Paul Danter on Welocalize Opal, AI Translation, and the Future of Global Content

Russ Fordyce 🏆 (00:01.506) Welcome back to the winner's circle, everybody. I'm here today with Paul from WeLocalize. They have been working in the language space for many, many years, and now they're working in the AI space and have just won the AI Excellence Award for 2026. Paul, welcome to the show and congratulations on the award. Paul Danter (00:02.358) Thank Paul Danter (00:19.156) Hey Russ, nice to meet you and yeah, thank you so much. It's really a huge team effort and the team that built our OPL platform, which we'll talk a little bit about today. We're really behind that and there's a lot of huge effort that's gone into that from that team. Russ Fordyce 🏆 (00:32.812) Yeah, we were talking about it before we started a little bit of kind of the annoyances of trying to be international and translating all, especially technical documentation or technical marketing is what I always ran into. That was always a challenge for me in my career. So give us a little bit of background about we localize and the Opal product. Paul Danter (00:53.483) Yeah, sure. Well, we localize as a company has been around for more than 20 years. We really been helping many large clients around the world, technical clients and otherwise, really be able to engage with their clients on a global level. So that can be everything from marketing materials all the way through to technical documentation. And you can imagine over 20 years, a lot of that work has changed and evolved as we've tried to streamline that process of being able to take take any content that they produce and then just make it available and make it resonate with the markets that they're trying to grow into. So we've continued to focus on that over the last few years. And as you mentioned, Opal is our latest iteration in that journey that we're on. We are working very much in terms of the AI in regards to how do we optimize the way that... AI can enhance that ability to be able to do that at scale and do it in an optimal way that represents a brand really well. Russ Fordyce 🏆 (01:55.374) So for those of you that, you know, we're not marketers 24 seven, 365, describe kind of Opal, you know, and why like an enterprise leader who's never bought any language services and is kind of trying to understand the value of that. Just give us kind of that elevator pitch on, on why somebody would want to even go into this area. I know like healthcare and other guys want to do it, but why is why for an enterprise team? Paul Danter (02:20.093) Yeah, well, enterprise teams are tasked with obviously unlocking global markets, right? So you need to be able to sell your products in global markets. And that means everything from marketing them all the way through to then supporting your user base once you've launched your products in those markets. And so over the last few years, really what we've tried to do is to think about the possibility that AI... offers us in terms of being able to do that translation work, the core translation work that's needed to do that work. And then to do that in an optimal way. And what I mean by that is that it's really how do you operationalize AI in a way that will allow you to take language that represents your brand, represents your tone of voice, and then be able to use that as part of your process. So you do sound differentiated, that you do sound like your company likes to sound when you communicate. communicate with your clients and then it effectively forms part of that process. So the way that translations always worked is that a human has typically been in the loop. So back in the day, that would have been a translator that would have actually done the work for taking some content from say English to German. And then over the last few years, more of that has been automated through things like neural machine translation. And then... the human being has been involved in more of a post-editing aspect of work where you're just really checking that the translation that's come out of the machine and then making sure it's valid and represents the brand correctly. Now what we've developed with Opal is how do you take the machine output from something like neural machine translation and then do that post-editing work using AI as well? And AI is really good at tone of voice and those sorts of things with the right amount of training. So Opal has really been developed to allow you to do that post editing and then to also score the end result of that. So you can kind of gauge whether or not that still needs to be looked at by a human to validate the output. And then the trick is also then taking that human validation and then using reinforcement learning within the AI model to improve your result. Paul Danter (04:31.286) And a lot of it is really based around the training that you need to do in order to make that model sound like your company. So there are lot of very generic sort of sounding models that are out there. And what we've been really working hard to do is to come up with the right mix of using the right model, the right training techniques to be able to ensure that when you do use that model to do the post-editing work to sound like your brand, it really does sound like your brand. and then the output can be more automated in the future. So we see that services post-editing work becoming less and less as time goes by and the models become better and better at understanding who you are as a brand and then the automation becomes more and more sophisticated and representative of what you're trying to do. Russ Fordyce 🏆 (05:21.336) Well, so it's super interesting because you guys had this legacy content, I imagine, this portfolio or library of jobs that you had worked on where a human did a full translation or a partial translation or a partial check. Was it AI that really, generative AI that really got you beyond those neural networks, machine learning content automation piece? Was it AI that let you really kind of you know, and I guess all the LLMs now kind of automating that quality decision. Paul Danter (05:55.938) Yeah, yeah, I mean, think the way I look at it is maybe I'm oversimplifying, but we in the language industry, I mean, it's the second L in LLM, right, is language. And there's a reason for that. A lot of those models were originally built around things like neural translation engines. So we'd been working in this area for many years. And the initial ways that those engines work were more of what I would describe as as predictive models. it's kind of understanding what you're trying to say and then translating it in a predictive way based on things that you've done in the past, translations that you know to be good in the past. What happened when, you know, in the last couple of years is this sort of generative AI capability came about and that's way better than sort of being able to think about tone and terminology and brand voice in a way that makes sense. So you're able to kind of move that downstream a little bit in terms of the quality of the output of the content that you're trying to work with. So yeah, a lot of it has been about sort of this evolution. And then it's about, as I said earlier on, like the operationalization of that then. So many of our customers, we're pushing through literal millions of words a day of translated content. And so you being able to do that in a way that works and that you have, first of all, the technology in place that can handle that kind of scale. but then also the kind of human capital that's required then for some of the post-editing work, some of the quality work that needs to happen as well to ensure that we're continuing to refine and reinforce the learning of those models. Russ Fordyce 🏆 (07:31.212) Yeah, I always envisioned kind of the early models of AI actually, you know, having that, you know, set from my iPhone kind of on the bottom of a translation so that you'd know it was maybe an AI translation if something got wacky. But it sounds like you guys are way beyond that now. Paul Danter (07:47.382) Yeah, well, think, yeah, that's part of the complexity of what we manage as a business, to be honest with you. So I think that what you really need to understand with a client is the types of content they produce and what is the right and appropriate level of intervention that you need to make as you process that content. So if it's something like a very basic support article for a piece of technology, probably the translation doesn't need to be indicative of your brand voice and If it solves the problem, if it helps me reboot my cable modem and understand what happens, then that's probably good enough, right? But then if something is very critical to a new product launch and the way you're positioning that against your competition in the market, that probably needs to follow a different workflow, a different path, and you need to apply different principles to that content. So it's the kind of engineering of those workflows and the right time to use the right technology at the right place that becomes part of this. this mix and really that's what this industry has been about for many years is how do you manage that sort of chaos really of content and then how do you apply technology in a way that's really meaningful to the right content at the right time so that it has the biggest impact for the business. So you're spending where you need to spend on content and then you're making do with content that's good enough to go out the door for certain aspects of what you want to produce. And I think that that's Russ Fordyce 🏆 (09:13.292) was Paul Danter (09:14.311) Like another thing that's really changing as part of this AI adoption that is happening across all of our clients is really thinking about what is good enough and what does quality look like? Traditionally, industries around language, around linguistics have been very academic. So there are academic scores of, know, does something use the right terminology? Does it sound right? Is the tone right? Is is the fluency correct, et cetera? And I think that that's actually starting to change now. I think it's becoming less of an academic scoring mechanism when you look at what comes out of AI. And what we're really working on is how do you govern your whole enterprise around where content needs to be adapted? How can it be fit for purpose? And we're even just starting to think about it, not in regards to the quality of the content that comes out, but really the risk of that content. And what I mean by that is if you apply this AI principle for this particular workflow for this piece of content, what's the risk if it gets it wrong? Is the risk that you're not gonna get as much click through on a particular piece of marketing or is the risk that an end user might not be able to use a support article to be able to fix something that they're really struggling with? So we're really starting to think about it in terms of content risk as all those flows come together to be able to produce an end result that actually has an impact on our clients business, which is the most important thing. Russ Fordyce 🏆 (10:42.22) Also walk us through a little bit if I'm a new enterprise or I'm new to you, new Fortune 500 enterprise coming to you. you're talking about millions of pages and I can imagine if you get a good client, you're really looking at just volumes of content if they're trying to localize everything. It sounds like you guys not only have kind of this AI, modern AI translation engine, but a system in the background actually helped me ship all that content too. Paul Danter (10:49.159) Yeah. Paul Danter (11:10.481) Yeah, that's a large part of what we do. So what we will actually do is kind of a content audit with our clients to really understand the systems that we need to connect to in order to be able to pull that content. And then the rules that we need to apply to make sure that we are handling that content in the right way and in the best way for their business. So there's a lot of work that we do with clients to be able to understand that and to be able to really work with them because often there's a, you know, I think Most businesses I've experienced working with have an array of different types of content that live in different systems and the source of truth for anything is really difficult to come by. So helping them understand where everything lives and then the best way to connect with that so that we can actually then handle that content in the correct way is really, really meaningful. Interestingly, the other end of that needs to be thought about as well, not just in terms of content, but in terms of impact. there are a number of clients we're working with right now as part of the Opal product set to be able to understand what is the impact of content, could it improve, and is there some kind of signal that we can send, surround whether or not it's working for you. So we are kind of then being able to perhaps regenerate content in a different way that we may be able to measure and may be able to understand that it's having a better impact on something for a client. So that very relevant to things like marketing. but can also be applied in many different ways across different content types. Russ Fordyce 🏆 (12:42.094) I imagine the support angle is huge right now and trying to get support out to, you everybody's trying to reduce costs of support and get that out of the human channel and into AI channels just for, you know, everybody's sake. It tends to land better for everybody until you get to needing, you know, tier two or tier three support. Paul Danter (12:59.778) Yeah, 100%. And yeah, and I think this is the critical thing right now. So the management of an AI process in this way, I think that we're all very excited, right? Anybody in tech about the possibilities of what we can do with this technology. But what we've tried to do with Opal is really constrain it into how do we make the best possible AI system that will allow you to post edit your content in a way that represents who you are as a brand and your brand voice as I mentioned earlier on. And actually that's important, not just from just can we translate content better, but it's actually important from a governance perspective as well, because I think that as everyone runs at these things, what we're seeing is that the control of how work gets done when you empower somebody with some AI tooling can just become a disaster, right? So all of a sudden you've got... anybody with an AI tool thinking that, hey, we can just push some content through this and let's see what the output could be. But you actually still need some control over that process or A, content's going to get regenerated multiple times by multiple people with no control over the quality of the output. And B, you don't even know what you're spending anymore. we often talk about how we think that companies are going to have this kind of moment in a few months where they wake up and they're thinking, we're spending all these tokens, like who's using all this? And how do we exert some control over that? So what we're trying to do with Opal is put some sort of framework around the way that you're able to harness it to be able to kind of manage that content at scale and have some kind of control as to, again, the right way to be able to handle any given content type and then also to make sure that you have a very clean channel through which that content needs to get pushed through in order to go. to global markets in a way that's meaningful for you. Russ Fordyce 🏆 (14:57.164) Yeah, I think you've summarized kind of the biggest problem with AI that I have and I see others having, which is it can do so much so quickly that you have to have the human resources and the process and all the protocols on the backside to be ready for all that content because there's just this tsunami coming at you, right? What do most people outside of the industry kind of underestimate about this work? Paul Danter (15:15.191) 100%. Paul Danter (15:20.599) Well, they're really underestimating the fact that while there are capabilities within AI to do certain languages and even certain languages well, first of all, training it and making it relevant to your business and your terminology and your brand voice, as I mentioned, is very, important. But then also being able to control just the volume of content that's going through that system. What we've experienced just recently, is this explosion of content because people, even in English, are generating more content than they've ever generated before. It's very easy now to be able to mine data within an organization and pull these abstract ideas together and formulate it into a piece of content that's actually quite compelling, but you're empowering your whole team to be able to do that across, often across very large organizations. And that's resulting in this... exponential kind of creation of content that needs to be managed. And I think that being able to kind of wrap your arms around that and put processes in place is going to be the next challenge for a lot of major enterprises. Because as it stands today, I think that there's an experimentation phase that we're living through right now that is resulting in huge amounts of content that's getting generated. And I think that that's exciting, but also at some point going to become a little bit of a load that businesses can't bear. Like where do you find the best content? And then how do you handle that content for global markets is a critical aspect of what people need to think about. Russ Fordyce 🏆 (16:59.598) Well, I was remembering with you my early days in international business of having things translated and it's always been kind of a pain in the neck. But with AI and with the neural networks you're talking about, it seems like it's gotten significantly better. Without getting into proprietary numbers, how did you and your customers go about measuring whether OPERA was actually working? Paul Danter (17:25.406) Yeah, so we were measuring things like time to market, so the turnaround time that you're able to be able to handle a piece of content. And of course, I mentioned earlier on that we've done a lot of work to speed up that process anyway prior to any of this technology getting implemented. But we're finding it's 50 % faster to be able to take content and have something that's validated and ready to go out the door, which is massive. And we're also seeing up to 26 % reduction in time for human beings doing any of that human loop editing as well. So the system definitely gets better over time. The more it's able to be trained on content, the more reinforcement learning that goes into it, the bigger impact that that's having right now as well. currently we're supporting over 35 languages in terms of the languages that are. where we're able to train and we feel comfortable with the quality of the output that they're able to produce. But at the same time, we're looking at the next 200 languages and we have a certification process by which we're working to make sure that the models that are trained in some of the more long tail languages already are capable of being able to do this work in the future. Russ Fordyce 🏆 (18:40.024) So when you kind of launched Opal, I imagine you had a couple of beta customers that were early deployments. What did you kind of didn't expect that you saw out of those early ones or what friction did you hit that you didn't expect? Paul Danter (18:56.176) A lot of the friction is actually just in terms of getting a lot of the things together that you need in order to be able to do the initial training. So that data can often exist in different places. There are some standards that exist in the language industry around some of the training data that's used for MTE, some of the terminology, for example, that has standards around it that you can use to train. But it's really kind of getting all that together and then getting the initial rollout done where you're really re-engineering some of the workflows that need to exist. Once you've done all that work and once we did that work with our initial clients, it was pretty interesting to sort of be able to sit back and then watch actually the magic happen, right? So you're able to then start to assess the output. We have a team of data scientists that then worked to understand and compare. the output that was coming from our post-edited AI content versus what would have been done by a human translator. And that's how we were able to come up with some of those statistics. And it's a really compelling story that this technology is capable of being able to do this work in a really high quality way. And I think that that really bears some... interesting thoughts around where this is going to be in the next two to three years. I think we're going to see continuous reductions of effort, maybe 35 to 40 % in terms of the human intervention that's required. And it's going to mean a much more automated system. And what does that mean for businesses ultimately? It means that they can experiment more, right? They can experiment in different markets. They can try things with content with less of what you described earlier on as kind of a painful process of getting translation done and internationalization done for content. can do it quicker, faster, and then really find what works for your product, your service in any given market and be able to respond and react to that. Russ Fordyce 🏆 (20:58.53) Well, AI is the topic for, know, AI quality is the topic we're all talking about, whether it's, you know, hallucinations or, you know, just bad performance. That's, that's, that's, you know, it's the gold standard for AI is we've got to get to that quality. That's, that's trustworthy, you know? So how did you get to, you know, these enterprise content teams, which are generally the least trusting people when it comes to quality, because it's got to be right. How did you get them to kind of trust this automatic decision? Paul Danter (21:29.368) Well, at first, what you do is we needed to make sure that we were using languages that we had certified that we were confident that we were going to be able to have a quality output that they were going to be satisfied with. We did some human intervention just to be able to test the output. We often do language quality testing anyway, so just making sure we double down on that effort to be able to prove that the output was going to be satisfactory for what they were trying to do. And then, then the results speak for themselves really. So once we had done all that work and I think it's really about the burn in period of the first sort of 12 weeks as you're working on these projects with clients, getting them to be able to see the output and then being able to kind of have a process by which you can introduce new languages and then do go through the same process of, of, of quality checking and then, and then establishing that it can work. And then AI is also doing a large part of the quality estimation output that we're working on as well within Opal. So. creating a mechanism by which you can really judge the quality of the output that the AI is producing when it post edits a piece of content and giving you kind of an indication, signal of the quality of that work and then routing that to a human in the loop as required is a critical component in that as well. So being able to manage that whole process end to end is really what gives clients confidence in their ability to do that. But as I said before, I think then it comes down to the different types of content that you're working with. So your tolerance for risk varies depending on really what that intended use for that piece of content is. So if you're a running shoe brand and you want to use your little catchphrase, you want to make sure that that brand mark is perfectly translated and has been validated by a human in market, you're still going to always want to do that. But again, if it's a piece of support content that maybe is lower risk and lower impact, if it isn't absolutely perfect, then you need to be able to handle that correctly as well. And I think that workflow process is a little bit of what many people underestimate, but is a really significant part of the work that needs to happen in order to be able to really harness the potential of technology like Opal in translation workflows. Russ Fordyce 🏆 (23:40.942) Well, and what I think you're 100 % right. And what I liked about from your nomination talking about the quality piece is you guys actually estimate the quality before the edit happens. Why did you think that that was a good way of doing it versus kind of that traditional translate and then post edit everything? Paul Danter (23:57.439) Well, yeah, so the flow that we really think about is, you know, how do you do some traditional neural machine translation, then post edit with the AI and then do a quality estimation based on that output. And we felt like that is the most prudent way to be able to check what the AI is doing, being able to validate the output and then being able to score it appropriately so that we can then introduce a human being to be able to make corrections. as possible. And those corrections are very different to what a traditional post-editing piece of work might be for a linguist. So you're not just correcting it, you're also annotating that work so you can point out where in the future the model may need to improve in order to be better. And so you kind of need to be able to put that piece in as well. So doing the quality estimation before it. helps us in terms of being able to do the reinforcement learning that's required to create a better model in the long run. Russ Fordyce 🏆 (25:01.794) Well, and the other thing I really loved from your nomination was the fact that you guys have this kind of continuous edit and continuous improvement loop. So like the AI is giving you feedback. You're going through that feedback sometimes with a linguist, a human in the loop, and then you're feeding that back into the AI to keep learning, which was like, Paul Danter (25:22.153) Yeah, well, I mean, that's a large part of it, right? I think that that's how these models get better. It's how the foundational models are trained also, right? So what we're taking is that sort of concept, and then we're kind of doing it on a level that really represents language for a client. And that's really, really important because most of the broader AI models... are just improving language generally. And what you want to do is be able to improve language that is specific to your business and how again, how you sound. So being able to do those corrections in that way have a really meaningful impact. Russ Fordyce 🏆 (25:58.336) Now, you mentioned it earlier that kind of the language services was kind of running production AI and machine learning and neural networking at scale long before we get into this generative AI stuff that we all now love so much. I think I loved your phrase too that, know, language is the middle name of an LLM. You know, what lessons from this category should other leaders and other verticals be paying attention to as they go and build out these, you know, models for their industries? Paul Danter (26:25.384) That's a great question. think that, like I mentioned earlier on, think what I've found, and it's actually, it's not a question that, or an answer rather, that is particularly profound, but it's actually, I found it to be true as technology has evolved over the last 20 years in this area, which is a piece of technology could be really, really impactful, but it's the operationalization of that technology within a really complex environment that becomes impactful to a business. So there's the potential and then there's the, how is this really going to get used at scale? And, and, and I think that, for me living in this industry for, for, for more than 10 years, you, you sort of see that, the company companies and enterprises that, do well, in harnessing that technology do so by really thinking about where does it fit into our process and. and how can we utilize it in a way and at scale that's gonna work for us. And it sounds really simple and really obvious, but it's actually very true. And if you don't really embrace that, then it's gonna really be just a potentially beneficial thing as opposed to a really beneficial thing to your business. Russ Fordyce 🏆 (27:43.51) Yeah, it's that tsunami, right? Almost any AI project I see is, you're going to get what you ask for. You better be ready for it. Now, you look, I look at AI and I think two years ago, we were dealing with like an eighth grader. seemed like, I think as early as last year, they were probably late high school. Now we're into college postgraduate work. think the AIs are just getting so much better and so much faster. Paul Danter (27:52.005) You better be ready for it, for sure. Russ Fordyce 🏆 (28:13.344) In five years, what does this look like in this multilingual world of yours? Paul Danter (28:19.813) Yeah, I think that you're going to continue to see the models get better. think they're always going to, I think that many companies are going to understand that sounding generic is not going to be right for them. It's not going to differentiate them. So very kind of traditional way of looking at, how do we, how are we perceived as a brand? So being able to take the potential of being able to engage with your, your client, you'll be able to engage with your buyers across the world in language that is meaningful and represents you is still going to be important. So you're still going to want to train it to sound like your brand, but you're going to be able to do that across, you know, 200 plus languages with confidence. And then it's going to be about really how do you manage, how do you manage that, that content? How do you make sure that content at source is correct, is written in a way that is able to be translated effectively. It's by the way, something that's been very important to the translation industry for many years. Actually, there's things at both ends of the translation process which have been really hard to influence, but AI is enabling us to start to influence those things. So one of them is, how do you make content at source better? How do you make the, which is traditionally English, right? How do you make the English created in a way... that it's going to actually be easier to translate and be represented across all these other languages. And I see that the next phase of this in AI is going to be the generative English approach or the source language approach. Let's not even presume it'll be English in the future, but can that source language, paragraph that really describes your product, can it be generated in a way that represents you, that doesn't sound generic and can be easily then adapted for other markets is going to be really critical. And at the other end of it, I actually think the performance of content and the way that that's measured and then fed back into the AI so that it can make improvements potentially automatically, right? this particular marketing campaign with this ad copy is not really working for me. I'm going to just, I'm going to, I'm already seeing that as an AI, I'm already seeing as an agent that the click through rate is terrible. I'm going to rethink about how that copy looks in that language. Paul Danter (30:43.712) republish it out and then see if the performance improves. And I think you're going to see these sort of living systems that are able to kind of make those kind of changes on your behalf, which is going to be incredible, right? Russ Fordyce 🏆 (30:56.142) Yeah, that what you just described sounds awesome to me. Like, I need I need that now. I know there's a lot of people doing automation and advertising, but multilingual, you know, you get in there and you Yeah, you're 100 % right. We say you've been at this, as you said, for more than a decade, and you've you've been at this kind of precipice of of LLMs coming to life, right? If somebody out there building AI, or you're a founder and operator in a Paul Danter (31:00.449) Ha Okay. Russ Fordyce 🏆 (31:25.902) What advice would you give them? What principles would you give them for working in these very brand sensitive regulated workflows? Paul Danter (31:34.496) Well, yeah, I mean, think, well, first of all, by Opal, no, I'm just, so, I mean, that's the, mean, there's, I think what you're gonna see is these expert models that are built around specific tasks. And I'm not just talking about language, I'm just talking generally, right? Where you're gonna be able to harness that trained model to do something for you. So I think that's really important. I'm actually, I think that the, yeah, I buy into this idea that we're gonna soon have a. billion dollar single person company that's operational, right? Where everything's been generated by AI. I think that these, anyone embarking on a new business that has the capability of using tooling like AI to be able to go build their company and to do it with a cost base that's nominal compared to what it would have been five years ago even. It's just so exciting, right? So, but I think what they're need to do is sort of harness these expert areas. of AI and bring them all together in a way that's really meaningful for whatever they're building. And I hope language plays a significant part in that. really passionate about language. I think as a human race, language is a critical part of what makes us who we are and of our cultures. And so I would hate for that to get lost, but this is maybe a chance to make sure that language can exist into the future because we're making more and more things accessible through AI being able to be part of that process of using people's language in a way that's meaningful and I think that's really exciting. Russ Fordyce 🏆 (33:05.228) Yeah, we, you know, in the, in the tech world, we've always talked about the digital divide of not being able to, you know, reach everybody and not, and everybody not participating in technology and the internet. And I think we've largely solved that, you know, or the infrastructure is there to solve it if we want to take advantage of it. Language was another huge divide, right? It's of US companies wanting to talk to other companies and vice versa. And you guys have really seemed to solve that. So congratulations again on the award. This is really, know, outstanding work. mean, from a business operator standpoint, this is just huge. Paul Danter (33:40.233) thank you so much. And again, yeah, we've, as a company, we put a huge amount of investment into our AI ML team, into the tech team that's behind all of this. They've done an outstanding job, not just in really the visionary way that they thought about how this needs to be set up for the future, but then also practically implementing that for our first clients. And they've done an outstanding job in that. It's been a great team effort. Russ Fordyce 🏆 (34:04.386) Well, Paul, thank you so much for joining us and really appreciate the conversation. Paul Danter (34:08.007) Thank you so much, Russ. Appreciate it.