E54: Agentic AI for Leaders: Sell Outcomes, Not Hype | Warren Schilpzand, Regional Director SambaNova Systems

So the first thing is that whatever you're doing, I believe this for any sales role, not just in the cutting-edge ones. It makes a massive difference if you're passionate about what it is you're representing. If you love it, your energy will be transferred onto those that you are talking to and they will, you know, value what you're talking about too, because largely at the end of the day, people buy from other people who are valuable Welcome to the High Performance Sales Leader Podcast.

My guest today is Warren Schiltzand. I've known Warren for a number of years and I've admired him in terms of his career trajectory, as he always tends to ride the technology transformation wave. Over the last few years, he's worked with some really interesting AI companies, And he has recently started as the regional director at SambaNova Systems. And I'm going to hand this over to him to actually explain what it does because it is absolutely at the cutting edge of AI infrastructure. So Warren, thank you so much for joining me.

Hi, Preet. Great to be here.

Fantastic. Now tell us for all our listeners, what does SambaNova do? Give me a little bit of background, and I think it's absolutely fascinating, but Over to you.

Danvanova is tackling the challenges within AI of how do you deliver with energy efficiency without a heavy reliance on water cooling and being able to do it at a price point that allows very, very high-scale AI adoption inside of enterprises, but also inside of other organizations as well.

Fantastic. And you were talking about how this came, know, with the founders and Stanford. So tell me a little bit more about that.

Yeah, so our founders are from Stanford University in the US, so Rodrigo Liang and Kunle Olukotan and Chris Wray. And Kunle is actually very famous for being the godfather of multicore processing out of Stanford. But one of the things that they wanted to do was to take a look at this AI space and say, well, what are the first principles for designing a processor to deliver And that if you start with that, you can then start to get to some very high energy efficiency capabilities with very high throughputs and being able to drive a lower price point per token per watt.

Okay, excellent. So let's unpack that a little bit more. There's a lot of talk about LLMs and training, but you're talking about inference and how that is impacting the systems that are built out there. So can you Tell us a little bit more about that.

So often when we think about AI, actually, for most folks, when they think about AI, think about using AI. So you're thinking about ChatGPT or something like that where you've, you're having a kind of a cool conversation with it. The other part of it is the training. There isn't a lot of very public-facing, not information, but public-facing stories about this. But this, You know, consumes huge amounts of electricity, huge amounts of water, and so forth, because this is a really big task. But when we train models, the output is the model. So we run this for a period of time. The main audience for this are things like, people like data scientists and so forth who wanna create a model for a particular task. But once that model is created, the thing that we need to do is to actually use it in our applications. It becomes an application development problem. Rather than a data science problem, because we have the model now. Now we need to go use it. And actually, the amount of throughput that's required here and some of the things that we need on that side are very different from what we need on the training side. So on training, we're okay for this process to take a long time. We've got huge datasets, we're doing— running lots of iterations and so forth. But on the inference side, it has to be fast. If a user can't use it or get an answer very quickly, they'll abandon using whatever system is utilizing it. We are going out to millions of users, not a small group of folks who are very specialized in training models. And so this inference space is actually the thing that most of us are going to interact with a lot, and we're gonna consume millions upon millions upon millions of tokens out of these systems. And when we do it this way, when we are using it, this is what we call the inference side, utilizing those models and delivering capabilities into applications, into user experiences, And doing that on a very large scale, ideally millions and millions of times over.

Okay. So help me understand, am I understanding this right, that inference is the speed of being able to interpret what the user is requiring? Or what is inference?

Inference is the execution of this model now, right? So if the model is very good at understanding the input and being able to generate an output, we wanna do that quickly. And we're using the model to do that. We spent and consumed a lot of resources over here to train the model to be very good at it. Now we actually want it to go do it. And this is the inference side. Go do the thing you're good at. So you can think about it like this is actually running the model like we would run an application. So if you think about training like the application development side, so we spent all this time to build something. Now we wanna use it on this side, and that's the inference side, is on the usage side.

Understood. And I can just imagine when we want everything in a matter of seconds or even yesterday, if you're looking at the prompt bar or any kind of application that, you know, I wonder how many seconds before people abort or lose interest and wanna go and do something else.

Yes. Yes. And, you know, for certain interactions, we need even higher speeds. So, for example, if you're interacting with something with voice, where you're expecting a conversation, the actual response time to you, the science says it needs to be about 200 milliseconds or less. Otherwise, you lose this illusion of the conversation. And so, you know, we do start to get very interested then in how fast we can run these things so we can keep the user engaged. Otherwise, they'll abandon it. Wow.

That is fantastic. So, To me, it sounds like you are at that early stage of the AI rush or the gold rush with the essential tools that people will be using to do this. So here you are in Asia, obviously will be leading customers on a transformational journey, I would imagine. So tell me some of the things that you are encountering when you talk to customers about maybe some of the You know, the wrong ideas or, you know, what are the things that you have to first get right with customers that they don't know about? Where does that start?

Yeah. So it's really interesting because you get a very broad spectrum. So you have some folks who are actually very advanced in what they've been doing. And they've actually been able to go and build out some really cool things. And then you have others that are at the beginning of their journey and still trying to figure out what is the use case. There are a couple of things that come into it. One is actually not a technology thing. It has to do with the business case behind what you want to do with AI. So the really smart organizations have figured out that what we actually need to do first is sit there and go, okay, how is this actually going to affect the top line or bottom line? And that's the goal that they've been working to. Others have said, we just need to do something in AI, and they've gone out to do something. They do it and then go, oh, we didn't get the benefit that we didn't really think about we should get, if that makes sense.

Yeah.

So, there's sometimes there's a little bit of that disillusionment. Oh, we tried this AI thing, it didn't really work. We're going back to the drawing board. We're going to stare at our navel for a little bit and figure out the use case now. But like anything that we do with technology, it has to start with how is this actually going to impact the business? And then you can start to design how your AI implementation and what you're gonna do with AI is going to drive to that aim. So that's probably the first one, is getting past those organizations that have had a little bit of disillusion, a bit of a false start, and starting to move forward. The second one really falls into the notion of, oh, for us to do something in AI, we really need to train or fine-tune a model. And in some cases, that's true. That'll help quite a bit. But the path that we're seeing a lot of organizations getting the most value from is Really where they're doing a lot of context engineering. And especially as you get into agentic, it becomes agentic context engineering where we're thinking really about the data that we're providing to the LLM for it to do what it needs to do. And can we actually encapsulate the learnings of how to use the LLM in the context rather than trying to retrain it? Remember before we said the training consumes a lot of energy, a lot of time. If we do it this way, we don't really need to go through quite that same process, and we can adapt a lot quicker than we can if we try and do a fine-tuning, because each time we need to go through a change, we need to do that fine-tuning exercise again. Whereas with context engineering, we don't. We're actually tuning the output of our AI engines through the context that we're providing.

Okay. So help me understand this through a use case, a typical or a common use case, what you've just said to me, you know, context engineering with Agentic.

So in Agentic, this is really the architecture of how we bring AI together. So the idea with Agentic, there's 2 parts to it. One is that the agent itself is able to decide the process. So we're going to say to it, you have these tools available, we want you to behave like this. And, but you essentially, we're gonna give you some guidance on how the process should work, but you're gonna make the decision about what you need to do, how many times you need to do it and so forth. So that's one part. And then we can group agents together into larger agentic systems. So agentic systems that may have hundreds of agents within it so that they become really super useful. Now, that's that part of the architecture. What we do then in this context engineering, as we talk to the— as we give these agents things to work on, is we are saying, here is a history of information, maybe about the customer or the user. Here is a history of information that we have about how this process has worked best in the past. Utilize this contextual information to make your decisions and do what you need to do. And so through that, it can learn. And so as an example, earlier this year, I was working with a bank on a simple use case, which was around expenses checking. And in this case, rather than training a language model, what we had was a 2-agent system where one agent was really designed to read the policy documents from finance. Not from IT, from finance. So reading the finance policy documents and knowing how to interpret that. So that's our context. This is the, the finance document is our context here. And so to change it, we only, if we want to change the way this thing behaves, we only need to change the finance document. And then we have the actual expenses coming through and the coordinating agent working through that to determine what it needs to do and is able to output that. So the thing that's really interesting to me about that is in the past, we would've had the finance document, it goes to IT, someone will either enter it into a rules engine or code it into a system, somehow codify that policy, and then we'll test it and it'll go out. In this case, you're going to take the finance document as the way to operate. So now I cut out a whole bunch of steps, and I can have finance now directly actually modifying those documents and rolling them out as per necessary. And obviously with things like, you know, the version check and, you know, are we okay with how it works and so forth, not those things notwithstanding. But that's something that's really hard to do with a traditional system. But it's quite easy to do with the— well, easy in a relative term with an agentic system.

So an agentic system is able to make its own decisions based on the parameters and context that it's been set.

Yeah.

So give me a glimpse of the future. And let's assume that you've got a sales leader or sales rep who likes a $1,000 Singapore chili crab, right? And let's say you've got issues with expenses. Somebody has a tendency to basically go for very expensive dinners. What would a system like that do? It's a very simplistic question, but I just want to make it really real for the audience.

Well, perhaps I can tweak your example a little bit because in the right scenario, it should do the thing that an expenses system does today, which would be, hey, that's out of compliance, and we're going to flag it, and it's going to get flagged somewhere, right? The more interesting thing would be to say, well, he's written a description about why it was a $1,000 chili prep, and the agentic system is able to read that and determine from that that actually a customer was present and there was 7 people there. And, you know, the circumstances called for it. And so therefore, actually, it does fit compliance because from a finance, the finance document says if you've got customers and this many and, you know, however many per head, however many dollars per head, Maybe it is compliant. So that's where it gets really quite interesting in terms of the way that this would process, because in a, in a straight rules-based system, you don't really have the opportunity to do that. It'll get spat out to a person who will then take it, review it, probably get into an argument with the sales guy, and someone will eventually either go, yeah, it's okay, let it go, or, or no, we've— this is the last time, right? But that's where I, and AI is really good at those things at, at the, figuring out and understanding what someone is doing with intent and then being able to apply that to things that are somewhat deterministic. So it's actually a really cool use case, Alan. I don't know if it's ever, if it's gonna displace all the expensive systems that are out there, but it's an interesting one.

That's amazing because, you know, I think you've made that, you know, the context really come alive in that example because it really depends who was at that lunch and what's the benefit of that lunch. Because if that lunch is going to lead to progressing a million-dollar deal, well, it's just a $1,000 lunch.

Yeah.

So, okay, that's really good. Tell me a little bit more about use cases and the rise of agentic. How do you see agentic playing out and the use cases being impacted? What are you seeing from the frontier of this technology adoption?

Yes. So agentic is where things are heading in terms of the way that we'll see AI largely implemented everywhere. And for the simple reason that it is very, very smart at what it does in terms of especially being able to reconfigure itself with how a process should orient based on what it's seeing in the context. What I think is quite interesting is we are going to start to interact with organizations and applications and services where we go, oh, that was really very intuitive, or that was a great experience, where we are not really going to be sitting there going, oh, that AI was really, really great at what it did for me. And I think this is something that a lot of organizations will need to get their head around a little bit in terms of how they do that. So, you know, some of the use cases that start to pop up are things like, you know, an application process, like a mortgage application process. There's a lot of information that needs to be collected. Some of it is somewhat subjective, you know, where you have to provide descriptions or, you know, examples of your spending or other things that need to go in and it needs to be assessed. And we are used to filling out those forms and we then spend some time on the phone with somebody who's gonna process it and then they're gonna go off and process it. And at some point, Later, it's going to come back. That could be hours, it could be days. It's going to come back, might be more questions and so forth. So how do we trim and shorten that process such that all the information that we need to be able to process it properly is done up front? Now, that may not be through a chat interface. It may not be typing into like a ChatGPT-style thing. I want to apply for a mortgage. Okay, well, let me now ask you the 100 questions. It might still be a form. But as we're filling things in, we're making calls out to AI, whether that's agentic or directly to an LLM, and making some assessments on the fly and saying, hey, that description that you put in here, we need some more information. You know, can you describe this? Can you describe that? And so forth. So that we're collecting all that up in the one session so that then when it does go to be processed, and it could be processed by an agent, agentic system, But even if it's being processed by people in the background, as much of that is already done upfront, we don't have to keep going back to the customer. And so, you know, saves, you know, from a mortgage provider's perspective, saves them time and effort, allows them to get an answer back quicker. There are some amazing statistics about how often a customer will go with the first respondent rather than necessarily the best respondent.

Yes.

So, you know, being able to get answers back very, very quickly is very important in that kind of scenario. We think about government use cases. There are so many things, you know, every time we talk about government use cases, everybody goes, oh yeah, working with the government's always difficult. But, you know, we've made great strides on doing some of those things, you know, like what some of the things Service NSW has done has been amazing, and I know that's being adopted throughout the country. But there are many, many other things that seem to go round and round in circles and people can't get— AI with agentic systems should be able to help streamline a lot of that process, whether it's through forms or just collecting the right information or helping people actually determine what they are after. Oftentimes they don't really know how to express it properly. So using agentic systems to be able to try and get to a level of understanding to match people with services from government as well. Some really key use cases. But we'll see AI turn up just about everywhere, just because simply this, the ability to have agents understand the process and do that in a way that's not as rigid as we have done previously in applications. So Previously, we would code it or use something that could flow us through a particular process. Now, if I can call an API and say, hey, here's the information, I just need an answer back to continue on in whatever my app needs to do, those things are going to really cut down the amount of time it takes to get out new features, new products, new capabilities, and deliver more intuitive user experiences.

So based on what you said to me, the real gains are potentially twofold. Am I understanding this right? Number one, at the human experience level, which is that, like you said, the first to respond or in a voice conversation, having a better interaction because as I can already see with The millennials and younger, if they don't have to speak to someone on a phone and they can do it with a chatbot or they can do it with a text, they're going to want to do that and they're going to want that right now. So that's one. But also, there's potentially some cost savings that could also be achieved in terms of removing inefficiencies beneath the surface. What are some of the greatest, I guess, advantages that companies are going to start seeing in and through using more and more agentic-based processes?

Largely, I think the ones that are gonna be most visible are the user experience ones. So where I can have a website where my website is able to interact with somebody like a sales assistant would, you know, it will drive folks to less cart abandonment, more Getting to a point where they go, yeah, that's the actual things that I'm after and clicking that buy button and moving through that. Because today, if you think about the online experience, largely it's, you've kind of gotta know what you're looking for rather than, but if I go into a shop, I don't really do that necessarily. Sometimes you do, but the sales assistant, oh, what are you looking for? Ah, you know, I'm thinking about this or I'm thinking about that and, oh, let me show you this or let me show you that. If we can start to bring some of that to our digital channels, that helps to increase top line. Remember before I said with the business case, we've got to figure out how it affects top line or how it actually affects bottom line. And so that's the first one is how do we, using these systems to really affect top line outcomes. So, you know, great use cases for things that anything in retail is going to be great for that. And as I said, you know, from a government perspective, it's similar, but with not a commercial Outcome on the back end. And then on the other side, which is more about how you affect the bottom line, I'd be really careful about using the productivity word, right? Because productivity is great and we're always striving for better productivity. The challenge with productivity in a business case is it's very hard to show the line item on the P&L that you have actually affected. Unless you go and fire the people or cut the contractors or do this or do that. And, you know, we just saw this week with CBA that that can be challenging, right? If I'm cutting people, that can be really challenging. But there are other things, you know, like again, using retail as an example, how do I cut down RMAs or returns, right? So if I can help someone choose the right thing upfront that they're happy with and they're satisfied with, they're less likely to return it. So, you know, that is an actual impact. If I can start to cut down returns, I actually have a financial impact that I can point to as a line item. That's just an example. So, but yes, and the kinds of things that we're gonna be able to build, the kinds of products that people can deliver will also be different because of the flexibility of agents. They're much more flexible in terms of the way they can work than traditional applications. And so that will drive Different products, different experiences.

So would that mean like with an Expedia, you know, would it be able to ask me questions and then tell me where I should be going rather than me just looking at Italy?

Well, it depends if you love Italy, then that's maybe where you're going to go.

Right.

And there are some great examples of that that have already popped up with travel agent type agents where yes. And yeah, this is a great one where actually a chat style interface is actually Probably preferable to a form-based one because we're used to having a conversation with a travel agent or with somebody about, you know, hey, I'm thinking about this or about that. Can you show me these options or show me those options? And so on and so forth. There's actually an example that I've used for a few years. A number of years ago, I was going to the US to take the kids to Disneyland and went and booked the airline tickets first. And at that time got back. a set of hotel recommendations, none of which were near Disneyland because your landing airport is LAX, right? So you get a lot of business hotels and other hotels in the LA area. So didn't think too much about it. Went off to another website, found a great hotel, booked a hotel. 2 days later, 3 days later, the airline sends an email. Ah, preparation for your trip. This is really nice. Here are some hotel recommendations. And now on this set of hotel recommendations is the one that we had booked through a different site. So, you know, as the airline, you may think, oh, this is a great experience I'm providing, but actually it's a negative experience because then it sent my wife into, well, I want that rate and now I'm gonna have to, but I don't wanna cancel it unless I know I can get that rate. And I don't wanna, I don't wanna book this one unless I know I can cancel that one and so on and so forth, right? So the extra work was actually very high. But you could easily have used AI in that scenario to say, what have we got? We've got 2 adults and 2 kids, right? Flying to LA. What are the chances? You know, if I use AI, I can probably, it'll go, yeah, probably going either to Universal or to Disneyland. And I can ask that question, are you going to any of these places? Yeah, I think I'm gonna go here. What kind of hotel would you like? And now I could go through that. So, and would've probably booked through them rather than through the other site. So yeah, that travel is a top use case because you're matching needs with properties and with locations and airlines. And it's more emotional than it is logical in a lot of ways, right? Because you're, what is it that you like? It's not what is it that makes the most logical? If we did that, we'd all stay at an Ibis, right? Sorry.

Yeah.

So the intuitiveness of AI to be able to interact with you in that way makes that a really great use case to capture more of the traveler's wallet, if you like. It's a bit of a harsh way of saying it, but—

So as we wrap up, this has been fascinating and I could ask you question upon question. How do we as professionals you know, in leading-edge technology companies who ourselves need a transformation and an OS upgrade in terms of how we're thinking, how we're talking to customers, how we are preparing ourselves. What would be 3 to 5 tips you'd give me if you were mentoring me in terms of being prepared for the world that is coming in 2026 or in the years ahead? How do I get ready?

So the first thing is that whatever you're doing, I believe this is for any sales role. Not just in the cutting-edge ones. It makes a massive difference if you're passionate about what it is you're representing. If you love it, your energy will be transferred onto those that you are talking to and they will, you know, value what you're talking about too. Because largely at the end of the day, people buy from other people who are valuable to them. And so if you're just turning up, if you think, oh, I'm going to get on this AI thing because it's the next wave, you know, it's the next technology thing, but you have zero passion for it, you're probably still not going to do terribly well at it.

Yeah.

So the first thing is developing a bit of a passion for the technology. And then the second one, which is very close to it, is a passion for how organizations are going to utilize whatever it is. So this is not about the tech itself. This is more about business, right? Your customers, what do they do? Are you talking to an airline? How do you become, you know, how do you embody that? How do you go and talk to them like you're an airline executive and then apply what you're bringing to it? The third one is always about having a sense of needing to move on to the next thing. And this is something personally I've always tried to do is learn something, teach somebody else how to do it. So that I can move on to the next interesting thing because this space is changing. Like if you have a look at the AI space and whatever's next, there'll be something after AI. AI is not the end. Whatever the next thing is, it'll be next. And you need to be able to have the capacity to move to that. So it's always about not ever getting to a point where you're saying, well, I know this and this is what I'm comfortable with. If you want to keep staying at the edge, you've got to keep moving. And that sometimes means relinquishing things that you've worked on or done in order for you to be able to have the headspace and the capacity to move on to the next thing. But it, again, it's about having that absolute desire and passion to understand it and move forward.

Thank you so much, Warren. I literally have so many more questions, but in the interest of time, I'm just going to say you really have expanded, you know, really, I think my takeaway was there's so much more that is coming down the road. And there's— if it really is improving customer experiences and outcomes, I think that rather than fearing what can come out of AI, I think there's so much to be excited about as much as people who are engaging with customers and the journeys that they're going with. So I just want to say a huge thank you. Thank you so much.

Thank you, Priya. It's been, been great having a chat about AI. Love it.