Jack Welde helps companies make more money by speaking their customers’ language — literally. Welde is the Co-Founder and CEO of Smartling, a disruptive translation services company that uses a combination of human and machine translation to help companies enter new markets faster.
Welde says in the interview that consumers are 75 percent more likely to convert when they are being sold to in their native language — even if they are comfortable with the language they’re reading. Smartling measures the accuracy of translations with data, as well as the translations’ effectiveness in reaching new customers.
Andrew Weinreich: My guest today is Jack Welde, CEO and Founder of Smartling. And we are in Jack’s offices, his brand new offices. Jack, thanks for doing this with us.
Jack Welde: Thanks for having me. I’m excited about it.
Andrew Weinreich: Jack, before we get going and we talk about Smartling, can you give us a quick background on you?
Jack Welde: First of all you and I know each other from Penn, where we both went to school many, many years ago.
Andrew Weinreich: We were in, would you call them competing fraternities?
Jack Welde: I think we were competing fraternities, yes. Ours was the slightly cooler and more attractive brothers. Yours was the better at math fraternity.
Andrew Weinreich: Fair description.
Jack Welde: So, we know each other from Penn. My background is, I started a software company right out of school with a couple co-founders. We ended up selling that company to Apple back in the company mid 90s.
Andrew Weinreich: What was the name of that company?
Jack Welde: Well, our company was called Trio Development. There were three of us, we weren’t that creative. But we ended up selling our company to Apple and they shipped a product called Claris organizer back in the mid 90’s. After that, I went and flew airplanes with the US Air Force for about ten years. I went to Penn on an ROTC scholarship, and saw the world and did a lot of really interesting things. And it was leadership and management experience like no other. Then after that, came back to entrepreneurship.
Andrew Weinreich: So, you come back from Germany.
Jack Welde: Yup.
Andrew Weinreich: And you get a job with RunTime?
Jack Welde: I got a job with RunTime.
Andrew Weinreich: What does RunTime do?
Jack Welde: Runtime built content management systems for the enterprise and had a really nice product that had attraction in associations, and banks, and a variety of different organizations that help companies to create better content and to manage that content on their website. And this was in the early 2000s when companies realized that they had to have a website and needed content that was easily updated. And so, RunTime was one of the first companies to build content management systems. My co-founder at Smartling was the chief architect at RunTime.
Andrew Weinreich: So you’re the COO of RunTime?
Jack Welde: Mm hm.
Andrew Weinreich: Runtime does not run its course, you’re in search of the Jack Welde new opportunity. Is that right?
Jack Welde: Yeah, I mean, look I learned a lot. I thought it was a terrific company, and we’ve worked with a number of really amazing brands, and I was excited about it. But, yeah, this was a non-venture-backed company, that had really for not being a venture-backed company an incredible amount of success, and was very innovative, But yeah, I was looking for something that was the next thing.
Andrew Weinreich: Where did the idea of Smartling come from?
Jack Welde: The inspiration for Smartling was that I was working at a number of different companies, that they reached a certain size, and they realize they have customers all around the world. And if they are shipping a product that is not truly local, think like one tiny little restaurant or you do people’s taxes for a particular state and county-
Andrew Weinreich: Was there one company in particular that you said, you are leaving money on the table?
Jack Welde: Yeah, I mean, eMusic was really a company that had a global product. We had a product, we were selling legal Mp3 downloads, number two behind iTunes at the time. And we had a presence in the UK, we had a presence in Canada, we had a presence in Europe. But when you looked at the success of each of those different regions as well as the US, it very much correlated to languages. And so, in the UK, hey, you know, they speak English as well, slightly different variation of the English that you and I speak here in the US, but they had great traction that was very similar to the US site. In Canada it was about half, because half the people were preferring Canadian French, and half the people preferred English. In Europe, it was dismal and so-
Andrew Weinreich: Did eMusic ever translate or no?
Jack Welde: They never did. Not while I was there. But it was something that was on the roadmap, and it was always on the roadmap, is always something that every meeting we always talked about, hey, we have got to localize the product. We have got to get it out there. And the challenge with this was, if you have a website, or an e-commerce application, or whatever, some sort of customer application that you built ten years ago, at the time you didn’t really think about, hey, I might actually at some point want to deploy this content, or sell to, or support customers around the world. And so, people build things in English. They built it in a western character set rather than a Universal Unicode Character Set that can support Chinese characters and Russian characters. They did not really think about, hey, how is the layout of my webpage going to be different for German, or French, or whatever.
Andrew Weinreich: We talked about this before, when you build a simple website, people typically would put the English literally in the code. And so, when you think about adapting for other languages. it is harder to reprogram than if the language lived outside the code.
Jack Welde: That is exactly right. And by the way, that’s not necessarily a dumb thing. I mean most startups and most companies that’s just getting started, they are desperately trying to get to product market fit, and if they are building things very quickly to try to see, do I even have a product that people want to buy? The idea of saying, hey, let us take our time and do this right and really focus on that at some point we might want to be in 10, 20, 50 languages. That is really hard to do.
Andrew Weinreich: So that’s the Uber problem. And I want to focus principally on what Smartling is doing now, but I am particularly curious before we do that about getting going. When you started, was this a vaporware product? You had a first client in mind and you said, yeah, I have got this whole thing built and I will translate your site, or did you just start building a platform to translate and then find a client?
Jack Welde: So, it’s a really good question. And remember that, when we started this it was about eight years ago, and it was really before mobile had completely taken off. I mean the iPhone had launched and android had launched, but really mobile was not at the point where it is today. Where it was like, hey, obviously many companies have to have a mobile application that is sophisticated and mobile first. So, we were really solving the website problem. We were solving the problem of e-commerce applications and web applications, that had never been built to support multiple languages. So, we started out with a product that acted a bit like a magic trick that could take an existing website that might have been built in English and had never been built for multiple languages. And we could sit right on top of that website and act like a lens that magically would convert it to French or Spanish or Italian.
Andrew Weinreich: You built the product without a client?
Jack Welde: We built a product without a client, absolutely.
Andrew Weinreich: So you built the product and then once you had a prototype then you went out and look for a client.
Jack Welde: Yeah, I mean we had enough experience between my co-founder myself. We had enough experience understanding this was a really difficult problem. And we came up with the idea, that if we could essentially sit and layer on top of a website, we might be able to remove a lot of the complexity of doing this. And so, yeah, we started testing, we started building it, and then, we went and looked for a client.
Andrew Weinreich: So what was the thesis behind Smartling?
“…the problem is, is that most companies have built their website or their e-commerce application or whatever for one language.”
Jack Welde: The thesis behind Smartling was that in this decade that we’re right now, that globalization was going to be much more important. That companies are creating products and services that have applicability around the world. That more and more content is being digitized, more and more companies want to reach customers around the world. But there is a problem, and the problem is, is that most companies have built their website or their e-commerce application or whatever for one language. And if it’s in the US, that language is typically English. And so, when you said, if the CEO of the company said, hey, we need to be in 20 languages because we are getting a lot of customers and a lot of traffic and a lot of prospects that are coming from countries around the world that probably prefer a language other than English, let me walk down to my development team and ask them, hey, how long would it take just to get French, and Spanish, and Italian, and Chinese, and Japanese, and Hebrew, and whatever on my website? How long would it take to do that? The developers might come back and say, oh gosh boss, 6 months,12 months, 18 months. And the reason for that is because they built it for English and they intermix code and text, and they didn’t put it into separate data store, separate files for every single language. And it is a ton of work, it’s a ton of work to go back and completely rebuild that. And the thesis of Smartling was, could we do it faster.
Andrew Weinreich: I got it. Let me ask you some rapid-fire questions because I want to just get a context of how big this space is. So, World Wide Web traffic. What percentage of it is non-English?
Jack Welde: Oh English is a very small portion right now. It is a very, very small portion right now. And part of that is because of how many people are in China, and so on. But yeah, English is not anywhere like 90 percent or 80 percent. It is closer to 10 percent.
Andrew Weinreich: I am trying to frame whether the problem Smartling has gone after is, you have Facebook, and you have Google, and you have these large American-based companies-
Jack Welde: Which are by the way, in how many languages? How many languages is Facebook? And how many languages Google? And how many languages is the Amazon?
Andrew Weinreich: Yeah, but those are not your clients.
Jack Welde: Some of them are.
Andrew Weinreich: So some of them are confidential, which clients are yours. Why is that?
Jack Welde: Some of those big companies that we were referencing that would have lots of languages are our clients.
Andrew Weinreich: Why would they want to be secretive about being a client?
“…you are 75 percent more likely to convert to a customer if it’s in your native language.”
Jack Welde: Who knows why people want to do certain things that they want to do. But the point is that big companies typically get to be big companies either because they are supporting customers around the world, and the vast majority of their revenue comes from outside of English, and say, the United States. Or, they are trying to drive in that direction, and language can be a very good driver. You know, people want to buy in their native language. It’s a pretty obvious tautological kind of thing that people want to buy in their native language. But it is an incredible number even though if you are comfortable with English, even if you are a French-speaking native who is comfortable with English, you are 75 percent more likely to convert to a customer if it’s in your native language. So companies either translate because they want to drive towards that direction, or they have already gotten bigger and they want to be able to support customers around the world.
Andrew Weinreich: So, you have an airline that flies between the United States and France?
Jack Welde: Sure, among other places.
Andrew Weinreich: Amongst other places. And
Andrew Weinreich: And how did that client deal with translating their site in real time or close to real-time pre-Smartling?
Jack Welde: In some cases they didn’t do it, in some cases they just simply didn’t do it.
Andrew Weinreich: So this hypothetical, if you’re a French citizen, you would have to buy reading an English site.
Jack Welde: You would to buy reading an English site. And beyond that you might be using currency that you’re not comfortable with like dollars. You might be using a credit card that you’re not comfortable with or payment type that you’re not comfortable with. Imagine today many, many Chinese citizens are very comfortable with Alipay, whereas you and I may not be. And they’re probably not comfortable with American Express whereas you and I may be. And so, it’s beyond just language, it’s really localizing it for the expectations of the particular buyer. But to answer your question directly, yeah people often just didn’t translate because it’s a hard problem.
Andrew Weinreich: Are there commerce sites, are sites that expect you to engage in a transaction today overseas that don’t translate? That seems like that would be a very antiquated notion.
Jack Welde: Absolutely, there are. Yeah. And sometimes just, it’s a British company that just assumes, hey, British English is going to be close enough and will attract whatever. Sometimes American companies start by deploying globally and they have a go to market strategy that says, “We’re going to go global but we’re to start with English speaking countries like Australia, and the UK, and parts of Canada and so on.” But absolutely. In fact, there is far more content on the web that could ever possibly be translated by all of the human professional translators. So that’s the real problem right there. And companies are making choices about what they can translate, whether their infrastructure will support those translations, whether they think they’re going to be able to attract customers, whether they want more traffic, whether they want more downloads, whatever those business decisions that they’re trying to make.
Andrew Weinreich: And you have something like 20,000 freelance translators?
Jack Welde: Something in that order of magnitude.
Andrew Weinreich: How many years are we away from computer translation, where we get to a place where it’s equivalent quality to a human translation?
Jack Welde: Yeah. So that’s a super interesting question. And in light of the topic of data, I think that’s very, very relevant. So if you would ask me that question eight years ago when we started, our message was basically human good, machine bad. Human translation is good and computer-generated translation is bad. And it’s been a pretty close variation of that over the years. Now in the last year and a half or so, machine translation, computer-generated translation like Google Translate, as most people have some experience with, has really undergone a bit of a renaissance. There have been a number of really important breakthroughs, in particular something called Neural. Whereby using a more advanced set of compute tools, and essentially using GPUs and think of it like a-
Andrew Weinreich: A GPU is?
Jack Welde: Graphic Processing Unit. Think of it like the equivalent of sort of a 12 lane highway versus one or two-lane highway. Neural and different algorithms that are associated with neural machine translation has had a breakthrough where the translation sounds more human, has style that’s a bit more human, can take better advantage of data and of all of the different corpuses of content and text. And I think people even professional translators are frankly becoming surprised when they look at computer generate translation, and in some cases, they are going, “Wow, that’s really good.” Or saying, “Gosh, I didn’t think that that was a computer that translated except for this one or two words that are funky.”
Andrew Weinreich: So how far away are we from a place where those 20,000 people- Is your business, Uber we’ve got hundreds of thousands of drivers, but we’re on the cusp of autonomous cars and all of those people are going to be out of work? Two questions for you, one, how far are we away from that happening in the translation business? And what does that mean for your business?
Jack Welde: I don’t think we’re terribly close to the point that professional translators, professional human flesh and blood translators in the world are going to be out of work. But I do think that things are changing. And I think that the very simplified message of human good, machine bad is changing pretty rapidly. And companies want to be able to translate more content into more languages. If you had all the time and money in the world, you would be in every language that any human speaks around the world. Right? But it’s too expensive to do that. And there’s been a couple really important breakthroughs. So for example, Google Translate I think sometime in the last 12 months or so, Google Translate got to the point where it translates more in a day than all of the human translators combined in a year. And so companies are starting to experiment a bit with more of a mixed methodology and saying, “Certain types of content still really demand human translation.” Think marketing content, the kind of content that supposed to pull at your heartstrings, and convert a customer and be an emotional appeal to the customer, the computer just isn’t terribly good at that.
Andrew Weinreich: Because it’s nuanced. Because it’s using words that you wouldn’t find in-
Jack Welde: Totally. Because it’s nuanced. Because there’s a clever, witty phrase in English that only makes sense in one particular language. I think about an example I heard of recently which is, Harley Davidson. And one of the taglines that they had on their website was something like, “King of the road.” Now to Americans that kind of resonates because we think about the song, and we think about like, “Yeah, I’m riding around on my Harley and I’m king of the road.”
Andrew Weinreich: That’s a tough thing to translate.
Jack Welde: It’s a tough thing to translate. And what is it going to be in another language? Monarch of the highway? That doesn’t even make sense. So that’s a really trivial example. But that type of content that designed to have an emotional appeal and marketing content, the computers aren’t particularly good at. Now the other extreme, imagine support content for a software application where there’s a Zendesk support center someplace. It might say something like, “To get started, click the login button and enter your username and password.” That might appear 50 times. The computer’s going to get that right all 50 times.
Andrew Weinreich: Your business model, you described two components of it to me earlier when we spoke. One component is the translation, the other component really has to do with really your infrastructure. So the infrastructure consists of the workflow and the hosting of these pages.
Jack Welde: Well, let me adjust it a little bit. What I would say is, there are a couple different components to Smartling. We’re focused on business process optimization for translation. We’re helping companies to use technology and data to be better at their translation process. What does that mean? It means that at the moment you create content wherever you’re creating it, in a content management system, in a marketing platform, in support content, in an image that somebody is creating in Photoshop, wherever you’re creating it, let’s shorten the distance between when that contents created and when it’s deployed in the hands of somebody who says, “I prefer French. And I want to read it in French.” So that’s one part of what we do. And so we do that through customized workflows, we do it through integrations that can capture content and route it to the right translation resources. Whether it’s human, or machine, or some combination thereof, put it back, score it, tell the company, “Hey, this has actually been effective for you. You translated this content into these five languages and it moved the needle for your business. You had more customers, you had more downloads, you had higher customer satisfaction.” Whatever the metrics you care about. There’s another part of our business which is infrastructure, where we act like a huge API in the cloud where you can send large amounts of content, whether it’s a website, whether it’s web pages to be able to send that into the cloud and transform it on the fly into another language.
Andrew Weinreich: So for that part- I want to take each of those. So for that part, you charge some amount per page served?
Jack Welde: Yeah. We charge a de minimus amount of money based on utilization.
Andrew Weinreich: Jack, what about for the first part, are you charging for analytics?
Jack Welde: Yes. Yeah, great because a big part of what we’re doing is helping companies to be more efficient at their translation process. Managing and creating and deploying global content is complex, and you need analytics to help you with doing that. It’s a feature of our enterprise product, is the ability to gather that data and then deploy and show people insights and analytics.
Andrew Weinreich: Do you charge a licensing fee or are you asking them for some form of profit sharing on the optimization they’re seeing? Walk me through a use case.
Jack Welde: Yeah, so you buy the platform. It’s a typical SaaS, software as a service. So we have an annual fee and that annual fee is for the platform. That platform gives you a set of functionality. It might be connectors to major content management systems, it might be different customized workflows for that content, it might be team management, it might be managing all of your prior translations and storing those away and it might be all the analytics and reporting to help companies to make better data-driven decisions about the translation process.
Andrew Weinreich: Ah, the analytics is one piece of it.
Jack Welde: It’s a feature, and then we charge based on- look, if you’re translating, you know, a million words a year, that’s going to be a lower price than if you’re translating a billion words a year through our system.
Andrew Weinreich: Before we get to that, the first, the platform, you’re charging me based on-
Jack Welde: It’s features, product configuration.
Andrew Weinreich: So it’s features. It’s not based on- there’s no volume component to that first part?
Jack Welde: Yeah, I mean we add on top of that. If you translate a lot then yes, there’s a utilization fee. Just like most SaaS platforms have, the more you use it, the more you’re going to pay.
Andrew Weinreich: So walk me through the analytics that I would see in this hypothetical airline. I translate it to French. Walk me through what the analytics are that I would see. 50 percent of my traffic originates the United States, 50 percent or originates in France. Prior to you, my French clientele were reading in English. So now, what am I seeing about the single translation I have? What are the analytics I’m seeing?
Jack Welde: Yeah, so there’s a whole bunch of different things that we can gather and help to derive actionable insights from this data. And one of the decisions we made early on was storage is cheap, anything that we can think of, it lets store it away. So every single translation that occurs, you have one sentence that’s being translated in from English into let’s say French. So when that human translator or however the translation process occurs, we store away about a hundred different variables and data that goes with it: what time of day it was, how many characters was in the English, how many characters were there in the French, were there any spelling mistakes, were there any grammatical issues.
Andrew Weinreich: Why do I want to know that? Why does the airline want to know that? I mean basically you’re doing the work, I’ve outsourced it to you, why do I want to know how hard it was?
Jack Welde: Awesome, because while we’re not going expose necessarily all of those individual pieces of data but, you heard me say the number of characters, right? So, typically, a language like German takes up more space so than English, for example. So if you look at a sentence that might have a hundred characters in English, in German, German just happens to have more expansion in its language and the words it might contain at that length it might be 120 characters that we would expect to see. And if we saw a hundred characters of English then we saw 18 characters of German, something’s weird, something’s wrong. Now that’s two of a hundred different data points. If the translator did some translation and then as part of the workflow went to an editor and the translator took a minute and a half to do this paragraph of this small paragraph of text translation, and then the editor took three seconds and we looked at that over and over again. I can tell you one of two things is happening, either your translator is so good that you don’t need an editor, or your editors probably-
Andrew Weinreich: Got it. So because there is such a huge volume of stuff to translate and it’s not realistic for someone here within Smartling to go over and quality check everything, you use analytics to assess the likelihood, the statistical likelihood of them doing it correctly.
Jack Welde: A hundred percent, and let me actually expand on that a little bit. First of all, just like writing and you may not be a translator and you know and some of your audience members may not be translators, but we all prefer certain things in writing style. You might say, “Hey, I like that fast writing over this other writing. I actually think that this will be phrased better like this.” And there’s an awful, awful lot of subjective feedback that goes into the process and Smartling never wanted to be in the business of saying you know, “I’m happy to meet you,” versus, “I’m glad to meet you.” I don’t want to get into those kind of arguments whether happy or glad is a better choice for this particular thing. So here’s how we think about it. I gave a couple really simple examples but we track about a hundred different variables that go into this. And instead of coming back and saying, “Oh, well based on these things, we think this was, that this specifically was a great translation or a poor translation.” Instead, what we’re doing is we’re bundling it together and we’re using that data and applying machine learning techniques, to say, just like you said a minute ago, probabilistically, there’s this statistical model or this is a probabilistic approach. This is a machine learning approach to say, that based on these different variables, it’s highly likely that you’re gonna have a good outcome. Or it’s highly likely that something is wrong with this.
Andrew Weinreich: Do you know what makes me think of?
Jack Welde: Tell me.
Andrew Weinreich: I remember one of the large photo sharing sites. This was probably 10 years ago. Had so many photos uploaded they couldn’t figure out how to screen out the inappropriate, the pornographic ones. And someone came up with the insight, that any photo that receives a critical amount of traffic, within the first 30 seconds or 60 seconds was, with a high level of certainty, inappropriate and they just deleted those. It was their shorthand for- I think what you’re saying is-
Jack Welde: Which is weird because your vacation pictures I’m sure are very very popular and they’re just kept getting deleted, right?
Andrew Weinreich: [Laughs] I think what you’re saying is, the opportunity for translating all of the world’s content is immense, but has always been constrained by the fact that the quality was poor and there was no ability for there to be oversight or review. And the great breakthrough you’re saying that you’re having now is this machine learning, as you’re calling it, based on a statistical analysis of what you would expect, is the way that you could scale translations.
Jack Welde: You’re on the right path, but I want to tweak it. Your example where you said, “Hey, certain inappropriate pictures probably a lot of traffic and let’s just automatically delete them.” That’s rule-based. A programmer went in and said, “If it gets this much traffic over this period of time, go ahead and delete it.” It could be wrong, it could be false positives, but somebody taught the computer explicitly how did you think about it.
Andrew Weinreich: That’s literally how they did it.
“Instead of teaching the computer here the things to look for, the computer could look at all this data and do a far better job of ascertaining that these particular data points are indicative of, this person may have cancer.”
Jack Welde: Yeah, so that’s rule-based. There are some areas where we can say, if this has a spelling mistake in it, it potentially could be a problem, and so there are certain aspects that we do that are rule-based. There are other parts, and here’s the analog that I like a lot. In the last year, I was reading about this super interesting medical case study where this was a radiology center that was essentially trying to teach the computer to take a look at cancer imagery and identify likely breast cancer patients. And so, there were about eight different factors that doctors and radiologists were using and so, what they were trying to do was teach the computer to identify them better the doctors. And the thinking was like that the radiologist get tired after launch and might miss something.
Andrew Weinreich: These are leading indicators of breast cancer?
Jack Welde: Of breast cancer, that’s right.
Andrew Weinreich: In other words, prior to someone having it.
Jack Welde: Yes, it could be an indicator that they may have had it and it might depend. It could be anywhere along there, you know, their spectrum. That’s right. But the problem with this was that they were teaching the computer to say, if you see any of these eight things, then this is a problem. Now, there was a bit of a breakthrough. The computer was actually slightly better than the doctors with this. Now the breakthrough was to say, “Let’s not teach the computer explicitly.” Let’s not say, if you get a lot of traffic to a photograph, then let’s go ahead and delete it. Instead what they did was they said, “We’ve got hundreds of thousands of scans and we also have the mortality rates of the women.” Did these women die? Sorry to be a little morbid. Did they die in three weeks, three months, three years, 30 years or still alive? And by looking at these scans and throwing this into a machine learning application, an important breakthrough happened. Instead of teaching the computer here the things to look for, the computer could look at all this data and do a far better job of ascertaining that these particular data points are indicative of, this person may have cancer.
And the other breakthrough, first of all, it was far better performance than the human radiologist, but even more than that was that the computer identified three additional markers, three additional things that radiologist hadn’t even realized that these are also potential indicators of cancer. Now, think about how incredibly powerful that is and think about that we may be talking about this for translation or maybe talking about it for medical. The real breakthrough that’s going to happen, and it’s happening already, is that every company out there will have a machine learning strategy and will say that certain things the computer is better at people, and certain things the people are better than the computers, and having these two things work together is really, really important. I shared earlier. I was a former military pilot for ten years. You know, if you think about a typical commercial pilot today, do you know how much time that the pilot that he or she has their hands on the stick or the yoke that are actually flying the airplane? About 10 minutes, five minutes on takeoff and five minutes on landing.
Andrew Weinreich: That’s a reassuring thought.
Jack Welde: But think about it, it’s the autopilot. It’s the computer that’s flying in the ascent, at altitude, and in the descent. Honestly, no pilot wants to be hand-flying an airplane for eight hours at 35,000 feet. They would rather have the computer do it. But the concerning thought would be, are we ready to have completely unpiloted airplanes? I don’t think we are. And I think people like you and me have comfort in the fact that there is a human out there that’s saying, “Let me check what the computer is telling me, what the autopilot doing. Oh, wait a second. You know there’s a thunderstorm up there. Something’s going on with my flight attendant. Hey, this person had a heart attack back in the back. Hey, I want to see if I can optimize for fuel or there’s turbulence here.” And so the computer is doing most of the flying but it is the human element that’s going into this, and this is where, in translation, we’re going to see more of a mixing of human translation and machine translation and a combination of these two things combined with data and machine learning to produce better insights and better outcomes.
Andrew Weinreich: I want to stay on the machine learning and the translation. If you’ve developed artificial intelligence with machine learning that can gauge the likelihood of successful translation, isn’t that interesting to use whether or not Smartling is the underlying translation workforce?
Jack Welde: Absolutely, super interesting to use.
Andrew Weinreich: So, couldn’t that be used by your competitors? Wouldn’t you want to say to your competitors that not only do we use this machine learning on top of our translation but we’re going to extend it as a service to you as well. And so, we will both sell direct to companies and we’ll sell to our competitors?
Jack Welde: Yeah. Well, I think there’s a couple of things. Number one, from day one, Smartling started gathering this data. And we were born in the cloud from day one, and the storage is cheap, and we have been storing this data away for eight years. And I would argue that we probably have more data stored that can be used for these types of insights than any other company out there. And so, unless you’ve started gathering this data and you’ve made it a mission, critical part of what you’re doing, you don’t have the data to look at. And machine learning algorithms, neural machine translation, all of these new AI oriented types of things are super data sensitive. So you’ve got to have the data in the first place. By the way, our point of view on this is, we provide a translation platform to help companies to translate content better and to have better outcomes using that technology and data. But we give our customers choice. If you want to use your favorite internal translators. If you have a favorite translation agency you’ve been working with. If you’re working with 20 translation agencies, we have a few customers, for example, that they use their massive community of passionate, 20,000 passionate users or whatever.
Andrew Weinreich: And they crowdsource.
Jack Welde: They crowdsource the translation from there. We’re really indifferent to that. And in all of those cases, we can still provide that data and provide those insights.
Andrew Weinreich: And the data assesses the success not the likelihood of success.
Jack Welde: It absolutely does.
Andrew Weinreich: So, I walked in here, I walked by your tables where your sales reps are. Will you have people calling on competitors? I appreciate what you just said, which is, we’re indifferent, I don’t know if you’re indifferent, but you’re okay with your clients using internal or crowdsourcing or other, but will you actually open up a division here that says, we’re actually going to call on our competitors and say, we can help you deliver a better product?
Jack Welde: I think we would call those partners. I mean look, we want to be the platform that companies choose because they believe that it will allow them to have a better outcome. So we have a lot of partners that we work with from here. And in an industry that is as large and as fragmented as the translation industry is, there’s about $40 billion that’s spent on translation software and services and interpreting services every year. And there’s something like 26,000 translation agencies in the world. And so, it’s a highly fragmented market. And so, we want to be partners with anybody out there that sees the value in a software platform that helps them to do this better.
Andrew Weinreich: Just so I understand, if we did get to that place 20, 30 years from now, where 50 years from now, some period of time from now where Google Translate completely replaced human translation, would we still need the machine learning to assess the quality and the accuracy of what this sort of utopian translation look like?
“Just like, you know, think back to the early days, where people thought, ‘do I need a website? Because I’m in the Yellow Pages, isn’t that sufficient?’ Now people are starting to say, ‘do I need a machine learning strategy?'”
Jack Welde: Well, first of all, I think by then the robots will have taken over and killed us all. So maybe not. But I think this idea that for the first time in human history, we can store amazingly vast amounts of data. And then, you can apply a variety of different algorithms and that there is a foundation of machine learning and AI work that’s being done all over the place around the world that you can build on top of, to provide the kind of actionable results and to pick up patterns that humans aren’t particularly good at. I mean, I think that is the fundamental breakthrough.
Look, you and I both came into technology pre-Web and pre sort of World Wide Web. And certainly, you and I both recognized that, wow, this is the first time where there is a lot of different parts that come together. The network called the World Wide Web, a browser that allows me to do it, a language that allows me to render these different kinds of things, you know different code that I can do certain things, rather than me as one company having to go build all of that. The same kind of thing is happening with machine learning right now. Where companies like Amazon, and Google, and Microsoft, and others are allowing people to deploy servers that are built with a foundation of machine learning capabilities with a much smaller footprint of code. You know, I’ve got a right 5,000 10,000 lines of code, not a million lines of code to do this because the foundation exists. Just like the Internet, just like the World Wide Web, just like the browser, just like HTML existed to help me to do this. That’s profound, and it allows companies whatever the thing is that you’re trying to figure out where humans are not particularly good at taking massive data sets and saying, here’s what we should do with it. Here’s the insight. Here’s how we should be thinking about that. We’re at a point now where every company has to have a machine learning strategy. Just like, you know, think back to the early days, where people thought, do I need a website? Because I’m in the Yellow Pages, isn’t that sufficient?’ Now people are starting to say, ‘do I need a machine learning strategy?’
Andrew Weinreich: One more area I want to drill down on before we break, so I appreciate the machine learning strategy. I appreciate the data that you’re offering clients which is the basis of that machine learning strategy about the efficacy of their translation?
Jack Welde: Yeah, totally. So, one thing is that part of our infrastructure product actually tells customers, hey, these particularly URLs, whatever they want to track, could be a shopping cart, checkout, could be some action-oriented page they want people to go do something on we can tell them the full breakdown of how many people came to this page and took the action that you wanted in this particular language, in French or Spanish or whatever else-
Andrew Weinreich: They can’t get that anywhere else?
Jack Welde: They can get it someplace else and it lives in Google Analytics, and it lives in Splunk, and it lives in places that sometimes it’s not terribly easy for the marketing manager or the localization manager to walk down to the IT area of the company and talk to somebody and say, hey, can you pull this data out? And so, yeah, we can pull that data out and show people how effective was this page for that particular language that that particular user was looking at. But really what we’re trying to do, is we’re trying to help companies to make better decisions about what they translate in the first place into what languages. And unfortunately, no company in the world can afford to translate all of their content into all of the languages that they would like to translate with human translators. And even with machine translation, from what we talked about before, you may not have the quality that you’re looking for this. So companies have to make Pareto type choices. And they’ve got to say, which of my content, which 20 percent of my content is going to produce 80 percent of the impact? So one of the things we can help companies to do is to say, let’s actually look at real-time what’s actually happening on your website, on your mobile app. Are you getting more traffic, more downloads, more purchases, higher purchase amounts-
Andrew Weinreich: And as a result, you should do what? You’ve already translated. So what’s the double down?
Jack Welde: You should potentially, well, you’re always creating new content. Every company is creating new content all the time and it’s all being sucked into a system like Smartling to be translated and then deployed back. And you might decide that, hey, we tried this out and we tried it and for whatever reason, we translated it into German but people prefer the English version of the page and they’re just not, you know, they’re just not converting or whatever else.
Sometimes we also find we’ve had customers where they’ve had a particular set of results on three different languages and they look roughly the same, and then all of a sudden, the German version suddenly takes off and French is considerably lower. And so, you can start to ask the question about, well is that because of something that happened with the languages or did you get different products that were only available in one particular country or region? Did you have a different go to marketing approach? Did you suddenly do a whole bunch of marketing and press for it in that particular area? And you could start to make better decisions about the efficacy of your translation in the first place.
Andrew Weinreich: I love your anecdotes, your insights about different markets or the efficacy of different programs. Is there a commercial opportunity for you to anonymize all the data and sell to your clients in the aggregate?
Jack Welde: Yeah, a hundred percent. And that would that would be like a benchmarking product. So, it depends on what people are trying to solve for. So, we have a lot of different reports, and different things that we can help customers to understand what’s working well in their process and what’s not working well. And I’m going to give a really simple example. A very typical thing that a lot of companies, particularly companies that are super passionate about their brand and really focused on their brand, what they typically will do is content gets created let’s say in English. And then, it gets put into a workflow. And the workflow is it goes to a translator, a human translator, then followed by a human editor. Now you’ve got to two different sets of eyeballs looking at, it looking for consistency, making sure that there’s no mistakes, making sure everything’s good. And then, it gets delivered back to the company. But then they put it into an internal review. And a lot of times that internal review says, oh Andrew, you’re in the Berlin office and you speak German. And Jaques, you’re in the Paris office and you speak French. And so and so is in the Beijing office, and you speak Chinese. Even though this isn’t really your job, would you mind reviewing these things. And guess what happens. That stuff gets bottlenecked really, really badly for weeks at a time because you have a day job. And you don’t necessarily want to go review the French or German or Chinese version of that content. And so it gets stuck and you’re delaying time to market.
It’s a really trivial example, but one of the many, many reports that we provide for our customers says, where are your bottlenecks? Where are you delaying getting content to market it in the hands of consumers that want to read your content in their language. So, that’s a good example where we can show people you’re bottlenecked in your internal review. How do we make that faster? But what’s even more powerful, is when we can say, let’s actually compare this to all of the other customers that we have in an anonymized fashion and say, guess what? Your internal review is taking five times longer than the average customer in your industry doing whatever.
Andrew Weinreich: Is that a paid service right now or is it something that you informally do?
Jack Welde: Well, it’s something that we are gathering the data around, and it’s something that we’ve been building for some time now, and it’s something that we help customers with on an ad hoc basis, but slowly but surely, we are starting to deploy this across all of our customers.
Andrew Weinreich: I had a more basic example.
Jack Welde: Yeah, tell me.
Andrew Weinreich: I have a finance-focused, high traffic blog.
Jack Welde: Have you translated that blog because you’ll get more traffic?
Andrew Weinreich: And let’s assume it’s not translated. And I monetize that with- we could come up with one or two hypotheticals, either I monetize it with a direct sales team or maybe even monetize it with a Google Adsense. It would be great if I could go to Smartling. And let’s assume I’m a high traffic blog, and I’m delivering common sense financial advice about amounts of money you should save, or how you should think about budgeting. So, may or may not be applicable everywhere, but there should be some applicability outside of United States and outside of English. It would be great if I could come Smartling and you could say to me, listen, the very first place you should translate is in Germany. And by the way, we have these relationships with how you would monetize it, but I can project for you, given run of network advertising, if you translate into Germany, what the drop-off would be? And so you could give people an instant sense of what they could benefit from translating, and I use the content example because if it was a commerce example maybe there’s other levels of complexity. I’ve got to accept in other currencies, need other business relationships, but for the content example, I just translate and people consume the content or they don’t. Based on the fact that you’re looking at data across lots of content properties, isn’t that something you should be able to tell me how much money I’m leaving on the table, which languages to translate to first, and exactly the steps I would go to make more money?
Jack Welde: Yeah. So I love the example. I think in reality, it’s a bit more nuanced than that, and I’d love to be able to say that and I don’t think we’re there. I don’t think any company is really there to do that in a truly automated benchmarked approach. But look, maybe that is a future where you can help people to be able to do this. There’s a number of challenges.
Andrew Weinreich: Why wouldn’t the data you have you’d be able to tell a content property, if you translate, here’s what you can expect?
Jack Welde: Well, you skipped over a couple steps where you said, and I’m going to monetize it in some fashion. Let’s suppose that you were monetizing it through ads. And in some parts of the world, ad networks just don’t perform anywhere near, like they would inside of or whatever.
Andrew Weinreinch: Let’s assume we allow you to put a hypothetical CPM in. In other words what pageviews could I accept? What you should be able to give me, I would think, without any other data points is based on what you’ve seen with comparables, what pageviews could I expect to generate in a host language with the only data input, what I am generating now and the content?
Jack Welde: If we were willing to use hypotheticals like this ad this ad network, this magical ad network, that we drop in that’s perfect for the Brazil market. And if we translate into Brazilian Portuguese based on a variety of other factors across our other customers that look like you, we think you’re going to get this many pageviews, and based on the amount of content that you have and the translation cost, however you translate that, we can literally turn that into, it’s going to cost you X dollars to do this-
Andrew Weinreich: Why isn’t that the holy grail for every content property?
Jack Welde: To some extent this is right. I guess what I’m sort of putting a little nuance around is the content idea that you have a blog post that you’re going to monetize with an ad network and honestly in just in many parts of the world, it’s just hard.
Andrew Weinreich: Yeah but if I’m the New York Post, there’s got to be some subset of people. If you were looking at enough companies, then again the New York Post is principally New York-
Jack Welde: Again, that’s exactly the problem.
Andrew Weinreich: But if you’re seeing enough companies that have New York content that if translated at some point and you might be able to tell me quick cost-benefit analysis, should the New York Post translate into Portuguese?
Jack Welde: You’re on the right track here. And let’s take a slightly different example, which is you have a mobile app. And it’s a social mobile app, and think about your background and social. It’s a social mobile app. And so what do you want? You want as many users as possible, right? And you’re not necessarily worried about monetizing, for example. I could tell you right away, without any data and without any machine learning and any whatever, the number one thing you should probably do is translate into as many languages as possible. Why? Because you got to get more downloads. And you’re going to move up the rankings in Apple, and there’s not a lot of content on typical mobile apps compared to blog posts and blog sites things like that. And so you should be in as many languages as you can possibly afford, because then, when people in any different country out there are browsing the App Store, they’re going to find your application downloaded and it’s going to increase your rankings from there.
If you have any kind of transactional system that like hotels, airlines, any kind of product that you can sell or a service that you can sell. And by the time you back out of any shipping or any other taxes and anything else that comes from that, if it turns out that this really truly is ROI positive, then you want to be in as many languages possible for that. So, you’re on the right path, and this is absolutely the kind of conversation we have with people.
In fact, where we typically start as we say, why you’re thinking about the fact that you can’t do every language across all of your content, so you’ve got to make some 80/20 type Pareto types of choices with this, think about what’s more important to you. Do you want users or do you want dollars, or Euros, or whatever else? Do you want users, or do you want dollars? And in some cases, you may look at parts of the world that just have a ton of people. Think China, but may not have the economic share of wallet as other places like, perhaps Western Europe. And these are the kinds of decisions people need to be able to make. Now, what Smartling wants to be able to then do is say, now that you have that thesis, let’s go translate, and let’s real-time, constantly, all the time, update and say, what’s actually happening here? Was that thesis correct? Should you continue investing in French and Italian and German and Spanish and whatever? Or are you actually finding that you’re missing a few languages, that if you were to do some of those, if you added Brazilian Portuguese, it absolutely would help your business. And by the way for whatever reason, Italian’s not converting.
Andrew Weinreich: Anything else you want to add before we wrap?
Jack Welde: I think this was terrific. I think this was a lot of fun. I think that Smartling is going to look like many of the other companies out there that is really taking advantage of data and using machine learning to help our customers to be more successful.