#279 Why Phrase Doubles Down on a Platform Strategy with CEO Georg Ell

20:44SlatorPod Host:of much more personalized localization. I feel like 2025 was largely a story

20:53Georg Ell:of unrealized ambition around AI. So we still had in the, you know, I mean, I think not even just the localization industry, I think the world at large, AI was still over-promising and under-delivering for much of 2024 and 2025 in terms of production-grade enterprise use cases. There are lots of good, lots of, well, lots of experimentation, lots of not good experimentation in the sense that much of it failed, right? There's a famous MIT study and we saw a lot of that play out too. So how do I feel about 26? I think more optimistic, actually. I think we're starting to see the models improve, maybe not so much directly specifically on translation.

21:39Georg Ell:I think that's still relatively flat. But in all the ways you can augment around translation, I think there's now a lot more capability coming on stream. To the extent that, and test me on this in the next two, three months, I think we're going to change our website and you're going to see the words translation and localization much less on our website. I've looked at all our competitor websites. They all say translation and localization everywhere. It's plastered everywhere. We're going to stop talking so much about translation and localization and about different things. And I think 26 is a year where some of this AI-driven personalization becomes more achievable and therefore we'll start to see production-grade use cases, production scale as well.

22:24SlatorPod Host:You also mentioned briefly the kind of multimodal and phrase studio. I believe you launched this like late May 2025. Any fives? What's been the experience so far? Take up, you know, problems, challenges?

22:36Georg Ell:This is an opportunity for some humility, actually. So we launched it with great excitement. We actually established a huge amount of interest. Like the interest blew us away.

22:51Georg Ell:What we saw over the subsequent months was lots of people saying, it's great, but. So lots of people were looking forward to experimenting, looking forward to trying and buying it, actually. Like we built a big pipeline, as they say. And then just lots of folks said, it's great, but not quite ready. And there was a few different reasons for that, many of which we've kind of addressed. So it was more or less a standalone product. It's now actually very integrated into the platform. So you can use phrase language AI across it. It integrates with translation memories and term bases and glossaries.

23:24Georg Ell:and it integrates with our kind of translation management system, like modules. There are new things we're going to launch in the next couple of weeks that are platform-wide, so they work across every element, including studio to do with personalization. So we integrated with Google Drive, which people want. There was like a ton of work we've done around quality and so on. And there was two phases to that improvement. The first phase was done before Christmas. The second phase will launch in the next three to four weeks, I think, thereabouts. So probably in a few weeks' time, we'll start reaching out to customers around early adopter programs for that next phase of improvements in quality there.

24:08Georg Ell:We're pretty excited. We've spent enough time talking to customers about the reasons they didn't buy it just yet that we think we've understood that and hopefully cracked it. So, yeah, definitely an opportunity for a bit of humility. but still like very excited about the potential. And I think a lot of people want us to succeed because people, like there are a lot of point solutions out there around video. There really isn't one that's integrated into a broader platform. So there really isn't one that would allow you to have like a range of different engines, access to terminology, translation memories, quality evaluation, human in the loop, you know, agentic work, Like all of the things that we wrap around any content type, there really isn't a solution that's baked in or can take advantage of all of that in the same way that Studio can.

25:02SlatorPod Host:Almost nothing. I mean, the UI of some of these is just atrocious in terms of not producing like the AI dubs or the subtitle. All of this is phenomenal. but the actual language part, if you wanted to tweak a 10 language version of whatever you're doing, it's like, okay, I have some type of box and I some manually kind of type and there's, you know, like even like something like a translation memory is in almost none of these or in none of these. So I think there's big opportunity. Is there like one part, one section of the market that has found it super useful? Like, can you give me a use case of somebody that, that, well, okay, that solves my problem now.

25:42Georg Ell:One I like is learning and development, which is producing a huge amount of content. And I think L &D actually, and just enablement, let's use that phrase instead, I think is an area where localization broadly can add a huge amount of business value, which might actually segue into something I want to talk about a little bit. But I'll give you an example. An engineering firm that uses our platform told me two years ago that only just under 2 % of their content was localized outside of English. And this is a company with 90 or 100 ,000 employees in lots of countries around the world, meaning that they would need to employ English language speaking field engineers to do, you know, the repairs in every country around the world.

26:27Georg Ell:So when you localize that enablement content, you have a bunch of business impacts, right? You reduce the time to hire, you reduce the cost to hire, you reduce the average salary potentially. You improve the fix right first time of that field service engineer because he or she is now operating in their local language and they're learning in their local language. And so if you bring multimodality into that as well, then again, you're improving all of those metrics that do with ramp and quality. So I think learning or enablement works internally. It works with field teams and office teams and it works with partners and channel teams and particularly as the world is changing so fast around us all the time the ability for an organization to ramp its people on whatever the latest thing is in their space becomes actually really important to survival let alone you know competitive advantage so i think a great opportunity for localization generally speaking and in particular something like video i want to talk briefly about voice

27:27SlatorPod Host:kind of voice AI, very hot over the past three to six months. And we hosted a very small voice AI meetup at our office in London. There's a kind of a meeting area where we had a few people there. It was very interesting. There was almost no overlap with the localization crowd as it were. And even though one of the demos by one of the companies called Newphonic, they had a kind of a very small model and the demo they gave was basically like live interpreting, but they didn't think of it that way. They just thought of it as kind of a voice use case. Why do you think this voice AI kind of startup ecosystem, community, whatever you want to call it, isn't really connecting with the localization crowd yet?

28:13SlatorPod Host:I'm not sure if I'm butchering this maybe, but that's the feeling I'm getting. I think our industry has been, you know, kind of a bit inward looking for a long time.

28:25Georg Ell:And like, it's funny, we were saying earlier how time flies. Everyone knows I stood up on stage as this sort of new kid four years ago. I feel a little more embedded in the industry now. But, you know, from when I first joined and now, we're still all talking about how we don't do a good job about explaining our value, right, as an industry. It's the same old thing. So I do want to talk about that. And I think this links to it in that we're still too often solving the same old problem and we're still too often measuring our success in the same old ways. And so we're not thinking about multimodality in different formats.

29:03Georg Ell:We're not as an industry thinking about reaching out and forming partnerships. with voice companies or generative video companies or marketing and advertising agencies in the way that we should be. Because we're so focused myopically on reducing the cost per word and reducing the spend on humans as a proportion of the total. The industry's... One of the challenges at the moment in localization is that every lock team is being asked to spend less. and that highlights the problem, right? The problem is that they're too often seen as cost centers, so just reduce the costs. And I think what localization teams are experiencing, like very sadly, is that when you have reduced all the costs, what's left is the team and then you reduce that.

29:50Georg Ell:So we as an industry, we need to tell a different story. And I mentioned some of it earlier. Language is in every interaction, right? It's like it's our input output. Um, so when you get that right, it's not like good enough, right? Um, it's about the, can, the impact that you want to have. So is it hiring people? Is it, um, basket completion rate? Is it, so I built a tool, like, like I'll give you an example. We'll put this on our website, uh, before long. Um, in my vibe coding in the evenings between the hours of sort of 11 PM and 4 AM, I built a value tool which will research a company, and then the AI will determine that based on that company, here are the 30 different KPIs that have nothing to do with localization metrics in which language might play a role.

30:42Georg Ell:So if you're a mining company, the AI will come up with things for mining. If you're a food and beverage, right, same thing. And then it will put together the maths, the calculus, the algorithm for what the difference might be of good language, bad language, right, and put a number on it. And the numbers are huge because like tiny improvements of many factors across a company. And so then you can work with that to help a localization team actually talk about the value that language has across a company more broadly than just, you know, is our budget reducing Eurovia? And then it becomes really interesting because suddenly they've got a right to talk to the people team.

31:24Georg Ell:They've got a right to talk to support and finance and legal and all these other folks. And I think what we as an industry need to do is come and bring solutions to those specific problems, those KPIs. Like, what is it that a platform can do that addresses that particular problem? And how and like it shouldn't require someone to log into an old fashioned TMS dashboard and be an expert in all the drop down menus. Like, that's not going to be how these things work. It needs to be much more native. So we have this concept of the spine and the edge, right? Something like phrase provides a central vertebra spine that has workflow, quality, assets, context layers, which are critically important.

32:07Georg Ell:Human workflows, approvals, chains, but business gets done out on the edge.

32:12SlatorPod Host:And these things need to connect to each other very, very quickly and easily so that people get all of that value, but they get it where they need it and when they need it.

32:20Georg Ell:and not going through some convoluted submission to a localization team who then come back in days and all that stuff. Okay.

32:30SlatorPod Host:Why does the industry, broadly speaking, still have this messaging problem? Why do we lead with the capability and the technical stuff and not with the thing that it solves for a customer? Is it to shyness? Is it like a certain kind of the inward-looking component? should just somebody be bold enough to just go ahead and build a website that basically literally doesn't talk about how the sausage is made, but the outcome, the positive outcome for the client, all these KPIs you just mentioned, basket completion rates, etc. Are we just too shy or many in the industry too shy or too reluctant or scared?

33:12SlatorPod Host:I don't know. I mean, it seems like an easy thing to do in a sense. You have a website, you signal just the things you're solving, the things you can improve and become, it's, yeah, it's less of a, you know, the good old, like the cost center conversation. No, no, you're a growth lever. You're helping a company, you know, hitting all these things. And then the spend is less important.

33:33Georg Ell:I think it's everything. It's everything you said. It's all of those, some combination of all of those things. And, you know, there are people in our industry that are not shy. And there are wonderful people that are. And there's lots and lots of people in our industry who have not had a tremendous, they haven't just had a lot of business education. Like they're linguists and they're enthusiasts. I was in an industry that often people work in because they love it.

34:06Georg Ell:And we're in an era where sometimes in lots of different ways you have to kind of disrupt yourself. and I think that's scary for people. And I think also a lot of localization teams are relatively thinly resourced and so there's a sort of sense of perpetually being only just keeping your head above water and therefore is there enough time for innovation and it's probably often felt like there just wasn't. It's lots of different things, isn't it? But I do think, look, we, I, my company, myself, have a responsibility here to our customers, to the industry. And I really think of it as an industry responsibility.

34:47Georg Ell:At Phrase, we'd like to play our part. I mean, we're not sort of messianic about it, but we want to play a role in helping. And I think we can do better.

35:01SlatorPod Host:The disruption part, I guess we fundamentally wouldn't have to disrupt the actual way we work. It's a lot of the messaging, right? And the existing customers that are happy with a sophisticated solution aren't going to walk away because you're trying to signal to all the people, all this vast blue ocean of people that have absolutely no idea about what localization is, but are looking for some type of tool or some type of AI readiness that a platform like Phrase would actually give them. So it's like, why is there this kind of fear that you're alienating your existing relatively small group of core localization advocates instead of signaling to the vast blue ocean of people that are looking for some AI tool and signaling to these guys?

35:47Georg Ell:A lot of companies are looking for an AI win, right? And, you know, we sort of, maybe that's one way to think about the conversation we were having earlier, which is that these engineers are trying to give the company an AI win. And what we need to do is get the lock teams to say, hey, we can give you the AI win. And it's like, actually good. So that confidence to come forward and be in the spotlight a little bit. And, you know, definitely to be clear, at Phrase, we want the LSIs to not just exist but thrive. We want lock teams to exist and thrive. To be clear, not every software competitor of ours wants those two things.

36:24Georg Ell:Like there are definitely software competitors of ours. I won't name them because it wouldn't be fair. But the CEOs have said to me very clearly, our ambition is to like destroy those audiences. So at Phrase, I just want to be very clear. We want to support those audiences. Here's an example, though, of the challenge. I was talking to the CEO of a medium-sized LSI, roughly 50 million of turnover, to give you a sense. And this person said to me, the problem is my senior people all think human first. And I said, we can help with that. We can help build the flows and show them. And he said, that would be great.

36:58Georg Ell:It would make such a difference to our margin. Cool. Let's connect people up. There was a call with the CTO. The CTO was lacking enthusiasm. and it's like stalling, right? So I'll go back to that CEO and say, look, I think it's going to take you driving this to make a change. And there's quite a bit of that in LSI's of different sizes. There's quite a bit of that inside some lock teams and some enterprises. There's a sort of defensive crouch because, look, I mean, I'm not, like things are changing around us so fast and it is a little scary. like because people don't understand where it's going they don't know what it means for them and you just see this slightly defensive crouch posture from lots of people um it's understandable but it's it's not safe you know like when you're running away from the lion like defensive crouch is not it's not this you know like to make run up the tree yeah yeah yeah run faster you know I mean, that's it.

38:01Georg Ell:So, yeah, I mean, like, it's incredibly exciting. Like, the things we can do with technology now are unbelievable and the things that we've delivered in phrase and we're about to. We actually have a launch coming up in, like, 10 days, which is why I keep saying we're about to because there's some really cool stuff there. And it is about moving things out of, you know, the lab and into production and it is about quality and confidence and scalability and all that. But yeah, I'm like super excited about what we're doing. But I think we all have to acknowledge that there's a lot of kind of uncertainty out there.

38:36SlatorPod Host:One area where maybe progress has slowed down a little bit is just general translation, like AI translation quality. Would you agree with that? I mean, obviously, we're at a very, very, very high level already, but I don't think we're seeing massive jumps anymore. Would you agree to that? Do you see that in practice as well?

38:54Georg Ell:I think it's true in terms of like the foundation model. I think where there's still like lots of improvements is all the stuff that you can add around that. So context is a really interesting one. People are building really good ways of bringing context in and then measuring the quality out in various ways and then using AI to do post editing so that the percentage that has to go to human review does continue to fall. And I think we are seeing that. So the customers that are using our platform in the most sophisticated way, they're releasing all of the capabilities and not just a subset. They are absolutely seeing kind of the percentage of human review falling.

39:35Georg Ell:And that is enabling them to do like more language pairs, more content types, more volume. And the nice thing is like the volume and the cost of doing it don't scale linearly. So they spend a bit more with us. They spend a lot less on humans. They get a lot more content out. And so I think that kind of experimentation moving into production is happening. But you're right. I think the foundation model is straightforwardly on their own, not a massive kind of standalone quality improvement.

40:00SlatorPod Host:You're based near London, right? I am, yeah. Yeah, I want to talk about London kind of as an AI hub because I don't know why, but X just started showing me literally in the last 48 hours. Maybe I must have clicked on something, but I've seen this huge buzz around London as an AI hub generally. And there's all these meetups and I'm seeing people queuing around the corner. And we've had our little meetup and SlidoCon London is shaping up to be very well attended, etc. What are your thoughts around London kind of coming in again as an AI hub?

40:34Georg Ell:Yeah, I mean, I think there's one specific thing happening there, which I'll touch on. But I think in general, London has always been the leading kind of venture capital location in Europe. Paris is probably next, particularly around AI. But London has always been up there. Google has big offices. Anthropic has offices. OpenAI has a small presence, but they're just recently committed to growing. And I think the government has generally been very supportive of AI. And there's some really smart people from the tech ecosystem that are advising government as well. So I think there's a very good tech ecosystem in London that's very mature.

41:21And outside of San Francisco, then New York, then London, probably that's the order I would guess.

41:28Georg Ell:I think very specifically, I read something myself this morning, which may partially explain it, which is that since Anthropic fell out with the Department of War recently, the mayor of London wrote a letter to Dario, the CEO of Anthropic, inviting him to relocate to London. And that letter was leaked to the Times. So it may be that X has sort of picked up on some of this buzz and is amplifying it.

41:54SlatorPod Host:Looks like it. Looks like it. Yeah. I mean, yeah, I mean, it's good for you. I mean, you guys are there. We're there. So I guess.

42:03Georg Ell:DeepMind is here. Demis Hassabis, based in London, I think, still.

42:09SlatorPod Host:So let's move to the Outlook portion of this podcast, which usually closes it. So let me, you know, what do you see in 2026 or in March? You said that, you know, 2025 was a lot of expectations that might have not been met. Now some of these expectations are being met, which then, you know, triggers some of the uncertainty, I guess, also. But, yeah, what are your thoughts around the rest of the year?

42:30Georg Ell:I think that the event horizon is definitely narrower, and in part because all of these models are changing the way that not just companies like us, how the products work, but also how we do the work. And I think all companies are going to experience that at different stages. I think that phrase, we're probably experiencing that sooner than most for a few reasons. One is we're small enough with 300 people that things can happen quickly. Friends of mine who work for big institutions, they might want to play with Claude at work, but they're not allowed to. There's a lot more guardrails. A company like us, we've very quickly pivoted to providing security, privacy, guardrails, policy and training.

43:15Georg Ell:And, I mean, we were doing all of that with ChatGPT from the very early days, but the recent model developments with Opus 4.6 in particular sees us concentrate a lot more on Claude. and just people around the business are doing the most unbelievable things. And what's interesting about that, it's kind of fractal, and you've probably experienced it yourself. If you sit down with Claude and you start building, before you know it, you've had three more ideas. And you start building those, and each of those kicks off three more ideas each. So nine. So you're like, it goes one, three, nine, 27. It very quickly spirals out.

43:46Georg Ell:So one of the interesting things for companies generally over the next couple of years is going to be how the nature of information work is, I think we are at that cusp where it's going to fundamentally change and very quickly. And so how companies figure out the implications of that, what does it mean for, you know, roles? I mean, someone asked me a good question the other day, but if I'm a PM using this, it's now very easy for me to kind of step on the toes of engineering or QA. It's like, what should we do? Should we stay in our swim lanes or not? So we have to redefine the swim lanes. Everyone becomes a bit more full-stack everything.

44:24Georg Ell:So if you're a PM, you're a full-stack PM. If you're an engineer, you're a full-stack engineer. If you're in marketing, you're a full-stack marketeer. These are the changes that are going to come to the world at large. And I think because of our size, because as a company, we're quite innovative, forward-thinking. We embrace these things. We want to be on the leading edge. We want to learn by doing. That's a phrase we're using a lot, no pun intended. and so we're experiencing that I think other companies will too at different rates so that's that I think the implications of that and how it plays out in technology at large and the product I think we're going to see different ways of interacting with our spine you know enabling customers to pull the capabilities of that spine into different places that they work in a more fluid way than they've been able to do in the past.

45:15Georg Ell:I think we want to open up the value that we bring beyond it being kind of curated through a lock team, but the lock team will still remain in control. I think we move away from the concept of translation and localization is our industry to something more around, we've got to still work out the sort of the right language to use, but I think it's like content delivery, content adaptation.

45:38SlatorPod Host:Language solutions. We call it Language hallucinations and AI. That's what we're signaling right now. I know it's hard, but like.

45:45Georg Ell:Yeah, yeah. And we're trying to work out how we talk about that in a way. You know, one thing that's hard is everyone was talking about TMS for far too long, and I cheered when you guys pioneered LSI instead. But then, of course, as soon as everyone becomes LSI, then you want to stand out again. So we'll probably continue to use the language LSI for others, but find an even slightly tweaked way of positioning ourselves in front of that.

46:09SlatorPod Host:You're an LTP. You're a language technology platform. Let the LSIs be the LSI. I recently read, I think, Milengo, an LSI, kind of sent out an official press release that they're now redefining themselves as an LSI, away from an LSP to an LSI. So it's getting traction. It is. And it makes more sense just now that we've been used to this kind of framework for a year. It just makes more sense to speak about it. Yeah. Lots of solutions, and you integrate them.

46:38Georg Ell:No, I think it has really caught on. I think you all did an amazing job. I told you that at the time, and I think it has helped reframe things. So, yeah, so you're right, LTP. Sorry, that's the problem with going to bed at 2 o 'clock in the morning. The words get tumbled sometimes.

46:56SlatorPod Host:The fractality took you to like 4 a.m. in the morning.

46:58Georg Ell:Yeah.

46:59SlatorPod Host:So maybe then it's about also deciding which problem, which fractal you're going to follow, even as a company. Like there's like a hundred ideas and like you could explore all of them, but, you know, you've got to be disciplined.

47:11Georg Ell:A hundred percent. A hundred percent true. And look, it's going back to the build and buy thing. This is also the problem that internal teams are going to have. So someone gets excited because they do a little demo of a bit of, you know, quick translation. But then they go down the rabbit hole. And this is, if you want to go down the rabbit hole, you know, it's deep. It's very deep. So I think this is what CFOs need to pay attention to because right around their companies, people are going to be building not just in language, but in all sorts of areas. They're going to be building things that they shouldn't be building as well as things they should.

47:49Georg Ell:And that's not even controversial. Of course they are. There's no perfect decision-making framework. And the question is going to be, should they really be rebuilding building blocks or should they take LTPs like us and others, build on top of the APIs, build on top of the MCPs, and build on top of the ecosystem and take that business to the next level? That's the interesting one.

48:15SlatorPod Host:Let me just exploit this for one more second. I mean, the question is, what's really the point? Like you want to save 10, 20, 30K a year on a specific tool because what? Because you now want to bring everything in-house and run it on some AI that you build in-house and that you will eventually need to maintain instead of focusing on your core value as a business? It doesn't sound very sensible, does it?

48:41Georg Ell:It doesn't.

48:41SlatorPod Host:It really does not.

48:43Georg Ell:I think there's two drivers. One is cost pretty consistently. And I think the calculus there is myopic. People look at a software spend and they forget the total cost of ownership or the salaries and the long-term commitment. And they also forget the risk, right? The sort of key person dependency that often arises as a result of this. The other aspect that I've heard people talk about is sort of control and speed. And that is where I think the LSIs do need to be very awake to that because if they're not able to match the speed and the kind of reporting and the control that the customers want, then it does become something that the customers want to take back away from the LSIs.

49:35Georg Ell:So we're trying to work with people to be a bit more tech forward and agile. And I think that's something customers do want.

49:44SlatorPod Host:All right. Well, another fascinating episode. Round three, looking forward to round four in, I don't know, 2028, 2027. Let's see. Thank you so much, Giro.

49:52Georg Ell:Before that, otherwise it'd just be avatars.

49:55SlatorPod Host:Yes. All right. Thank you so much. Thanks for taking the time.