 Hi, this is George Gilbert. We have an extra special guest today on our CubeCast. Aman Neymar, Senior Vice President and CTO of DemandBase, started with a five-person startup, Spider Book, and almost like a reverse IPO. DemandBase bought Spider Book, but it sounds like Spider Book took over DemandBase. So Aman, welcome. Thank you. Excited to be here. Almost good to see you. So DemandBase is a next-gen CRM program. Let's talk about, just to set some context, for those who aren't intimately familiar with traditional CRM, what problems do they solve, and how did they start, and how did they evolve? Right, that's a really good question. So for the audience, I mean, CRM really started as a contact manager. And it was replicating what a sales person did in their own private notebook, writing contact phone numbers in an electronic version of it. So you had products that were really built for salespeople on an individual basis. But it slowly evolved, particularly with Siebel, into more of a different twist. It evolved into more of a management tool, a reporting tool, because Tom Siebel was himself a sales manager, ran a sales team at Oracle. And so it actually turned from an individual-focused product to an organization management reporting product. And I've been building this stuff since I was 19. And so it's interesting that with the products today, we're going, actually, pivoting back into products that help salespeople or help individual marketers and add value, and not just focus on management reporting. That's an interesting perspective. So it's more now empowering, as opposed to reporting. Right. And I think some of it is cultural influence of the, over the last decade, we have seen consumer apps actually take a much more predominant position, rather than traditional earlier in the 80s and 90s, the advanced applications were corporate applications, your large computers and companies. But over the last year, consumer technology is much has taken off. And actually, I would argue, has advanced more than even enterprise technology. So in essence, that's influencing the business. So even ERP was a system of record, which is the state of the enterprise. And this is much more an organizational productivity tool. Right. OK. So tell us now the mental leap, the conceptual leap that demand-based made in terms of trying to solve a different problem. Right. So demand-based started on the premise around marketing automation and marketing application, which was around identifying who you are. As we move towards more digital transaction and web, web was becoming the predominant way of doing business. As people say, that's 70%, 80% of all businesses start using online digital research. There was no way to know it, right? Majority of the internet is this dark, unknown place. You don't know who's on your website. You're referring to the anonymity and not knowing who is interacting with you until very late. And you can't do anything intelligent if you don't know somebody, right? If you didn't know me, you couldn't really ask what would you do. You'll ask me stupid questions around the weather, right? And really, when we, as humans, can only communicate if you know somebody. So the innovation behind demand-based was, and it still continues to be, to actually bring around and identify who you're talking to, be it online, on your website, and now off even your website. And that allows you to have a much more sort of personalized conversation. Because ultimately, in marketing, and perhaps even in sales, it comes down to having a personal conversation. So that's really what, you know, which if you could have a billion people who could talk to every person coming to your website in a personalized manner, that would be fantastic. But that's just not possible. So how do you identify a person before they even get to a vendor's website so that you can start, you know, on a personalized level? Right. So demand-based has been building this for a long time. But really, it's a hard problem. And it's harder now than ever before, because of security and privacy. You know, lots of hackers out there, people are actually trying to hide or at least prevent this from leaking out. So in the, you know, eight, nine years ago, it was, you know, we could buy registries or reverse DNS, but now with ISPs, and, you know, we are behind probably Comcast or level three. So how do you even know who this IP address is registered to? So we would, you know, in the, about eight years ago, we started mapping IP addresses, because that's how you browse the internet to companies that they work at, right? But it turned out like that was no longer effective. So we have built over the last eight year proprietary methods that, you know, knows how companies, you know, relate to the IP addresses that they have. But we have gone to doing partnerships. So when you, you know, log into some websites, we partner with them to identify you, if you self-identify at Forbes.com, for example. So when you log in, we do a deal and we have hundreds of partners and data providers. But now, the state of the art where we are, is we are now looking at behavioral signals to identify who you are. In other words, not just touch points with partners where they can, where they collect an identity. You have a signature of behavior. It's really interesting that humans are very unique. And based on what they're reading online and what they're reading about, you can actually identify a person. And certainly identify enough things about them to know that this is an executive at Tesla who is interested in IoT manufacturing. So you don't need to resolve down to the name level. No. You need to know sort of the profile, the persona. And that's enough for marketing, right? So if I knew that this is a C-level supply chain executive from Tesla, you know, who lives in Palo Alto and has interest in these areas or problems, that's enough for Siemens to then have an intelligent conversation to this person, even if they're anonymous on their website or if they call on the phone or anything else. So, OK, tell us the next step. Once you have a persona, is it demand-based that helps and put together a personalized profile and lead it through the conversation? Yeah. So earlier, well, not earlier, but very recently, we've been building this technology, which is a very hard problem to identify now hundreds of millions of people, I think, around 700 business people globally, which is majority of the business world. But we realize that in AI, making recommendations or giving you data in advanced analytics is just not good enough. Because you need a way to actually take action and have a personalized conversation because there are 100,000 people in your website making recommendations. It's just overwhelming for humans to get that much data. So the better sort of idea now that we're working on is just take the action. So if somebody from Tesla visits your website and they are an executable by your product, take them to the right application. If they go back and leave your website, then display them the right message in a personalized ad. So it's all about taking actions and then, obviously, whenever possible, guiding humans towards a personalized conversation that will maximize your relationship. So it sounds like sometimes it's anticipating and recommending a next best action. And sometimes it's your program taking the next best action. Because it's just not possible to scale people to take actions. I mean, we have 30, 40 sales reps in demand-based. We can't handle the volume. And it's difficult to create that personalized letter. So we make recommendations, but we found that it's just too overwhelming. So in other words, when you're talking about recommendations, you're talking about recommendations for demand-based for... Or our clients' employees or salespeople. But whenever possible, we're looking to now build systems that, in essence, are in autopilot mode. And they take the action. They drive themselves. Give us some examples of the actions. That's right. Some actions could be without... If you know that a qualified person came to your website, is notify the salesperson and open a chat window, saying, this is an executive. This is similar to a person who will buy a product from you. They're looking for this thing. Do you want to connect with a salesperson? And obviously, only the people that will buy from you. Or the action could be send them an email automatically based on something that it will be interested in. And in essence, have a conversation. So it's all about conversation and an ad. Or an email or a person are just ways of having a conversation in different channels. So there was a... It sounds like there was an intermediate marketing automation generation. After traditional CRM, which was reporting, where it was basically... It didn't work until you registered on the website. And then they could email you. They could call you. Inside sales reps, if you took a demo, you had to put an ID in there. That's still... So when demand-based came around, that was the predominant between the CRM we were talking about. There was a gap, there was a generation which started B2B marketing. It was all about formfills. And it was all about nurturing. But I think that's just spam. And today, their effectiveness is close to nothing. Because it's basically email or outbound calls. Yeah, it's email, spam. We all have email boxes filled with this stuff. Yeah. And why doesn't it work? Because not only because it's becoming ineffective, and that's one reason. Because they don't know me. Right. And it boils down to if the email was really good and it related to you, it related to what you're looking for, or who you are, then it will be effective. But spam or generic email is just not effective. So to some extent, we lost the intimacy. And with the new generation of what we call account-based marketing, we're trying to build intimacy at scale. Okay, so tell us first the philosophy behind account-based marketing, and then the mechanics of how you do it. Sure, yeah. I mean, really, account-based marketing is nothing new. So if you walk into a corporation, they have these really sophisticated salespeople who understand their clients, and they focus on one-on-one, and it's very effective. So if you had Google as a client, or Tesla as a client in your Siemens, you would have two people working and keeping that relationship working because you make millions of dollars. But that's not a scalable model. It's certainly not scalable for startups here to work with or to scale your organization to be more effective. So really the idea behind account-based marketing is to scale that same efficacy, the same personalized conversation, but at higher volume, right? And take, maximize, and the only way to really do that is using artificial intelligence. Because in essence, we're trying to replicate human behavior, human knowledge at scale, right? And to be able to harvest and know what somebody who knows about pharma would know. So tell, give me an example of, let's stay in pharma for a sec. And what are the decision points where based on what a customer does or responds to, you determine the next step or demand-based determines what next step to take. What are some of those options? Like a decision tree, maybe? You can think of it, it's quite fetish in our industry now, it's reinforcement learning, which is what Google used in the Go system, AlphaGo, right? And we were inspired by that. And in essence, what we're trying to do is predict not only what will keep you going, but where you will win. So we give rewards at each point and the ultimate goal is to convert you to a customer. So it looks at all your possible futures and then it figures out in what possible futures you will be a customer. And then it works backwards to figure out where should you take you next. Wow, okay, so this is very different from the- You can plan six months ahead. So it's a planning system. Okay. Because your sales cycle's a six months ahead. So help us understand the difference between the traditional statistical machine learning that is a little more mainstream now. Then the deep learning and the neural nets and then reinforcement learning, like where are the sweet spots? What are the sweet spots for the problems they solve? Yeah, I mean, there's a lot of fat and things out there. In my opinion, you can achieve a lot and solve real world problems with simpler machine learning algorithms. In fact, for the data science team that I run, I always say start with the most simplest algorithm. Because if the data is there and there's you have the intuition, you can get to a 60% F score or quality with the most naive implementation. 60% meaning you- Like accuracy of the model. Like confidence model, okay. But how good the model is, how precise it is. Okay. And sure, then you can make it better by using more advanced algorithms. The reinforcement learning, the interesting thing is that it's ability to plan ahead. Most machine learning can only make a decision, they're classifiers of sorts, right? They say, is this good or bad or is this blue or is this a cat or not, right? They're mostly Boolean in nature or you can simulate that and have multi-class classifiers. But reinforcement learning allows you to sort of plan ahead. And in CRM or as humans, we're always planning ahead. A really good salesperson knows that for this stage opportunity or for this person in pharma, I need to invite them to the dinner because their friends are coming and they know that last year when they did that, then in the future, that person converted, right? If they go to the next stage, so it plans ahead the possible futures and figures out what to do next. So for those who are familiar with the term AB testing and who are familiar with the notion that most machine learning models have to be trained on data where the answer, the answer exists and then they test it out on, well, they train it on one set of data where they know the answers, then they hold some back. And test it and see if it works. So how does reinforcement learning change that? I mean, it's still testing on supervised models to know it can be used to derive. You still need data to understand what the reward function would be, right? And you still need to have historical data to understand what you should give it and you sure have humans influence it as well, right? At some point, we always need data, right? If you don't have the data, you're nowhere, right? And if you don't have, but it also turns out that most of the times there is a way to either derive the data from some unsupervised method or have a proxy for the data that you really need. So pick a key feature in demand base and then where you can derive the data you need to make a decision, just as an example. Yeah, that's a really good question, right? We derive data as all the time, right? So let me do something quite interesting that I wish more companies and people used is the internet data, right? The internet today is the largest source of human knowledge and it actually knows more than you could imagine. And even simple queries, so we use the Bing API a lot, right? And to know, so one of the simple problems we ran into many years ago, and that's when we realized how we should be using internet data, which in academia has been used, but not as, and it's not as used as it should be. So you can buy APIs from Bing and I wish Google would give their API, but they don't, so that's the next best choice. We wanted to understand who people are. So there's their common names, right? So George Gilbert is a common name or Alan Fletcher who was my co-founder and is that a common name? And if you search that, just that name, if you get that name in various contexts or co-occurring with other words, you can see that there are many Alan Fletchers, right? Or if you get versus if you type my name, my name, you will always find the same kind of context. So you'll know it's one person or it's a unique name. So it sounds to me that like reinforcement learning is online learning where you're using context, it's not perfectly labeled data. Right. And I think there is no perfectly labeled data. So there's a misunderstanding of data scientists coming out of perfectly labeled data courses from Stanford or whatever machine learning program. And we realized very quickly that the world doesn't have any perfect labeled data. Then we think we're going to crowdsource that data. And it turns out we've tried it multiple times and after a year, I realized that it's just a waste of time. You can't get 20 cents or 25 cents per item, workers somewhere and wherever to label data of any quality to you. So it's much more effective to, and we were a startup so we didn't have money like Google to pay. And even if you have the money, it generally never works out. We find it more effective to bootstrap or use unsupervised models to actually create data. Help us elaborate on that. Yeah. The unsupervised and the bootstrapping where maybe it's sort of like a lawnmower where you give it that first, that's right. You know, tug. I mean, we have used it extensively. So let me give you an example. Let's say you wanted to create a list of cities, right? Or a list of the classic example, actually it was a paper written by Sergey Brin. I think he was trying to figure out the names of all authors in the world and this is 1998, right? And basically if you search on Google, the term has written the book, right? Right, just the term has written a book. These are called patterns or Hearst patterns, I think. Then you can imagine that it's always preceded what a name of a person was an author. And so George Gilbert has written the book and then the name of the book, right? Or William Shakespeare has written the book X. And you'll see it with William Shakespeare and you get some books or you put Shakespeare and you get some authors, right? And then you use it to learn other patterns that also co-occur between William Shakespeare and the book, you know, and then you learn more patterns and you use it to extract more authors. And in the case of demand-based, that's how you go from learning, starting bootstrapping within, say, pharma terminology and learning the rest of pharma. And then using genetic terminology to enter an industry and then learning terminology that we ourselves don't understand. Yet it means, like, you know, for example, I always use this example where, you know, if we read a sentence like, Takeda has inlicensed a molecule from, you know, Roche, it may mean nothing to us, but it means that they've partnered and bought a product, right, in pharma lingo. So we use it to learn new language and it's a common technique. We use it extensively, but so, you know, it goes down to, while we do use, you know, highly sophisticated algorithms for some problems, I think most problems can be solved with simple models and, you know, thinking through how to apply domain expertise and data intuition and having the data to do it. Okay, let's pause on that point and come back to it because that sounds like a rich vein to explore. So this is George Gilbert on the ground at Demand Base. We'll be right back in a few minutes.