Art of Artificial Intelligence in Marketing Optimization
Learn how artificial intelligence is empowering marketers to communicate with customers more effectively.
– [Dan] To be frank a good deal of what my job is these days is, whether it’s writing articles or doing these interviews, it’s getting beyond the marketing and jargon on homepage of Artificial Intelligence companies and figuring out what’s actually happening, what pieces are actually moving, and what business results are actually being driven. A lot of the time shaking those answers out of founders is not an easy job. Luckily, in this interview it was actually relatively simple.
I speak this week with Yohai Sabag who is Chief Data Scientist for Optimove, which is a marketing artificial intelligence and sort of marketing automation company based in Israel, also have offices in New York as well. I speak with Yohai specifically about what humans are needed for in the process of sort of automating and optimizing marketing and what they’re not needed for. What facets can be automated and what facets cannot and I think he spoke pretty frankly with some actual examples as to what elements of the marketing mix namely sort of optimizing individual campaigns to tailor to actual market segments can be automated and which elements namely copywriting and some particular high level decision making about market segments and approaches to those market segments that need to be done by people. And so how do those two continually loop? How do the humans feed the machine and how does the machine feed its information or results to human to continue to make better results?
So leave the world of intuition and be able to optimize in a way that really only machines can. Still definitely a labor heavy process for humans though luckily a lot of the number crunching is done by the technology and I think this interview paints a pretty good picture as to where marketing Artificial Intelligence really is today. Certainly a space where there’s a lot of investment, certainly this is a space where there’s probably a lot of hype. I think this is a pretty realistic view as to how this stuff actually works and how it actually drives business results. What does it really look like when AI is applied and fully utilized in the marketing department today. So without further ado, this is Yohai with Optimove.
– So first and foremost, Yohai, I wanted to ask a lot of the value prop of kind of marketing technologies today. I know a lot of what you guys work on at Optimove has to do with calibrating marketing messages to the particular client, really having kind of that audience of one. I think this is part of the big promise of AI in marketing. It seems so difficult to calibrate the emotional tone, the right product to offer, the right email subject line, all these very nuanced personal elements to a specific person. How have you had to put kind of in layman’s terms how that process shakes out? How that is even possible with machine learning in AI?
– [Yohai] Well, so that indeed challenging. The solution for this challenge it’s a marriage between art and science and that’s very important to note because in automating the marketers should come up with creative and high quality content that we can communicate to the customers and with interesting offers and the science should work for the marketers and find the best match between offers and customers and how we do that? The scientific part actually can be divided into two stages. At the first stage, we want to get to know our customers very well and deep and in the second stage, we want to be able to measure every interaction that we create with them. Every campaign that we send them, we want to be able to measure the effect of it in a very clean manner. That’s the infrastructure to do that and the first part I talked about it’s actually personalization which is a big bother today in marketing.
– Yeah, yeah.
– And here in Optimove, for the first stage, what we do we have a model that slice the customers all the way down into micro segments. We start in high level segmentation to life cycle stages. The most common one are new, active, churn, and then under each life cycle stage we slice the customers in a more specific manner to segmentation layer with specific segmentation layers. For example, like RFM or their product preferences and promotion usage, etc, or even if they active into social media.
– And just to being clear too, Yohai. Just to make sure we’re on the same page, when you say RFM, are you talking about Recency Frequency and Monetary?
– Okay, yeah. Because I just….
– That’s exactly what I’m talking about.
– Part of my job is to call out the jargon stuff for the folks who have tuned in. That’s for everybody who’s listening, RFM or Recency Frequency Monetary serve a way to segment a database of customers by how recently they’ve purchased, how frequently they purchase and what their sort of value is monetarily, how large were those transactions and I guess what Yohai is saying is that that is one of the layers, as you were saying Yohai, that you can put on top of the data base.
– Yeah, that’s precisely true. RFM segmentation is a very strong predictor to predict future value…
– Big time.
– …to predict churn rate and that’s a very common way or method to segment customers into top tier customers, mid-tier, low-tier, etc. And as I said before, we segment the customers via many segmentation layers. We create many point of view on that customers and then the next level of segmentation is the combination of that clusters. What we actually do is we create a customer DNA of a very specific customer persona which is very granular and homogeneous that stand as a learning unit and, for example, we can choose like the following combination of customer persona. We might see active customers that are mid-tier customers, they purchase twice a month, with average order of $50, they had a good shopping experience lastly, and they didn’t return any of the items that they bought on their previous purchase, and they’re not active into social media. That’s might be an example of that kind of customer persona and there are many more. So that’s the first part. The second part is the more scientific part. The one that will allow us to measure every interaction with them and to get the clean uplift of that interaction and here, we have to get into a purely scientific mindset and I mean that in every campaign that we conduct we want to follow all the steps of experimental design and that’s very common in the academic, but not so common in the business world and what steps do we have to follow? First, we have to define a goal for that campaign. For instance, if we now target our active customers, so we have to define what do we want to achieve? Maybe we want to prevent churn, maybe we want to up-sale or to do a cross sale, or if we target churn customers so maybe we want to win them back. So first we have to define a goal. Next, we have to define a leading KPI or a set of KPI to measure that goal and that might be the value or the level of risk of churn, etc. And the last thing and maybe the most important one is to isolate a control group and I said it slowly because it’s very important and although it’s important and very basic, many companies tend to overlook it.
– You know, I was going to say, I think, probably one of the reason it’s overlooked is because it’s awfully difficult to do. You brought up an interesting sort of part at the beginning of the process, Yohai, where you throw a lot of different segmentation layers on top of the databases, it’s obviously, pretty well calibrated depending on the client and you finally serve homogeneous groups that have certain qualities in common whether that be, you know, behavior on social or purchase behavior or whatever the case may be and then you’re mentioning having a control for those groups that we can measure sort of which methods are working. I think part of the…of course, that would be necessary. With your science hat on, you really have to have a control that gets treated in sort of a particular in uniform way so you can measure if what you’re doing is right or not. I guess it’s a lot of…businesses probably find it difficult to kind of set aside and sequester a small group to treat one way and then split up multiple campaigns and treat the other bunch of folks in that cluster with sort of these variable methods of marketing and methods of messaging. I think the act of doing that micro slicing is probably too much of a challenge for a lot of businesses, or you probably see a lot of people kind of get lazy and not take the time to do that very very granular work.
– Definitely, but I don’t expect from the marketers to find the best match for each of the customer personas. Sometimes they have hundreds of them. It can work on a higher level and the machine later on will do that matching job. So even if a marketer like even if they only have like 20 groups of RFM, even then we see it many times, they don’t isolate a control group for each interaction that they create and that’s…it’s very sad actually because there is no way to conduct or to conclude what was the real effect, the clean effect of what we did if you didn’t isolate an appropriate control group. And that’s also very, an extremely important concept because when you isolate the control group so if you have huge numbers, so it might be sufficient just to do it randomly, but when the groups are getting smaller so it’s very important to balance between the group and to isolate a balanced control of it that both group would be comparable. An example that I like to give in that case is a boxing match. Like a heavyweight boxer can’t compete with a lightweight boxer because they’re not comparable. It’s the same here. The top tier customer persona is not comparable to a low tier customer persona and if we have like in the test group the ratio of VIP is significantly higher than the ratio of VIPs in the control group, so the test is not valid and we cannot conclude anything of it.
– Completely agreed. I think, again, the hard work of doing that front end slicing is often skipped. I can speak from experience and speak to my own laziness in this regard in various e-commerce companies and it’s tough to get down to that level, but once you can, of course, you can start to vary things and see how your results shake out. It sounds like, and kind of getting around to the beginning of the question here, it sounds like once you’ve done that level of slicing, once you’ve isolated your control groups within those homogeneous clumps, you know, again top tier customers, low tier customers, frequent purchasers, infrequent purchasers, whatever those groups are that you can isolate controls in, only then can now we start to vary our messaging. Of course, varying our messaging by having people vary the messaging, has been done for quite some time. You know, big companies like LL Bean and these other big database marketers that were doing this stuff before they had computers, you know, know how to slice up a list and rotate offers. So the difference here is the fact that we’re engaging Artificial Intelligence in making some of these calls around what messages should be rotated in front of what people. That’s different than a database marketer scratching his chin and planning it out.
I take it, Yohai, and maybe this is done differently than I’m imagining, the marketers do a lot of crafting of a whole bunch of variable messages or parts of messages in various campaigns and then the job of the AI is to sort of match those right pieces with the right people. Again, the beginning question was how do we calibrate the offer, the emotional tone, the timing of the message, all this stuff to an individual person with AI? It seems like such a hard job. I take it you do your segmentation, you build your marketing pieces, and then there are some way that the Artificial Intelligence is rotating that across people. Talk us through sort of how that process works.
– Okay. So that’s precisely true. Now when we’re combining the very detailed customer persona that stands as learning units, with the ability to measure every interaction, that’s exactly the infrastructure for an optimization process for a learning machine. Now and in Optibot, we have our own bot that’s called Optibot and what it do, it get as a starting point, a target group that might contain dozens of customer personas and a set of actions that might be available for them and then it makes an optimization process that eventually find the best match between the very specific and homogeneous customer personas and the actions that the marketers provide.
– And of course my guess is that this is an ongoing kind of iterative process here where you’re testing various campaigns, you’re figuring out what tends to work with our most active buyers, what tends to work with people who are active on social, what tends to work with people who buy, you know, within this department, for example, you know, the health department, or beauty department, or whatever the business model is.
– Yeah, that’s exactly…
– There must be some kind of looping. I’m interested in how that works as well.
– Yes, that’s exactly that. It’s just an interactive manner. It’s make more and more experiments. It’s repeating experiments and then it’s summarize the conclusions from that experiments. It summarize the results and it’s implement that conclusions on matching between customer personas and actions. That’s exactly what that’s iterative process does. It keep making experiments and summarize their results and matching between customer personas and marketing offers in order to increase the uplift, to increase the effectiveness of what we’re trying to get.
– Yeah, there’s a real kind of man and machine sort of marriage here where the creative creation process, you know, the segmentation and hard thinking on the front end is certainly aided by people in addition to data. That’s obviously very important. The building of the Lego bricks which we could call the pieces of marketing, if you will, the subject lines, and offers, and discounts, and coupons, and, you know, body copy of an email or postcards, or whatever it is that we’re measuring, I’m not sure if you guys just stayed at digital, whatever the various and sundry channels are, those are hard thinking done by people then it seems as though the iterating through that stuff and coaxing out what works best with what segments that’s sort of going to be the machine’s job. I take it if we’re rolling through different holidays or rolling through different years, the company’s growing or coming out with new products, the process of continuing to feed the machine with new ideal marketing copy for Christmas, for Thanksgiving, for whatever these various and sundry holidays might be, or events might be, or offers might be that that process kind of continues. It seems as though there’s a pace and a rhythm between man and machine here in terms of feeding the machine and then letting it run, letting it send out what it wants to send, calibrating that, feeding the machine more pieces. I hope I’m getting the right mental picture of sort of the collaboration here.
– Yeah, you definitely get the right picture. I said it in the beginning, marketers are playing a vital role in that game. They definitely play a vital role in that game. They feed the machine, as you said, they feed the machine with creative, with content to communicate to the customers, with interesting topics to start the conversation about our customers as the machine find the best fit between those interesting offers and the customer personas and keep validated all the time.
– Got it. So the humans are looking behind… I’m painting this sort of mental picture for the folks listening in because I think it’s important for them to get an understanding of sort of where the technology stands today. So the various offers and creatives are rotated across the segments, the people obviously have access to some kind of dashboard where they are able to see how those response rates went. And then there’s some part of the process it sounds like, Yohai, where the machine may be in a certain week, or on a certain day, or whatever the case may be will begin rotating out messages that it believes are going to work best for different segments based on what’s worked sort of historically maybe to those similar segments. Is email marketing a good example to talk about here? I want to get a concrete example that we can mention. Is email a common channel for you, Yohai?
– Yeah, yeah. Email is very common channel here.
– Okay. Yep, yeah, I love using emails.
– And the example you gave is very relevant because the conclusions that the machine gets might be very insightful because if we see that for some segment we can’t find the best fit, so the machine can recommend to the marketer, hey, I need another action here. The offers that you gave me doesn’t fit. We can’t get a positive effect on that specific segment of customers. Please provide me some interesting content. And sometimes the segments will give some inspiration for our content because if we see that a segment of a sale seekers, we don’t get a positive uplift on it so the machine will ask the marketers for a more interesting and relevant offers for that specific segment and he will say, oh, that guy is a sale seeker so maybe I’ll offer them a coupon or maybe I’ll offer them some discount and then this process is always…it’s kind of a cycle. The machine gives feedback to the marketer and he provide it with more actions.
– For sure. Yes, okay, got it. It’s neat to note that there’s some kind of alert process there where it’s like, hey, you know, the following attempts aren’t really getting it done. We need additional material to try across this list sort of be our control, if you will. Is there a point where, let’s say, a given, I’m not sure if there is this…if this is even sort of possible at this point, if it’s more of a suggestion thing or if there is some autonomous action. Is there a point whereby the system might like, you know, you send out emails every Monday and Wednesday, for example, will the system maybe suggest what Wednesday’s email should be to a certain segment of let’s high frequency buyers? Or is the system at a level where marketers could set things to rotate the best newsletter offer you can on Wednesday go ahead and take a swing at it and allow it to run by itself? I know a lot of people obviously would want their hands on, maybe some people would be interested in the autonomous side of it, is it more of a suggestive intelligence at this point, Yohai, or is there some kind of autopilot of meshing the creative elements we have with what we believe is going to work best for this segment and just getting that email out the door on Wednesday with what we want autonomously? Or is there at this point anyway always going to be someone in the loop to look at the suggestions and pull the trigger?
– At this point it work more on matching and it takes like the content that it had, the offers that it have in the bank and try to match between them and the customer persona, but you definitely can do other experiments and try to change, like we can treat time as another variable in the equation and try to play with it and try to check whether an email that goes out Wednesday might go Monday and works better or even try to send it Thursdays. It depends like, we call it how much freedom we give to the machine, how much ranks of freedom, and according to that, the machine will recommend to calibrate the time or the specific offer to different spots.
– Yeah, and that’s suggesting and recommending, I think, seems to be the value of the AI sort of in the system here and being able to sort of match the right messages. It seems like, again, right now a lot of human in the loop, but I think that’s probably a good thing given where AI is today. Yohai, just being mindful of where we are on time, I have one last question. This should be pretty short, but I’m really interested in your perspective. You guys have been growing Optimove for quite a while now, you are selling this technology to all sorts of businesses, you know, here in the U.S. and overseas and whatnot. I’m interested in what the sort of machine learning applications in marketing are that you really think will be much closer to mainstream, let’s say, five years from now. There’s a lot of things that you folks do and a lot of things that other folks in the marketing space do. You know, churn prediction, suggesting subject lines, working on segmentation, you know, AI sort of to coax out some unique clusters and segments here that we could target in specific ways. There’s, you know, copywriting Artificial Intelligence, there’s all sorts of different approaches here to marketing and AI. What do you sort of core application, core functionality, if any, do you suspect will be much more common much, closer to the mainstream in five years? Something that really is going to get a lot of traction as an individual function or feature. What jumps out to you in that category?
– So I think of it as that machine would be more proactive. They will recommend on more actions that the marketer might do. They can tell him, hey, you neglected or overlooked a segment of customers over here. Let’s target them. Or I need more actions or more interesting offers to a set and I guess that the principle of ever increasing personalization would be more significant. I mean that like more data sources would be available and these machines would be able to digest and to aggregate data for more sources, both structured data and unstructured data. So we will increase the level of our knowledge base, we’ll increase our knowledge base, and of course, the response time to the customer behavior will go shorter and shorter. We already see many real time applications and applications that react ad-hoc to a customer activity or that can update their knowledge base on the customer on a real time and it leads us to a more focused and better timed offers. And in addition, I think that the capabilities of AI and machine learning provide more and more powerful tools that leads us to some kind of an arms race or marketers and if in the past it was like some kind of an interesting trend so it’s not the case anymore and now it’s really a professional need. Every model in the marketing department must have these abilities, but as I said at the beginning, I think that marketers won’t go out of the picture. For sure the next five years, like I don’t know what’s going to happen 10 years from now, but for the next 5 years, I think that marketers would continue to play a very important role in that game. They will have to come up with quality creative and we will have them on the picture.
– Yep. No doubt about that and I think that the tangible takeaways here in summation, Yohai. It sounds like you guys listening in you have at least 10 years until the robots get you. So for at least for marketers, certainly still playing an important role and I think Yohai speaks to an important trend that everybody should be mindful of. It seems pretty obvious that to some degree or another this degree of continued calibration is only going to continue and personalization in terms of real time and also of software getting smarter with its suggestions. I think that as a trend in the coming five years, having machines that are pulling from data sources and able to suggest actions that drive business value, that’s, you know, not exactly something that we could have done, you know, 10 years ago, 5 years ago in many regards and I think 5 years from now that will likely be some portion, some function and feature of a lot of our major marketing software out there. Your Marketos, your Infusionsoft, your HubSpots, etc. So we shall see what time will bring, but, Yohai, I’m glad you were able to share your perspective with us. Thanks for being here on the podcast.
– Thank you, my pleasure.
– That wraps up today’s episode here on the TechEmergence podcast and thanks for tuning in. If you’d like to stay in touch with the latest interviews with C-level executives, and top researchers, and thinkers in the domains of AI and the intersection of technology and intelligence, then make sure to subscribe here on iTunes or visit us on our main website at techemergence.com, where you can see all of our interviews broken now by category, as well as articles, news, market research and trends in Artificial Intelligence. If you found this episode particularly thought-provoking, feel free to leave your thoughts in a review here on iTunes or you can feel free to reach out to us at our main website. Thanks as always for tuning in and I’ll catch you next week.