Understanding Agentic AI with Rita Castillo

[MUSIC] Welcome everybody to another episode of the podcast. I'm your host, Bobby Grille. And today we're going to get deep into the world of a gentick AI, what it is, how it works here at ServiceNow. And we have a wonderful expert joining us, and that is Rita Castillo, who is vice president of AI Design here at ServiceNow. Rita, thank you so much for joining us today. Of course, thank you so much for having me. Walk us through your background in the world of AI, and just how this world is changing so much and your role in it. Yeah, I'd say I have a very lucky job because I get to focus on AI. And it happens to be a passion of mine, something that I've always been interested in. Even from a little kid, I was always like that kid that liked sci-fi. So for me, this is kind of like where I've been very happy to have my career progressed to. I've done in, you know, the tech sector for a long time from consumer to biotech, and then to enterprise with ServiceNow. And at ServiceNow, I really got to dig into AI because ServiceNow, like early on, started with AI. And so I've been able to progress a journey of early, you know, scripted machine learning all the way up to the gentick now. And for me, I was like, this is something I want to do. And so I slowly built my team up at ServiceNow until now, this is all I do is I just drive a design team for AI and I love it. I mean, you're one of the few people that has this background of knowledge in this. And I want to, before we get deep into this, kind of bring us up to speed as much as you can, because I know there's a lot to cover. But from this idea of, you know, science fiction to science fact now, how much do we need to know about AI going forward? Learn as much as you can, but don't feel overwhelmed. You don't need to become a data scientist. We have amazing data scientists, especially here at ServiceNow. But what you want to know is that it's going to be an everyday part of your life. And it is something that you should integrate into your life and learn about. It's not something to shy away from or think I'll just do it later. It definitely is everywhere. And it's at the gas station pump. I mean, it's all over the place. Yeah, it's crazy how much we use it now in that short time that AI has actually been anything. I mean, you know, we joke about it and, oh, and sky net and all these things. But now it's like, no, it's just kind of part of everything. It really is fortunately solving a lot of our problems, right? Yeah, I definitely say it. It's solving a lot of problems. It's helping us find new ways to do things. It's taking care of some of the stuff that we just really don't want to do all the time over and over. So it's really advancing automation, which is something we've already embraced. Right? Automation is a good thing. Automation is repetitive processes that you don't have to do. And that you can expect to be done for you. AI is really leaning on top of that and helping us with these things that we already trust. But it's also taking it to the next level and handling more complex things for us. Well, you mentioned automation. And that's the big thing. I think a lot of people forget that it's not a new idea. It's just now it's something else making that automation work like you said. Give us the introduction to a gentick AI and how that really flows into this foundation of making AI successful. So I think you really have to have a strong foundation if you're going to be a company that's going to provide AI to people. You can't just turn it on and be like, here's some AI. You really have to have something for it to build on. And so you really have to start with something strong. Like, well, serve is now is really good at workflows. Right? We're really good at data connection and data across silos. And like I say, we really nailed automation early on. And so we have scripted workflows and deterministic workflows, which means that they happen when you, like exactly the way you say they should. Right? Things that are repeatable, things that are automated. But, you know, we've moved into predictive workflows. You know, and that means that now we can kind of think about like, oh, based on patterns of automation, what should we do next? How do we pre-fill forms? And this is kind of more classical AI people call and machine learning predictive intelligence. Those are kind of products that we've had for a while. And that has been really helpful because that like progresses workflows, like we're saying progresses automation. And then we started doing conversational workflows, which is when generative AI and LLMs came in board. And it became a huge consumer thing, right, with chat GPT. People started understanding, hey, I can do things with natural language. And I can talk to machines and it serves now. That's really what you're doing is you're using natural language to do these workflows that we're really good at. And we can already do. But you're not having to go to the drudgery of a form or a list or anything like that. And then that next magic moment was a gen tick. Really, a gen tick is like, oh, not only can we handle deterministic, well-known experience workflows, but now we can handle complex situations and variability and do that more autonomously on behalf of humans. Well, let me, let me give an example and you tell me if it's right. Because I know a lot of people like myself are still trying to figure this out. You know, as much as we think we know about this stuff and forms and workflows and things, it's like, there's that gap. So if I've got a scripted workflow, I've built all of this into a process either service now or in my life. For some platform, you're going to fill out all these forms. The predictive is you put in your address, we know it's in California, we know the zip code. We've that information's there, right? Yep. Okay, oh, you're good. I'm halfway there. Conversational is, hey, where does this guy live based off of some information that I kind of just kind of spit out, right? Yeah, that's a change in the UI. You're not looking at a form, you're not filling it in. It's asking you and just doing it in behind the scenes. Okay, and then a gen tick does in that really weak example, like, what is a gen tick then do? So a gen tick would understand if you are, let's say you're in a different location or has context of who you are or that you've moved recently or that you have two addresses in the system and you're likely to be at this one instead of that one because you're on vacation. And so a gen tick will be able to switch those variations. So if you're like, oh no, I'm not, because the conversational one might say, oh, do you want me to send this to you at your home address? And you might say, oh gosh, no, I'm visiting someone else. I needed to send to this address. It would not be able to really handle that very well. You'd have to say, like, please enter this, you know, all the things you need. A gen tick would be able to look and be like, oh, are you at this location? Right? Because it knows more about you and it can switch the, switch the conversation to, to pick up those changes. And if you're like, yeah, and I want it two days instead of one day and maybe if that wasn't put into the prescripted one that you could change a shipping speed, right? But the gen tick one would be able to check into the system and check policies and see if they can do that for you. Okay, that's completely insane and amazing at the same time. That, I mean, I just kind of, I was thinking about an example. I can't think of anything that's smart. But when you explain it, that's really amazing. I mean, hands down because I mean, all that information is out there. It does take time. You've got your calendar sync. You've got all these things. It sounds so, you know, what's the scientific term for duh? You have to do it. Oh, wait. But this is, this is breakthrough. It really is. And the thing is, all those things I just described are something that if you were talking to a human or a live agent, they would have done for you. But again, it takes time. They have to find it. They might have to get approvals. And you're taking a human's time, right? When this is something you could self-service. And so that's the beauty of it is that, well, you know, that's a very important workflow for you. It's something that can be handled with the existing information. And can be handled immediately for you. You don't have to wait for anybody to help you. So that's that idea of autonomy then. Yep. So how does that now come into this world? How are we looking at autonomy and agente AI? Agente AI really is just, it's an artificial intelligence system that can act proactively and autonomously. But it's really acting towards a defined goal. And it's really acting on behalf of humans still. So while it can make decisions on its own, it can act, it can look at the context of the situation. It really is still bounded by ethical guidelines and principles that humans put in place. Because there are things that we decide are okay for a machine to do for us. And so but there are thresholds, right? We still set thresholds for security and things like that. So while it has autonomy to do these things and make decisions and send out tasks and ask for approvals, it's to the speed of the ability for it to do that and to predict what it should be doing and to be asking and planning that workflow is really the autonomy where it comes into play. How do we get to this level of autonomy? And I know that's that's terrible. Like how do we do it makes no sense. But what is the steps or the thinking behind to get to autonomy? So early on we were thinking about when we were designing it. So when we're designing it, we're like we need to make sure that while it's an autonomous system, like I said, it still works within our guardrails and things like that. And early on we were like, okay, can we draw a line between read and write activities? Right? So if it's a read activity, it can be autonomous. We don't, I doesn't have to ask can I go read this policy? Can I summarize the policy? It's not creating anything, it's not generating anything, it's not changing anything without permissions. But the more and more we let the systems do these things, where we as humans will trust that they're doing the right thing. And so we'll give them more autonomy over time. We'll say, "Okay, you can do these things and you can do these things." And I'm okay with this. You've repeated it enough time. As in your success threshold is a certain state that we are comfortable with giving these lower level tasks over to machines, but we always build in safeguards. Right? So one of the principles is that we can't, we don't want to allow anything that's not reversible. Nothing that a human can't review or a human can't edit. You know, critical high-risk situations are not things that we would push to autonomy. Okay, that makes a lot of sense. That kind of is that we've heard a lot that toddler analogy when it comes to AI. It's like, well, you can do all of this, but let's make sure you're not breaking anything and not hurting something. And where it's going to be pretty good. You had mentioned, and that's one thing I want to talk about with workflows is deterministic. And again, for a lot of people who are kind of learning this, that seems like a new concept or it's a hard concept to understand. Explain that idea of deterministic workflows as it comes into agentic. So if you think about deterministic workflows, it means that we can determine what the outcome will be. Exactly how it'll be performed, what's used to do it, what tools are used in the process. Then exactly what the outcome should be. They're like password reset, ordering a computer. Very, very well-known, very same, same, same, every time. We've done this. We know it's going to happen. Yeah, you're cool. You can handle it. Yes, exactly. We trust it. It fits into our regulated environments. This is exactly what we want to happen. Okay, okay. That's because that's the one that I know when I talk to people about it and people are asking me and we're all talking about AI and trying to make our own AI. Because I know that's a bigger part of this or not make our own AI, but use AI to solve our problems. It's that interesting concept as I need to think all the way through this. However, there's a learning ability of agentic and AI to a point that it's going to, I got to do it once, but then it's going to figure out how to do it better, faster, correct? Sometimes yes. Okay. And so the thing is that it, well, it might seem to like pull it out of thin air, right? It's not the best way for it to work. It needs to work and be trained on things like you said, like it needs to watch someone work. It needs to look at how things happen in a system. It needs to study a system where the last 90 days of how humans have done it or how it's been performed by the automation in the system. So it can learn what kind of things should it be doing, what kind of variations could exist? Where might it need to make these other decisions and action points? And so that's really the autonomy and the learning come in by looking at things we've already done. So that comes into this idea of more agentic. Again, you're the expert on this. So I've only done, I've only, like most people kind of scratch the surface of it. You have the assistance we're using on the workflows. And then we have this now autonomy and these agents were all starting to see. And we're going to spend a little more time talking about agents, but there's a step after agent, which is a little bit more in depth that it's like a team or something. Yeah, so there is our the analysis system that we talked to, right? That's our interface layer, it's our engagement layer with AI. But behind the analysis interface is an orchestrator and a team of AI agents that are working. And so there is this hierarchy kind of where it's like, hey, now assist, I need this thing or help me with this incident or help me replace my computer. And so it knows based on intent and intelligence that this is something that's going to require a little bit more complex task or it's going to require a deterministic flow and it can help decide. But when it says conflict situation, so orchestrator actually creates a plan. This is agent to kicking into gear here. We're then it like actually picks the right agents to put together this, team basically for you so that they'll go off and do what they're good at. So the agents are made to do very specific tasks, to do specific tools, they use workflows, they use scripts, they talk to each other outside the system. But the orchestrator at ServiceNow is really kind of wrangling them, picking the right ones with a right job and then putting them in the order where we'll get the data sequenced back so that we can give the human back the right answer. Okay, and I was you're explaining that that you know, I think all of us who started to use AI more and more, it's like, oh, I have a colleague who's doing something similar to me and we've got two different workflows and two different paths, which are probably 80% the same except I do it left, they do it right or they like green, I like light green and there's something a little different that it's it's going to be you have to determine right, which is the right one to use. And so there's another layer to do that. Great, that's correct. Yeah, yeah, it is quite and it's looking at all those different things and there is a bit of complexity and personalization that is kind of future looking and things that we're looking at too that where it knows who you are and how you think and how the ways that you work and so it will start to it can reflect back the ways that you like to work or your team likes to work as well. So there's a lot in how we instruct that orchestrator in order to make sure that things come out the way we'd like them. Okay, okay, so that so let's kind of go through the the hierarchy as it were. We've got our tools at service now every company has one but we're going to talk specifically service now. We've got all these tools and the skill sets and the workflows and all of this this data that's the what the who is the let me get this right the AI agent correct. Yes. Okay, and so the agent is we've broken that down to HR and item and CRM and those things. Correct, that's the next one is who. Yeah, and the thing is that the way that we're building these here at service now is that we create out of the box agents that are good at these different industries right HR and CSM and technology workflows and then our business units also create ones that are even better specifically for their verticals too. And so they know these agents already know a lot about what's going on with that customer was going with that industry how to use our tools and our workflows. And so yes, those agents are prebuilt for those specific use cases. Okay, and then the next level is the orchestrator. Correct. Yes. So coordinating those and picking the right ones is the orchestrator. So if you think like a pyramid it's like the orchestrators at the top. Yeah, so pick the agents and then the agents pick the tools of what they're going to use to do the job. Okay. Okay. That that that is actually very very clear now because it's again, I you know, I know I'm dumping this down but I'm the dumb person in this conversation because some of this stuff it's like you know with anything AI it's like well it's doing it for you. How? Why? Where's the next thing going to go that's so much of us are yeah, we we trust it and and at service now I know from experience and talking with many experts like you there's a huge amount of of of protection and trust that goes into all of this. But that brings up to the next thing of this this you know evolution of where agentic is going now that we've started this great evolution and almost revolution where are we moving toward now with agentic? Oh that's a great question because there's so many places we can go and I think for us we are building more and more into ensuring that we build that trust with our customers. And so especially Sarah's now we have a huge desire to provide ethical and responsible AI regardless of its agentic or regulatory AI but for us to do that it means that we have to really understand and secure what we're putting out and make sure that we abide by our policies our customers policies but still make sure that they work you know really really hard and really well so the quality and the value is there but we want to be that trusted company that we already are we want to continue building that trust. So that means we work hard on security we work really hard on reliability you work really hard on accuracy and we have a flexible system so that we work with whatever the customers LLMs needs are and we work with data inside and outside of service now but we are glad to say that it comes into our one single platform and comes through our secure walls and things like that so we can ensure that trust. And that makes sense because I'm going to pull another example here which I know you've talked about in other talks this idea of you know we hear we use the example I spilled coffee on my laptop a lot which is I hear it so many times and I go man I spilled coffee in my laptop the other day I am now part of the problem but not the solution but you know bigger things that you mentioned trust where somebody goes well I need to change my bank account for direct deposit and that's a everybody no matter when they have done this is gone your fingers just go my fingers won't let me type I can't do this I'm afraid to touch my important information but that's changed correct yeah I think that that's our goal is that you don't you're not afraid that you are comfortable with it and that you can trust that the system is working in a beneficial way on your behalf and that you're not worried where is this data going. It's my personal information who's going to have this if I type it in here or if I say it with voice, how is that going to, what's it going to do? What's the system going to do with it? And so for us, that kind of thing is is hard because it's also an evolution of control. So if you think about early on, virtual agents and chatbots, if you were to use a customer banking one or you would go to a customer website, you would type it in. Like you said, your fingers would be doing the work and you're typing it in, you're checking your card, you're looking at it. You know, you're like, "Did I, how many zeros are there?" I don't know. Is it bank? Is it my savings? Is it my check? Is it my other checking? I'm trying to get some. So that was there, but that's still very manual and it's like a lot of work and there's, you know, and say you were doing something with your bank account or you want to change something like your direct deposit, you have a lot of work to do even with a chatbot helping you with the form. And then there was an evolution where we gave an answer, right? And so that was like our first gen.2BI answer was like, "Oh, we can tell you how to do these things." So you still do it yourself, but you get a little bit better information of like where to go, what link to use, what you're going to need to get in before you start, you know, like so really strong, helpful answers. And then they can walk you through it conversationally. And that alone's a revolution. Yeah. Yeah. You're actually helping me do something scary. Yes, exactly. And I feel a little bit more confident because I'm like, "Oh, okay, I know what's going to happen. I know what I need. I feel confident in the answer. It's telling me I know the source it's coming from. I can see it. But it's still you have to do it. Or you have to spend a long time talking to it. And there might still come up to places where you don't, it doesn't know what to do, right? So like that's that, you know, adaptability and flexibility where agentic comes in. And so agentic systems then would say, "Oh, like this isn't going to be easy for you. Not only do I know how to do it, and I know you, but I can do it for you, right? But the thing is that we build it in a way that we're not a black box. It's not like, okay, I'll talk to you later. I'm going to go do it. We tell you what we're going to do. And it's going to say, I'm going to do this, this, this, and this. Are you okay with that? Oh, wow. Right. And so the human is still in control. The human is still saying, yes, or they could say, yes, but don't do this or yes, but I moved. My address is different, right? And so and the agentic will be able to do like a switch and be like, oh, not only do I need to help them with the direct deposit, I also need to help them with the address change. And that's where agentic can do that. I can take two different intents and do all of them and stitch it together with that orchestrator. And then the beauty is that once it gets through helping with all that, because it knows you and knows these systems, it can actually proactively suggest something else. Like, do you have a different phone number two, or do you want to add anything else? So these are the things that agentic can do that that we're hoping will be in a way that helps you feel comfortable working with the system and trust it because of the value you get back. I love how you put this into a personal solution that so many of us have, but this extrapolates much bigger to business because you're you're talking about a workflow that service now is helping a customer's end customer. Yes. Right. Okay. Good. Just had to make sure I get the right the right path on that one, the customer's customer. But that's chain, you know, we're also talking bigger things in business transformation and big big business platforms. Just briefly kind of walk us through that aspect of it on that high level world of a platform. Yeah. I mean, you can think about like large scale workflows, like large like a, you know, like management change and large scale, like supply chain things, large large workflows, or even like a large amount of workflows. Like say, it is, you know, password reset being pushed across the whole company, right? Because there were some security breach or something that we need to help, but agentic is something that could help build out that whole thing across them every really quickly. You know, your IT department's not going individually and fixing everything for everyone. But you can also think about like how will this flow across the system? Right. How do I get it across HR and IT and CSM and all the other silos that we have in order to make sure that we have large scale transformation or workflows. The agentic systems work across that. And then imagine you're doing something where you updated all your pay systems, right? And everyone has to reconnect their stuff and work day or something. Right. This is something that agentic and also do is talk to work day by either using our workflow data fabric, which means it can talk to the data in work day or with our AI agent fabric, which means it can talk to work day agents. And so it can do these things on a large scale across employees with us, you know, big companies, but also between businesses in order to make sure that all of it comes back together. When you look at something like this, when you're looking at agentic AI and you're looking at workflow and your experience in the industry from all sides of it, what are you excited about happening as this moves forward? Oh, I think there's just so many things we can unlock. I think every day that's something that I'm excited to think about is what's the next thing that we can do with AI? What is the next thing a gen fix going to do for us and for humans? One of the things I like to think about is when I design something or my team's designing something that needs to be useful, the very first thing is usefulness and purpose. But for me, AI and agentic, it unlocks human potential, right? There's a, but it also accelerates it. That's the thing that I'm excited about is what else are we going to be able to accelerate so that we can do more and we can have more space to create. And whether that's create for, you know, humanity's sake or create for my business transformation, I just think it gives us more space to find those places where we can provide and push value as well. Oh, that's amazing. That makes, I think all of us feel good. Of course. Thank you so much for having me. I'd be happy to talk to you anytime. So there you are, everybody, a good explanation of a genetic AI, both in the world of AI and here at ServiceNow. Rita Castillo has been our guest. She is VP of AI Design here at ServiceNow. I'm your host, Bobby Frill. And if you need to know more information about ServiceNow platform, head to www.ServiceNow.com. Thanks for listening.