's recommendations being given as well right now. And uh, they're also talking about what else could be the extension of that Wiley agent. Can it go and talk to the university as well? Agents can typically do much broader. It can do deep research, it can do look up on the web as well for more information. So we are uh, trying to bring in and enrich the first use case was that for Wiley Agent. So now we are trying to increase the scope of that agent to more.

Speaker A: And what about on the marketing side? Because that's kind of sales orientated, right? You got customer service sales. Anything that you're seeing that folks are loving on the marketing side?

Speaker B: On the marketing side today is largely a uh, combination of the personalization element and then the combination of what you can do around targeting those. And Salesforce has this product called Marketing Cloud which is about personalization tool. Over the many years, for those people who have not seen or never use Marketing Cloud, they would probably be like, oh, how can you do personalized segmenting targeting for a vast segment of array of customers? And for a brand like Air India, which is one of my home countries from where I am, it actually can target very personalized messaging for that brand. And they do this while not just servicing the customer, but also offer them options around holidaying and tourism. M and this is a combination of how we brought together both the sales and the marketing journey for them. And they understand deeply about hey, do you have kids in the family? Uh, this is your anniversary. So they can actually personalize that. And as the agent is naturally converging, probably say hey, I see that you have an Anniversary coming up in two months time. Do you want to book a travel? We can give you a 10% code right now if you do that today. So this is selling in the flow, but it's a marketing agent.

Speaker A: Yeah, yeah, yeah.

Speaker B: Nice.

Speaker A: All right, so I, I wasn't going to ask you this and there's nothing wrong with me asking, but it just came to mind as you were telling all these stories. So this was like as of recording right now, there's a couple weeks ago where ChatGPT announced an integration with HubSpot.

Speaker B: Right?

Speaker A: Yeah, yeah, I don't, I know we don't want to say the HS word while you're on. Right. We had same thing. Is there going to be an integration coming out like with, with Salesforce? Is that something that's on the radar? Because a lot of people are asking

Speaker B: that very good question and I think, I think this is typically a point in the industry where we are seeing different branches of those conversations happening, technologies coming together and partnering with OpenAI in particular, we have been working with them um, deeply as one of our most trust model providers within Agent Force, uh, to an extent where uh, within Agent Force Today there are two layers at which OpenAI is actually being used. One is at the public model level layer, but at the second layer you probably, for those who have seen the Agent Force architecture, something called as the Atlas layer, uh, which is the brain of Agent Force Atlas reasoning engine. So the Reasoning engine is again today powered by OpenAI, while we definitely want to extend it to other partners. So our relationship with OpenAI goes long, long back. But what, what you typically saw there with HubSpot is a more headless approach of integrating HubSpot within OpenAI as a scope. And I think those partnerships are naturally going to open up for us too because what we are trying to do with AgentForce right now is we've opened up the Agent API, which is headless, making it headless. And then now we are adding our own MCP capability. That is AgentForce will become an MCP server. And once AgentForce becomes an NCP server, you can plug it into any public model interfaces, you can plug it into OpenAI, you can plug it into Cloud Desktop or Gemini.

Speaker A: You name it Gemini. Okay, awesome. So what's the timeline on that? If you. I got to ask, right? You don't have to tell me, but I got to ask.

Speaker B: The MCP server roadmap, uh, is on the horizon. So before, just before Dreamforce, which is another three, four months away, we are actually going to be launching A couple of MCP capabilities. And with that what we are just also testing is because MCP is an open source protocol, like Anthropic developed it, but they open sourced it. What companies have to realize is there's one thing with mcp you can open up three different things. You can open up tools, resources and instructions and prompts. But what we are very carefully doing is how much of tooling, how much of prompts and instructions and resources do we open up? Because it's also about the criticality of the guardrails and security of the data at the end of the day. And today when we integrate with OpenAI the, we are very productive of not sharing any PII data that's stored within Salesforce. And that's where it's a boundary check condition for us right now as we roll out those MCP servers. So just before Dreamforce, you'll probably see us doing a bunch of those announcements.

Speaker A: Okay, nice. Well, we're really excited to see that, so very much.

Speaker B: Yeah.

Speaker A: Do you want to show us like I know there's some, some pretty cool things on the AgentForce side that we talked about before the show started. We'd love to see, you know what that and share what that looks like. So why don't you share your screen so everyone can see what I saw?

Speaker B: Yeah.

Speaker A: And we can we go through some, some, uh, some.

Speaker B: Yeah, absolutely. So one of the things while I'm bringing it up is the own homegrown agent force deployment that we call us probably Customer Zero as dog footing our own technology. And the first time while I bring that up, it's on our public health portal, health.salesforce.com today we receive millions of hits, particularly customers coming and asking for hey, I need to find this document. I need to understand how to migrate this process. So what you're about to see, uh, just before we go there, this is actually the AgentForce metadata platform which is sort of built from the scope of what we have been building for the last 25 years on top of our SaaS layer. And of course Agent Force relies heavily on everything that's built around our cloud infrastructure and all the other different clouds come around to operate and data cloud becomes one of the key foundational layers for Agent Force. And uh, we talked about different sets of agents, Sales agent, marketing agent, service agent. There will be different Personas of those agents coming into life as we evolve further as well. But what I want to show you is uh, a scope of what's available publicly and you all can go to help salesforce.com and try this agent yourself. I've authenticated myself so that I can actually ask more specific questions even without authenticating. You probably can come in here and ask a bunch of questions. So I'm going to ask a uh, question like hey, is data cloud needed for agentforce? I get asked a lot most of these times because they say hey, I have my own data platform because if you remember this one, I use databricks, I use Snowflake. Do I still need Data Cloud for AgentForce? Well let's see what the agent answers to that passion. And what you're about to see is an agent that is authenticated my own uh, login and it's going to tell you what the customers are going to miss if they don't have Data cloud. And while it's trying to pull some of this detail, what you can see is it give three very specific pointers of uh, data Cloud, why data cloud is needed and what are those key pointers. But you can also see that it gives some additional link. These are uh, probably what are like to call as citations if you folks have used tools like publicity today it actually gives you a link of hey, I found this resource but this is actually where that originated from. So today we can actually pull that origination story from deep linking of our own documents. So this help.salesforce.com was built by indexing over 700,000 different help articles. And it's about a six to eight months of process to build it up to this level. And we have about 92% accuracy on this agent today. One thing I'd like to call out is people definitely talk about three key pillars before rolling out into this agent takeout, AI world, people, process and technology. There's a fourth P that people don't talk about much which I shout at the top of my voice. It's patience. Like these things take time to actually build an agent and then get it up to a certain level of accuracy. The first time we built health salesforce.com, we ourselves were at about 50% accuracy. And then it took us a few more months up to the eight month mark to actually optimize how we set up data. How does this agent consume those data points and then fine tune that and bring it back. Now we have about 93% accuracy with over 75% to 80% accuracy. The call deflection that doesn't even land up in our contact center with human agents being involved. So we've saved a lot of time. And this is a story that we are trying to translate it back to our customers.

Speaker A: All right, like that man, I like the patience. Uh, it's true. Everybody wants everything yesterday. So that makes sense. So what was the other dashboard you were talking about or Command Center I believe is what you said we're going to show.

Speaker B: So what we launched, uh, what we launched three days ago was uh, a newest version of AgentForce launch. There was a major launch in our uh, San Francisco offices, AgentForce 3.0. And we particularly announced three major key features around. This one was called uh, as the Command center today I want to give you a sneak peek of that command center. And then we talked about the favorite topic that you picked in what are we doing around mcp? How is our uh, MCP roadmap and other marketplaces coming up? And the last one was like what are we doing to drive faster adoption? And how can somebody build agents quickly using AI itself? So let us go and take a look into the Command center itself. So what you're seeing on screen is the Agent Studio and within Agent Studio I actually can go and invoke a particular menu item to actually launch Command Center. So within Command center what you're expected to see is a bunch of agents which are deployed. So you can see there are sales agent, HR agents, service agents, etc. So I'm going to launch the customer service agent. And what you can see here is at a very top line view, you can see what's really happening with my customer service agent, how much call deflection has happened and what is the average cost per resolution. Because remember at the end of the day you're all m, uh aligning it back to the dollar spent. So you need to know how much money, how much is the ROI that you are counting in and what is that bottom ranking topics, how many of those queries coming in are really not performing well. So let's take a look into some of that. So let me go into the cancellation request which is for. This is for a product centered company and I would like to know what was that really bad chat session that a customer had? And we're going to see why this was poorly scored. And what you're going to see as I peek in is, is a session log that tells me there was this particular query in the moment at which this was this latency of 134 milliseconds, just two interactions. And this wind chat only lasted for 24 minutes. 24 seconds. And you can actually see, yeah seconds. The quality score is very low primarily because you actually Saw that the customer actually downvoted the response. At the end of the E chat session or the agent search session, we ask, how did you like the session? Was this response accurate? And most of the times we want it to be accurate. That's where I said 90, 92% accuracy. Depending on how customers have set up agent force and the data that they give to agents, the response scores could be high or low. Continuing the journey back again into the agent studio, what you can see in the command center is a part of the case deflection. But here what you should definitely see is what you've saw is a sneak peek into one of these sessions. But we can particularly also go in and look at other details. All of this is really giving you a single pane of glass, which was previously not possible. And customers had to spend time across several different dashboards. Now we decided, let's pull it all together. Give somebody, even somebody within the company, like maybe the chief customers officer wants to know how their agents are performing. And what you'll see over a period of time is we'll be able to bring in how human agents and AI agents are able to perform in the same dashboard. So that's the dream and that's the vision that we are going towards. So this one is so customers can actually see a portion of this coming through their hands and their orgs right, uh, now.

Speaker A: Okay. Nice, man. Yeah. Love the visuals. I love that you could do the drill down and understand specifically. And is this like a voice agent that's handling those calls or what kind of agent is it that's handling the support calls?

Speaker B: Good question. So the voice agent is uh, actually on target to be GA for Dreamforce. So currently we are doing a bunch of pilots for voice. And one of the reasons why voice for us has been a delayed launch compared to like one of the questions that I get asked is, yeah, Jay, why is Salesforce not yet launched? Wise my voice is available on the public models out there. I can actually chat with my OpenAI or Cloud or Gemini. One of the reasons why it is so hard. Voice is such a hard problem to solve and we want to get it right, is because Salesforce has a lot of data that our customers store in their own respective arcs and that data is their business context data, which is not publicly available. But when we actually put a voice model on top of it, we want to be sure that we understand the modalities of it. And that's where a lot of extensive testing is going on. We are working with those customers to make sure working them with them through the pilots to see, hey, have you seen if this voice temperature modality suits you? Because some of these customers are also operating across different geographies of the globe.

Speaker A: Uh-huh.

Speaker B: What if as I'm talking to you in English, I switch to few words in French or Spanish just in a few sentences? The voice agent should be seamlessly able to handle that part of the communication. So these are the nuances that were thrown as curveballs. Like without actually forgetting the context of the conversation, the agent should seamlessly carry on. Today, if you try doing this with OpenAI, uh, voice model, you'll actually see that it will skip a context. We'll say, let me restart the chat, you'll see it doing that. So that's a very hard problem to solve. And we do not want a business context data to be lasting that.

Speaker A: Yeah, I've seen that happen. That, I mean that would piss a lot of people off if they're dealing with it, right? Because they're like, I just told you, like, I remember when I was doing that with a cable company and like, I, I would. These were live agents. These weren't even like fake. Not fake agents, AI agents, I should say, fake agents. Oh God, what am I doing?

Speaker B: Ryan?

Speaker A: Um, these were live people that were agents. Right. And they'd ask me the same question over and over, like, what's your name? What's your customer number, what's your address? Let me transfer this department and then do the same thing over. And it was like the most infuriating and frustrating thing in the world. So.

Speaker B: Right.

Speaker A: Um, all right, why don't you show us the, or did you show us the whole thing?

Speaker B: Is that everything on this, that was part of it. But what we've typically trying to do, I show you the command center. This is typically what is the agent Exchange, which has actually got, um, a lot of people excited. So we actually announced launch of 35 agent exchange MCP launch partners. What that means is we are working with these vendors to particularly get their MCP servers plugged into our ecosystem. So as you saw, we could actually see some of the well known entities there. This is actually going to become a much more open wide ecosystem so that our MCP access and their MCP access should become access that. It is very much easily a plug and play ecosystem. And then you can imagine the extensions of innovation that can continue beyond. Nice.

Speaker A: Yeah, I mean obviously this unlocks a lot, uh, when you have this capability and then what would you say, like, I guess since you work with like the biggest customers, what would you say is the most demand? You're seeing, like the top three use cases that you see patterns of people like, hey, I really want to use agent force to do that because I'm just curious, like, because I've heard use cases all over the board when it comes to this.

Speaker B: Yeah, I think the top use cases still largely comes around customer service for us today. And within customer service, I would say there's different tranches, like there's the uh, different levels of severity, but also different levels of complexity. In terms of being an agent, being able to take an action, can it just go and do those things? Like if I have an agent that is deployed for one of the world's largest retailers, and if a customer has to come back for a, uh, return of that order, how does that agent handle that in terms of going back, but also going back all the way to the logistics chain and warehouse to actually look up if there is even a possibility for them to take back the order? Because sometimes, you know, some of these orders are out of process and they say, hey, the vendor is actually not in the scope of returning an order with us anymore, so you probably will have to wait. And then, you know, you've gone through that process before. Uh, they expect us an agent to be able to do these other additional things. So at this stage I probably would like to classify the different types of agents that Salesforce works with or ships for our customers. Three different types. One is the foreground agents, the ones that are actually based, uh, which you'll typically see here. The other one which we have seen is employee agents which are internally facing. This could be your uh, HR agent, finance agent, which typically in our day to day you could use internally within your company. AI concierge. And while I say that today within the company Salesforce, we have over 52 agents deployed. And some of those agents talk to each other. Sometimes some of the agents don't talk to each other. I have a payroll agent. If there's, uh, the other day I had some uh, issues with my pay slip, so I had to go back and rectify some records. So I just sent a request and I found out it ended up talking to another agent behind the scene and then found out clarification and came back so I didn't have to do that and worry about the handover. So the background, uh, the employee agency, the other one, the third category is the background agents. Background agents is like uh, for those uh, who are probably outside of the scope of Salesforce, OpenAI launched about something called as OpenAI operator where uh, you could give it a task. Hey, I need to go on a holiday booking. Come back to me with a couple of holiday hotel and flight options. It will go do its research, come back to you with like, hey, these are the three options. Do you press on this button 1, 2, 3 for it to go into Expedia, booking.com, so on. Right, so those are background agents you fire and forget. You just give it a task, forget about it. It'll come back to you whenever it's completed that task. And that's the third category of agents. And uh, we've done some investments by building some of that, but we also acquired a business called Convergence.

Speaker A: Yeah, I heard. I used them before, uh, you acquired it. By the way, do you want mine? Stop sharing. So we can see your bright and shining face as well as my big shiny face. Yeah, so I was curious about that because I wrote a post about that that went pretty crazy, um, about that and I thought that was smart. And then you also. Was it informatica that you. It was like an 8 billion dollar acquisition. They were about a week and a half apart. I thought they were related. So yeah, walk us through that with convergys, like why you decided not ah, you but why the company decided maybe you're a part of it. I don't know man. To acquire them and what was the thought behind that? So their technology worked pretty good and they were a really small team. So I would love to hear that.

Speaker B: I'm sure you've used their proxy tool.

Speaker A: Yeah, yeah, yeah, yeah.

Speaker B: So that's exactly what it was. So let's talk about convergence for a bit. So convergence came into light when we were looking at what are the other different types of agent categories that our customers would want. And as I told you about the foreground agents, internal facing agents and then background agents was typically top of mind and proxy was one of those agents that they were building from a B2C scale. And these are really a smart bunch of techies here, all in UK, about 20, 25 of them with a great pedigree across different backgrounds that they came from. And we started talking to them, I think about the art of possible. And then we saw the larger vision of how we could bring them together and really work on that idea of extended scope of background agents. So in the near future, as we just started onboarding them and bringing them into our uh, closer technology stack, uh, you'll actually find us doing a lot more work with background agents.

Speaker A: Okay. Yeah, that's like uh, there's a big interest in the market with that. The whole concept of like running while you're sleeping, you know what I mean, where they're leveraging that and um, people just want to be able to push the button and have shit work in the background.

Speaker B: Right, right.

Speaker A: So I think there's. Yeah, there's a lot of massive utility for that. Very much some of the.

Speaker B: One of the. One. No, one of the customers I like to code because you said fire and forget are the sleep agent that could do the job while you're sleeping. Is Adecco is one of our largest customers. They, they are one of the world's largest staffing companies. They do uh, staffing for non um, it jobs primarily. It could be a plumber, carpenter, teacher, you name it. Like anything under the sun. And they receive over 10 million applications for jobs every year. Can you guess how many of them are actually screened manually by a human?

Speaker A: Out of the 10 million job applications a year? No clue man.

Speaker B: 4%.

Speaker A: Okay, wow.

Speaker B: 96% of them just goes to the bin, just goes to the bank. They can't even possibly look into those. And these are human.

Speaker A: Oh they just throw them away.

Speaker B: They thrown away, thrown away. They can't screen. They can't screen. And again when I say the can't screen like they look at the first screening of course happens, but then they don't keep them for future jobs. Like they just, it goes back and also possibly they do not have the time to fulfill these jobs for the end customer because there is so much volume of applications. Can there be something done around AI? So we started bringing this job screening agent for them which is live right now in UK as a first market. Very, very successful with that. They have been able to screen a lot more. The percentage is still pumping up but they want us to be able to do a lot more background tasks uh, in terms of being able to let's say hey, I want to have this new job category that I'm working for, let's say the uh, department of IT and they want to hire a bunch of these back office uh, uh, carpenters or construction workers want to set up this data center for us. But we have these 20,000 resumes that we just want you to screen and give us the best category of output. And we just need you to do this for a period of two days and that is a request. If Adeco uh, receives, can they just give it to an agent that just goes behind, takes whatever time it takes to come back and do this. The beauty of background worker agents is you do not have to worry about the idea of timeout, idea of API limits of something running out of the bounded scope of hey, time has passed and you've lost the connection or the agent has stopped responding because you just fire and forget async approach. And that's a, uh, very interesting area of problem that we'll be continuing to solve for a lot of customers.

Speaker A: Excellent man. Well, speaking about the future, uh, I'd love to hear your thoughts. I was thinking about this before the episode. I was like, all right, so A.J. knowing what you know about agent force, agent agents and basically AI as a whole, if you had to bet all your savings on one AI trend for the next three years, what would it be?

Speaker B: I would definitely see the ability for people to be able to do more around the second layer, which is really key because, well, we'll definitely start seeing a splurge of agents coming out in the market. No second thoughts. Uh, as Salesforce, we want to be definitely massively successful and capture a majority share in that. But we do also know that there will be other agents that customers and other people will continue to build. Developers will have their own agents, so there's going to be a plethora of that. The layer about that is where it excites me the most is the supervisory agent. There will be a, ah, non human entity that will be needing to manage a bunch of these agents so that they don't run out of scope. So think of this as almost a manager agent. And what our CEO Mark Benioff says, very exciting is this is probably the last line of executives in the company or anywhere in the world who only have humans reporting in. So within our org charts we will soon start having agents showing up and it has already started to happen at Salesforce and I'm sure other places too. So I'm really excited about AI building other AI systems. And I think that's probably the number one focus where I'm super excited about. I'm also equally excited about the opportunities that the world is doing around discovering a lot of other ideas around path to AGI. How far is it? Uh, what does that look like? But we like to call it as enterprise intelligence, Enterprise artificial intelligence before the AGI, mostly to what we could apply to our own businesses, I think. Yeah, of course the robotaxis are going at a different scale, so that's uh, not in our scope. But robots. Yeah. If you speak to Mark, our CEO, he'll say, yeah, I'm um, Deeply excited about what we could do with robots. So we have started doing a lot of that. So if you attend any of our conferences, you'll actually find robo ducks going here and there. But what they're doing is actually helping you. You can, uh, a question saying, hey, which is the next best data, uh, session that I can attend? And it'll ask you to scan your badge and it says, looks like you're a data developer, so maybe I'll give you a developer Centric data session as a recommendation.

Speaker A: So, yeah, I saw that. I went to Dreamforce last year and, yeah, saw those robots rolling around, you know, it was kind of funny, um, entertaining. So, okay, that makes a lot of sense. And the management of, like, digital employees. And by the way, Benioff's such a great speaker. He's very entertaining. I love hearing him talk, um, at conferences and stuff. I saw him and Jensen over at Nvidia. They were going back and forth. That was a lot of fun, seeing them talk to each other. Um, but you mentioned AGI and it was funny. I had an experience. This was literally yesterday, right? And I, uh, consider it kind of like my AGI experience, if you will. Like, I had my chatgpt moment when I. It was almost like my origin story, like I told you, where it's like, I was like, holy shit, this is going to change everything. And I had my same thing with, like, AGI and agents yesterday. So I'll tell you the story. It's. It's rather quick, but I want to hear if you had one of these yet or something similar.

Speaker B: Go for it, Go for it.

Speaker A: Wow. Okay, so this is what happened. I've never talked about this publicly. I talked about to one of my clients earlier today. So this is like breaking. So I was like, I'm in. I'm working through the Lovable Ship program, right? Where basically it's like, it's a, uh. It's kind of like a challenge. They have to create a product in six weeks. And one of the things that was Surface, they were all in week two, but was all about product research and ideation and validation and all the core components of product design, which I'm sure you're super familiar with. That's not my background. Right. So basically what I did is I was like, all right, how do I attack this? So I pulled up, uh, Perplexity Labs, right? Which is freaking amazing, by the way, with.

Speaker B: They're doing amazing. Yeah.

Speaker A: I pulled up chat CPT03. I had three monitors, by the way, So I had labs running on one. Doing, um, I had. Doing market research. I had. O3 is kind of more like an executive thought partner that I was going back and forth with, but then also pulling up deep research reports off that. And then I was using Gemini to build, like, actually build code in their canvas. So, like, I was going all the way from, like, I basically had, you know, an analyst, like a product manager and an engineer running all at the same time. And, um, I'm not a product person. I've never had any experience doing product. I've never coded before. I mean, I've done no code tools and I've spent way too many hours in LLMs. Um, but then at the same time doing the research analysts. And then I was going back and forth and they're all operating at the same time. And I like, I was like, holy shit. I just did like weeks of work in like an hour, right. Of like everything they connect me to. And then what I even started to do is I'm like, all right, maybe I can take this one step further. This was later in the day. I'm like, all right, I gotta try this. So I used the same prompt, okay, across. For research on a product idea across Manus, Genspark Labs, Google Deep Research and OpenAI Deep Research. Right?

Speaker B: Okay.

Speaker A: Got all those reports and then had, uh, OpenAI's O3 model analyze the reports like they were a leader. And those are all the analysts that provided those individual reports and come up with a, like, synthesis of like, recommendation and plan. So that was my AGI moment I literally had yesterday was like, holy shit, this is going to change everything. And that was just manual. There's ways obviously do it automated too. But once I saw that, it just like, blew my mind. And I'm like, this is. This is crazy. So if you had anything like that or something similar where you're just like, like, whoa. Like, just really just blown away.

Speaker B: Yeah, like, particularly like that. That is fantastic, by the way. Like, given your background and you being able to superhumanly power through that, that is empowerment. Like, that's effectively democrat. Democratization of AI, right? Basically giving you superhuman abilities. For me particularly, I would say my format of AGI would be like, I've been trying to create a digital avatar of myself that actually can go into meetings.

Speaker A: Nice.

Speaker B: Similar to what the, uh, Zoom CTO did particularly.

Speaker A: Oh, I didn't hear about that. I didn't hear about that.

Speaker B: Internally into some of his company calls their avatar is so accurate because the Zoom CTR apparently doesn't talk much out Publicly, he's like, a silent listener, but he's like. When he speaks, he speaks very intently. So he's like, yeah, to just stay silent most of the times. So mine, when I'm on my internal company calls, I'm, like, all the time talking, and, like, they have to raise their hands multiple times to shut down, not ask me to shut up. So it was very hard for me to switch my Persona, uh, and suddenly become silent. So this is one agent that I've been trying to build, and I've been experimenting with a bunch of dolls. So white coding is definitely one of those options. But I haven't really experienced the, uh, near AGI level principle until I saw what Google did with VO3. That was actually blown away. And the reason I'll tell you why it is blown away is particularly what I was trying to do is most of my calls when I'm actually, I'm outside, not at my desk, is on my phone. I. And I switch on the camera, and I actually see me moving and the world around it. With VO3, what, uh, I'm experimenting right now is can I actually show a dynamically changing world around my. Me as a background and me blending in around that? Like the sunlight and the waving hair. Ah. When the wind comes in. So I'm trying a bunch of those. And I believe it's getting much and much better. I would say it's the reality. I'll tell you why I'm saying this. I showed it to my wife the other day, one of this clippings that I created of my altar. I'll see if I can pull this up. And she's like, when did you actually go? Because I've been traveling, doing a bunch of travels. I was in Paris, I was in Zurich. So she knows I visited the Paris office and the other Salesforce office, which is right. Right opposite to the Eiffel Tower. But she didn't know that, uh, actually didn't go for an evening cruise, which, because I didn't. But it was an agent with an avatar, uh, that has a sign river in the background, and me taking the cruise in the evening and me talking. So I just captured that clipping. Looks so surreal with that lighting and the face light up and the camera angles, like, man. I was like, okay, it. It can get better. It can better. I can send. I can send that video to my parents, and they wouldn't even realize.

Speaker A: Like, let me ask you a question. Is this like, you took a picture of yourself and then you're integrating it into Veo is like basically saying like hey create me doing this or something like that.

Speaker B: Yeah, VEO API is one of those. But then I've been playing around with a lot of other video libraries as well that can actually add additional componentization

Speaker A: to hey Jen or something like that or.

Speaker B: Yeah, that's one of them. But uh, there are a couple of these open source ones that uh, uh on Hugging Face which is an extension of that model. So I've been trying to chain them together because Gemini can give you only so much with veil. But then I've been trying to see what else is missing and some of this is also pushed me into the direction of like because I also end up talking to a lot of YC founders going back to my startup roots. A couple of bunch of them are building some really innovative stuff that could actually complete that last mile to reality and I got to try out a few of those. But uh, product which are not yet ga hence I will not drop their product names. But it is really interesting. Some of those experiments do happen post office hours. I would say like one of these hours.

Speaker A: Yeah. No, I mean that's amazing that you're doing that. I love that you're, you're chaining together multiple solutions to try and integrate you into video. That's a super creative use case. I'd be curious like there's some funny things you could do with that right. Like I saw like we were joking around. I shared it with my kids and my wife. There's like we got a French bulldog and so there's one on the French Bulldog Olympics and it shows all these like French bulldogs doing and they look so natural like jumping off the diving board like of how a French bulldog would act. So it's like it's even getting like the physics and the nuances down of like the individuals. So it's getting freaking wild. It's going to be weird.

Speaker B: I think there's so much potential of like even companies like Google have to think about what is the next scope of search engine optimization that the ad revenues even like one of them uh in back at home in India just using the three tweets created a video that went live and they were able to generate a few hundred dollars from that video campaign and they didn't have to have to press a button to do that auto generated from those three video feeds based on whatever is the political situation happening. And you could imagine something like this would actually be challenging to a lot of people doing those day in day jobs. But then the way I'm looking at it is it's actually an empowerment to augment and give them more power to do other things as well.

Speaker A: Yeah, very true, man. Good considerations. All right, well, I'll ask you one last question, then we'll bounce. I know we're way over time, man. So what would you say is the wildest prediction you have about AI that most people call you crazy for?

Speaker B: Wow, this is a tough one. M the first time GPT launched and this is like a little backstory to this. The first time GPT launched, I knew we are going to go exactly to that moment of me watching that movie her. Like, I knew that's coming. That's coming in the next 24, 36 months. Like knew that is coming. Because for me, the transformer models were probably something that I started reading about transformer models back in 2018 because I spent a lot of time building predictive systems back in the day. And when Transformers started coming out, I was like, oh, uh, yeah, maybe this is an extension. People are just going to do multiple other experiments similar to what they did with NFTs for all, you know, on blockchain and things like that. Oh, uh, this is probably just going to be a fad at some stage. What I think is AI accessibility to AI is going to be so cheap that it will almost be zero value. Like it's almost going to be. You don't pay anything to breathe air today, like it's free, but then you have different use cases that you could be put in a different situation with. So I think the extension, the access to AI will be almost zero cost, but where people will charge money or people will start visiting businesses, the deployment of AI. How do you actually deploy this to make it something personal? And I feel that's the directions value of tokens going 0,000. Yesterday, today you saw just today, this afternoon, one of the models that LLAMA announced, uh, was 0.$10 for a million set of tokens. It's like, when is it going to come down to zero? Just zero? Like when are they going to just say zero?

Speaker A: Here's what I think. I think right now it's a race. So I don't know if that's going to continue because it's like BC subsidized right now and eventually people are going

Speaker B: to want to get paid.

Speaker A: So I think they're trying to get people hooked on the platform or whatever they're using and then they're going to start escalating prices. Like I think it's gonna follow like so, and we'll use, uh, let's use LinkedIn as an example, right? Like, I'm a LinkedIn user. I remember two years ago where you would do a post and if it would get, let's just say 30 likes and 10 comments, you would get, I don't know, 30,000 views out of that. Right? Or 40,000 views on average. Or maybe not that high. Maybe it was like 10,000 or whatever. Yeah, uh, you, um, do that now and. Because they're optimizing more for like, like, all right, we got enough users now. They're more optimizing for revenues. They're. They're basically showing more ads and they're giving less organic reach. So I think the same thing is going to happen. Like, everybody's like, yeah, the price is going to go down to zero. But like, these guys got to get paid sometime, right? And their investors got to make some money. So that's, that's the. I think there's going to be a reckoning, right. Once people are really hooked and they got all your data and they got everything, it's going to be a hard switching cost. And I think that's when shit's gonna get kind of weird, right? So I could be wrong. I could be wrong. But.

Speaker B: Yeah, but I'm also curious about, uh, what, what are we going to do about alternate realities as well? Like, uh, if you remember the movie Zero, Ready, Player one and that becoming.

Speaker A: Yeah, ready, Player one.

Speaker B: Yeah, ready, Player one. While Zuck was building the virtual worlds, I was for a while thinking that he's possibly going in that direction. But now with all these virtual alternatives coming up and the, uh, ability for you to naturally mimic the real world, I think there could be virtual worlds built out in different contexts, like.

Speaker A: We'll see.

Speaker B: That'll be an interesting one.

Speaker A: I think that's what Fang Fein Lee is building, is like kind of that virtual world model view or whatever.

Speaker B: Yeah, yeah. One of, uh, favorite, uh, so. Oh, yeah, because you mentioned favoritely, her spouse actually works for Salesforce. He's Sylvia. He's a chief scientist of Salesforce.

Speaker A: Okay. Oh, chief scientist. Really? Oh, yeah, I remember hearing about that. I didn't make that connection so well. Hey, man, we're way over. I don't want to keep you any longer. This has been amazing. So much fun to nerd out with you on, on these different topics. Uh, where can people find you? Where can they find more about what you're doing? And, uh, we'll wrap things up, man.

Speaker B: Yeah, you can find me on LinkedIn. And, um, I'm pretty much act very active on LinkedIn, much active than I'm on X and other social platforms. And, uh, if you're in London or in the Bay Area, you'll also find me actively speaking in various community forums. I. I love to come into the builder section, where people are building things from scratch. I'm also very actively and mentoring, advising some of those founders as well. Yeah, definitely a part of the community. So if at some stage you need something to be just check. Idea check. Hey, I need some pointers on this. Always open to aj.

Speaker A: It's been a pleasure, man. Appreciate you being open. Um, and thanks for being on the show, man. This is so much fun.

Speaker B: Thank you.

Speaker A: All right, and we will see you all on the next episode.