Video Transcript

– [Pini Yakuel] So hi, everybody. I’m so excited to be here. And I’m going to talk today about AI, which is essentially all the rage these days, and everybody’s talking about it. So hopefully I can give you some…an interesting insight on, you know, what’s AI, the relationship between AI and CRM and campaigns and kind of like how this all meshes together. It’s a very thought-provoking session. So I hope you enjoy it, and I hope you learn something from it. So the title is “Bridging the Gap Between Art and Science.” So let’s get started.

All right, so a little bit about myself, so my name is Pini Yakuel. I’m 39 years old. I’m Israeli, originated in Israel, from Tel Aviv. Today I live on the Upper West Side in New York. I moved there 13 months ago. I have a dog, a wife, one baby and one on the way. And essentially this past year has been very exciting in terms of kind of like learning about the American culture.

And I’m the CEO and founder of Optimove, which is the company that, you know, I started out of university. It’s my first job. And I was kind of like growing the company, and the company was growing me. I’m always in a position that I’m not qualified to do what I’m doing, even at this point.

All right. So Optimove in general, Optimove is a very unique story in the Israeli tech scene in the sense that it’s essentially a bootstrap company, meaning that we built the company in the old-fashioned way. We didn’t raise money. We grew the company through our clients. It started 2009. The first three years was we were selling, you know, the Optimove model and the Optimove vision through services. And in 2012, we launched the brand Optimove, which is today trusted by 250 brands. We have… It’s a global company with offices in Tel Aviv, London, and New York. The headquarters is in New York today.

And what we do is we specialize in helping brands grow through their existing customers. So we get started from the point of visit of an acquired customer or registered customer. We start from that point. And from that point, we foster loyalty in relationship with the customer by mainly leveraging data science and machine learning. So I’m going to explain today in great detail about how we’re unique and different and how this all relates to AI and marketing and retention between the two. Yeah, we work with businesses like Deezer, Happy Socks, Freshly, Stitch Fix, many strong brands both in e-commerce and gaming, apps, and financial services.

So as they kind of like…I like to think about us as philosophers and thinkers. And throughout the years, we’ve been, you know, kind of like meddling with this problem. It started from my days at the university. I was doing my degree in stochastic modeling and, you know, machine learning. And I started a company with a PhD. And we’ve always had this dream of kind of like connecting the data science part with the art part. And we’ve always kind of like been thinking about how is it possible to… It’s kind of like, you know, the two sides of the human brain, right, the left side and the right side, when one side is responsible for creativity, and the other side is responsible for analytics and critical thinking.

How do we make the power of data science more accessible to CRM practitioners?

And the same thing happens essentially within Optimove. And connecting the two has always been kind of like an overarching problem that kept me at least awake at night for at least five years. But if we kind of like break it down… Yeah, so if we break it down, it comes to how do we make the power of data science more accessible to CRM practitioners? And if we break it down even further, it essentially becomes two questions.

So CRM practitioners, you know, their job is pretty difficult because today they have to leverage and understand data science and be, you know, knowledgeable in terms of data. But on the other hand, the art portion is still very strong. So they have a lot of tools, a lot of things to master very quickly.

Utilize technologies and automation to become CRM masters

So the first thing is how can marketers better utilize technologies and automation to become CRM masters. We’ve noticed that many of our customers, the difference between kind of like the gurus and the masters is very big. So on one hand, we want to help them using data science to be a lot stronger. And on the other hand, we want to help them overcome the price of generalization or the cost of generalization.

And essentially what that means is that the cost of generalization is the reason why we don’t do a good job as marketers. But this is a phenomena that’s far greater than ourselves. Essentially, if you think…the best way to think about it is to think about what happens in medicine.

So a personal story. I told you guys that I have a baby on the way cooking. So we were doing all the tests, and everything was going great. And all of a sudden, like the worst 24 hours of my life, we got one test, and they say that…a blood test…and they say that some hormone is elevated, right? So what does this mean? What could it mean? And they say that statistically it could mean many bad things

So we had the worst 48 hours of…you know, of our lives. We’re very… You know, there’s a lot of tension and things like that. And then we do another test. And the other test says, “Oh, guess what, everything is fine.” You know, that’s great news. But the reason it happens is because of the price of generalization. They don’t really know. So you are… Once that one of the tests is slightly elevated, it puts you in a bucket that could mean many, many things, but it’s all statistical.

Overcome the cost of generalization in CRM campaigns

So the same thing happens with your CRM campaigns. When you’re sending a campaign to an aggregate, to kind of like a broad group, that campaign on an aggregate level could be very successful. But, essentially, when you break it down, there’s probably many microsegments within that campaign that actually don’t like the message. It doesn’t resonate with them. It even annoys them. But it’s all being masked behind the…you know, the overall result of the aggregate.

So this is the cost of generalization. And we’re trying to overcome these two things. This is the problem. How do we make people, you know, use technology, use automation to become faster and stronger And how do we…in terms of, you know, scalable. And then how do we reduce the cost of generalization using science? How do we understand our customers a lot better in going micro.

So the fallacy… And I want to tell you about the evolution of our thinking because this problem has been kind of like unfolding for many years. And the result of that problem, I’ll give you a little hint, is our bot. So eventually we have something called Optibot, which is how we, you know, overcome all of these things. And I want to talk to you about the approach that we took to solve this problem.

Why doesn’t complete automation work?

So at the beginning when we got started, being, you know, data scientists, we started from a very automatic utopic vision, which is like kind of like this machine. It has a lot of microsegments on one side and all the possible marketing actions on the other side. And that machine constantly iterates and tests and does a lot of, you know, exploration versus exploitation. And we do a lot of testing. And then we find the best marketing action for every customer every time for every micro segment. By the way, it has to do to our name, Optimove, the best thing you can do for every customer.

And the reason I call this the fallacy is because this couldn’t work, okay? This couldn’t work, although it sounds very utopic because it means complete automation, right? Everything is tested. We just feed this machine with new actions, and that machine would always find the best action for every microsegment. And you’re not going to pay the cost of generalization because the microsegments are small, right? So they’re very relevant. They’re very precise. They’re very specific.

And essentially the reason this couldn’t work is because there’s a disconnect, okay? So why this couldn’t work, because the world that we live in as Optimove and generally speaking is in the nexus of data science and the art of marketing. So we are right there in the middle. And this is why we can understand and I can talk today about this disconnect.

How does Optimove help start a conversation with your customer?

Think about these two words…worlds, sorry, for a second. So the world of data science, it’s a world of precision. You know everything. There’s no…it’s a black-and-white world. Everything is about correlations, causations. How do I kind of like find predictors that will help me to improve my business? How do I summarize and organize the data in an insightful manner so I can get new insight? This world is very rigid. It’s clean. And it abides to very, you know, known rules.

Whereas, the world of the art of marketing, how can you tell if a marketer is good or not, right? If you know people and you say, “She’s a great marketer. He’s not a great marketer,” what’s the…how do you know, right? It’s a world of art. It doesn’t obey to, you know, specific rules. It’s a world of psychology, sociology. It’s a world of copy, of design.

And over there, the thing that we realize and the reason there’s a disconnect, when we are running CRM campaigns, essentially what we’re doing, we’re starting a conversation with our customers. And that’s the biggest thing. So if you think about it, if you guys are doing CRM campaigns, every campaign that you do, think about it for a second, you’re actually starting a conversation with your customer. And you can start a lame conversation that will not resonate with your customers, and they’re not going to open the email or look at your message because the conversation will be, you know, bland[SP]. And maybe it’s not even a good time to start a conversation, or you don’t have good enough of an excuse to start a conversation. Or it could be very creative.

And what we’re looking at Optimove all the time is going to the data scientist and finding some things that will spark ideas for conversation starters. But our research shows that a ton of the conversation starters are actually happening within kind of like the human creativity realm and are not yet fully translated into the data science realm. That’s why there’s a disconnect. That’s why the fallacy…

– How are we able to make this connection? In order to model the disconnect, we identified three types of campaigns. And the three types of campaigns are in terms of how do they kind of like explain the relationship between the data science part and the art part. So we call them kind of like… Bear with me. These are three interesting names. We call them mute predictors, verbal predictors, and mute conversation starters.

The relationship between data science and art: three types of campaigns

So in one extreme, we have predictors. We call them mute predictors. Why mute predictors? Because they do not help me to start a conversation with my customer. I find something super cool in the data. There’s a crazy correlation or an interesting causation. I find a unique segment, but that revelation does not help me to start a conversation. That’s why we call it mute predictors.

On the other side, on the complete opposite, we have mute conversation starters. Something is happening in politics. The season’s changed. The rain is coming. I’m launching a new brand. I’m doing something within my business that has nothing to do with the database of the predictors that I find there. And this is called mute conversation starters.

And in the middle, these are the ones that we would like to produce more and more of. These are verbal predictors. So these are predictors that not only that they reveal something very interesting in the data, but also what they do is that they help me to spark an idea for a conversation, to start a conversation with my customers.

Campaign 1: Mute predictors

All right, so a few examples. Here’s an example of returns. And it’s a mute predictor. So we know from our research that essentially if customers are returning items, not people that return older items, but if they have mid return ratio or high return ratio, we know that they have a very high future value. But what am I going to do with it? I’m not going to send an email saying, “Hey, I noticed that you returned some items, and I know now that you’re a great customer.” It’s not a conversation starter. It doesn’t help me to start a conversation.

It does help me, however, to build operationally. I do understand that I need to build a very seamless return process because this is something that resonates well with my customers. I need to allow them to enjoy that. And I see it in the data, but it doesn’t resonate back to CRM.

Another example would be, in terms of kind of like the diversity or variety of people, so if I have customers that are buying both women’s and men’s items, these are very high-value customers. But again, it doesn’t help me to start a conversation.

In the same way, another mute predictor… So it’s about the first purchase. So if I… We always see it in the data. The first purchase, if there’s a lot of items from many different departments in the first purchase, the likelihood of a second purchase is extremely higher. But again… So maybe I can…product-wise, I can build the process of a first purchase to push customers towards buying another item or another category. But again, I’m not going to start a conversation around that.

Campaign 2: Verbal predictors

Examples around verbal predictors is… As an example, here, I find a cluster of VIPs who are at very high risk of churning. I see people that buy for a lot of money. They visit many, many times. This is from one of our clients, Happy Socks. And I can see that they didn’t visit for a while. And the model tells me that churn risk is extremely high. So it sparks an idea. I’m going to do a special presale just for these customers with a very aggressive discount.

I have a reason to start a conversation because these are VIPs that I haven’t seen for a long, long time, and I’m going to show them that I want them back and I kind of like value their business. So this does help me to start a conversation.

Another example, it’s around the show “Westworld.” So essentially… It’s pretty ironic, but this is an email that I received. So HBO probably discovered that if I complete to watch a complete series, right, if I watch the whole series, my future value is higher. I’m not going to churn. So what they do when they see in the data that I’m not…I reach Episode 4, and I’m not watching anymore, they send me an email to watch…to come back and watch the show. And I didn’t because I find it a bit scary, which is a bit ironic, given that my topic today is AI. But this is a true story. So again, I’m using the predictor to actually start a conversation.

Campaign 3: Mute conversation starters

And last but not the least is the most popular types of predictors and conversation starters is when essentially, as an example, you know, the fall is coming or the summer is coming. So this is Adore Me, another one of our clients. So they’re going to send an email with bathing suits or worn pajamas depending on the season.

Another example would be… This is a Happy Socks Local Hero campaign. So they came up with a very interesting grand marketing campaign. They chose kind of like 15 local heroes in 15 countries, and they’re designing their own limited edition. So I’m just going to use the data to find the target group. But the reason to start a conversation came from marketing, essentially from the fact that they came up with this creative campaign.

Same goes for nelly.com, another one of our clients. They launched a new brand. I have a new brand. It’s a big news for me. It’s a reason to start a conversation with my customers. And maybe I’m going to help them to tailor the segments better. And Deezer, a streaming service, so if Pharrell Williams is launching a new album, it’s a reason to start a conversation. It didn’t come from the data. It didn’t stem from the data. It came from the fact that Pharrell Williams have a…just is launching a new album. And essentially, I can use the data to find a better target group like hip hop lovers or people who like to explore new types of music. If they never listen to Pharrell, maybe now is the time to send them this.

But essentially, again, this data was not in our databases. So the Optimove predictive model could have not mined it. We may be able to do it in two years’ time when we grab data from kind of like semantic data. And it has to do a lot with semantic data and types of analysis that human beings are a lot better than machines. And the last but not the least, obviously, the king, so a very unfortunate conversation starter, but still, when Leonard Cohen passes away, so they blast their entire base because it’s a big artist.

So to kind of like to go really fast and explain, so this led us to choose the marketer and understand that we still need the human being to help us with designing an experiment. So designing a very smart experiment is still something that the marketer could do a lot better than the AI. But after the marketer does that, we’re going to use the AI to make the science and automation work for the art. So we’re going to put the marketer in the center, and it’s kind of like a combined approach. Once you design an experiment, we’re going to use our predictive models and our bot to help you get the best results out of that.

Optibot: looking at the customer from many perspectives

So kind of like to set the stage really quickly… So essentially the Optimove model is kind of like a model that goes all the way into microsegments. So I talked a lot about microsegments. It starts from stages, layers, then microsegments. And then what it does, it goes…kind of like looks at the customer from many perspectives. We have the microsegments, and then we look and see how customers move and migrate between those segments. And this is the basis for our predictive model, our predictive engine, and the basis for all of our journeys that are based on these microsegments.

But the whole thing, this whole evolution of the problem and our thinking process led us to develop Optibot, which is the first marketing optimization bot and essentially does two main things. So the first thing is we help the marketer become a CRM master by surfacing one-click actionable insights. So there’s a lot of insights that the marketer just needs to approve. It has to with methodology, with data science, with different types of segments and new ideas for campaigns that the marketer just seas. And it’s kind of like a concierge that helps the marketer become a lot better.

And the second… So this covers the automation part. The second part is the part of the science. So we have this really exciting thing called the self-optimized campaign. And this is truly grand. Essentially what it does is after the marketer designs an experiment, what we’re able to do is, you know, slice up the big aggregate. So if you’re running a campaign for all of your active customers, but you have ABCD, you have a few options, then we start to show and mix and match the best possible action not on the aggregate level but on a microsegment level.

And this creates amazing uplifts because we just…we’re not paying that price of generalization. And the reason we have the U.S. map here is to explain… Imagine if we could have had Hillary as the president of California or New York. Of course, that would have been problematic in the democratic system. But essentially, we are paying the price of generalization. We’re choosing one winner for the entire aggregate. Whereas, we could’ve gone down to the state level or even a district level. And if we go micro, we don’t have to choose just one winner. We can’t do it in politics, but we can definitely do it in CRM.

So, guys, that’s it. It was very fast. It went really quickly. So this was artificial intelligence in the service of marketing. The way we went through this problem is that we try to make it, you know, machine-only. We saw and understood the deep, you know, forces that work underneath and make kind of like the marketer and the artist still very much important, and we’ve chosen to make a combined approach, using the science and automation to help the marketer focus on the art.

– Now, Pini, the all-important question, how do people get in touch with you. We know where you live. We know about your family. But if we actually want to, like, talk to you, how do we do that?

– All right, so very easy. On our website, there’s an info email. I get it myself. And if you want, you can email me or just reach out to the company, and all the people can, you know, get in touch with me directly.