AI翻訳コンテンツを各市場で成果につなげる方法 ― Acclaro のアンドリュー・トーマス氏

Welcome, everyone, to the Emerge AI in Business podcast. Today's guest is Andrew Thomas, Head of Marketing at Aclero. Aclero is a content intelligence and localization company that supports global organizations with multilingual content. Andrew joins us today to discuss why content that is translated accurately can still underperform in some markets and why he sees closing that gap as a matter of content governance rather than translation. He explains how connecting marketing performance data with the localization process lets teams keep adapting content in each market, why brand style guides and terminology need to be built into AI translation tools and where people who understand the local culture still need to step in. Today's episode is sponsored by Eclero. In this episode, we cover why marketing, localization, and marketing operation teams need to share data and work together for AI translation to deliver the outcome each piece of content was created for.

To go deeper on this topic and learn how leading organizations approach AI investment more like a venture portfolio, and why interdisciplinary collaboration is critical to defining the right data for AI success, download our free PDF report beginning with AI at emerge.com.aik1. That's e-m-e-r-j dot com slash a-i-k-1 to download your copy. Now the conversation with Andrew. Andrew, welcome to the Umoge AI and Business Podcast Studio. Hey, thanks for having me. Glad to be here. Here's a puzzle that a lot of global marketing leaders will probably recognize as we go through it. And that's, let's imagine every page went out in all 12 languages on time. Every word was signed off.

And for some reason, Germany converts while Brazil just sits flat and nothing happens there. And if you had to ask someone on the team why this is happening, nobody really knows how to answer you. So as enterprise organizations expand globally, what are the biggest operational challenges they face in terms of maintaining things like consistency, accuracy, and speed across multilingual content workflows? Yeah, great question. It's something our clients struggle with and we help them to solve. And essentially, it comes down to the simple reality that every culture is different, has different values, and people respond to content differently in those cultures. And while AI has been an amazing blessing for scaling up our content operations as marketers, it can sometimes miss cultural nuance. So if you don't have a process that includes humans in the loop that actually understand the culture that you're developing the content for, oftentimes you will see these types of results where the content misses the mark.

So that's why it's so crucial for companies, especially when they've implemented new processes that take advantage of these new capabilities that AI offers, that they've really integrated the right people in that process and the right quality gate checks to ensure that the content delivers the outcome that it was intended to achieve. And really, if you think about it, it's not even a translation or a localization issue. It's quite simply a content governance issue, but with that global perspective. Yeah, absolutely. And I want us to get into the idea of content missing the mark, because I feel like that sounds like an expensive miss. And many of our listeners might think, you know what, it's very normal for our content to land in a specific country better than another, or to have better conversion in, let's say, Germany or wherever. So if we have a VP of global marketing asking their team today,

is our Japanese landing page doing what it's meant to? Not is it perfectly translated, is it doing what it's meant to do? Who in the building should be able to answer that question, and what will they base that answer on? Well, it's a really great question, and I think it probably depends on the company's structure. Ideally, the marketing team can actually answer that question themselves and that they have a good relationship with whoever owns localization within their company or with any vendors who might be providing that service to them so that they can dig into these issues around content. But, you know, you raise a good point. People should not just accept that content performs better in one market than another. I don't believe that that's ever true. I think that content that is adapted for each market can be just as performant in every market. There's underlying exceptions to that rule, of course, depending on what you're offering. You know, there might be issues with the actual offer itself, the product or service that you're trying to provide, and that might be an indication why the content's not performing.

But all other things being equal, the content should have similar performance globally if it is dialed in correctly for the local market. So in the perfect world, a marketing team has a great relationship with whoever owns localization as the process within their organization. They are able to integrate the data from both the translation process with the marketing content performance and feed that back to the localization team so that those teams can actually make changes to the content to increase its performance. If you actually connect all of the data that your typical MOPS team has with the underlying systems that are managing the overall localization process, you can actually create this positive feedback loop where the localization teams can essentially continue to adapt the content in market to maximize its performance.

Not unlike marketing teams today do for things like SEO. where SEO teams continue to modify content until the content is performing correctly. It the exact same idea It just applied to the global content and working with different people I see that And I want us to really get into this because from what you saying and from the research that we have done it seems like solving this pain point has not really been a priority for a lot of organizations, not because the pain point is not there, but because they need to change processes that they don't even know about. So before we get into what the solution to all of this is, I also just want to take the listeners through a bit of painting a picture. So when a newly hired salesperson in Indonesia tells HQ, I need content right now, what happens inside a typical company and how long does that take? It obviously varies greatly based on industry and market expansion. But a typical story is a company wants to grow globally.

they decide, oh, we think that there is some real business opportunity in this other country. And so they hire a local salesperson in that country just to see what the local interest might be. And as you say, as soon as that salesperson is hired, one of the first requests that they're going to have is the content that they need to support their sales efforts. And it's also led by their idea of what they need, which isn't necessarily what the customers necessarily want, right? It's all just driven by the salesperson's perspective. So they'll make a typical request to headquarters, and if there is an existing translation process, that process will be followed. The timing of that is greatly reduced these days thanks to AI translation, but there's still some waiting period. They get their content, and then they start to use it in country and see what works and what doesn't work. And quite honestly, what they will often find is either they ask for the wrong kind of content or the content isn't performing the way that they want.

And then inevitably they come back with more requests and you just end up eating a lot of cycles with this kind of back and forth. And, you know, what I would advocate as a more modern approach, thanks to the power of AI translation, is to actually test localized content in market before you deploy local teams, you know, salespeople, support people, that sort of thing, to see how much market interest there is for your product. Because the process is so much faster and cheaper than it's ever been, it's really worth giving it a shot to just put the content out there to see if you have potential customer interest. And based on that content's performance, you can then make decisions on where you want to make investments for market expansion. And more importantly, because you haven't wasted too much money or time on producing the content just through straight AI translation, where you do see interest, you can now decide to bring humans in the loop with, you know, working with companies like Eclaro to improve that content and actually make it fit for purpose for that market expansion.

It's a much more efficient way to test global growth, honestly. Yeah, so it's no longer just about translating the same marketing campaign into each language that would be appropriate for the country. It's making each ad or each information piece appropriate for the value systems of that receipt, not just the language. And you've mentioned now how AI is helping here. So what AI capabilities are having the greatest impact on multilingual enterprise workflows today? And how do we work those into our processes? Yeah, I mean, most of it's probably already familiar with people who are using, you know, any kind of off-the-shelf foundational model today, whether it's ChatGPT or Cloud or Gemini, you name it, translation is baked into all of these models. So you can request content in practically any business-related language today that most companies would need and get that content quite quickly. Now, that obviously has made a huge impact on the translation industry in terms of speed and scale and, yes, also cost.

But where it causes issues is actually very similar to things that I think a lot of marketers who might be listening to this podcast are already familiar with when it relates to brand. As soon as you have all the power of the AI at your fingertips, and it's not just at your fingertips, but it's literally at the fingertips of everybody in your company, you have the rise of shadow use of these tools outside of whatever processes you might already have in place. You might have people creating, quote-unquote, marketing content without ever going through a marketing team that doesn't bring in your brand guidelines and, you know, your preferred terminology and, you know, any other aspects for marketing that you would care about. The same exact same thing is true for translation into other languages. If you have a localization process, if you have a localization team, what they are currently struggling with today is shadow localization, where lots of teams, including us fellow marketers, might be using these AI tools to translate content without going through these other official processes.

And, of course, the reason why they go through, the reason why people do that is because they don't want to be slowed down and they don't want to be hamstrung by something that feels overly bureaucratic or might take too long. But the resulting downside to that is, of course, a lack of quality control, a lack of governance, and that can oftentimes result in a lack of content performance. But if the groups that need to work together actually align, there are ways to actually embed, just like you can embed branding guidelines into your foundational models for the rest of the company to use. You can similarly embed localization data to ensure that translations are performed with a higher quality that are aligned to your company's specific terminology or your company's previous translations. There's lots of tools that we use that actually can help enforce that. So rather than trying to fight against shadow localization or shadow marketing, in my opinion, it's much better to subvert it by embedding your own cloud skill or your own specialized copilot agent,

you know, whatever tools you might be using that actually are plugged into the back end systems that the localization teams are using today to ensure quality of translation. When you put those things together then you get all of the power of AI with much less of the risk And you will then begin to see your content become much more performant in the various markets that you want to put it out in I want us to just go back to the brand part because that's something very interesting that you mentioned. We can use AI to keep our brand sounding the same even in all the different languages. And I think that's something that some of our listeners might be fearing because we see a lot of AI-generated marketing content in whatever language you speak. And we've seen that a lot of it looks so similar. So if they want to trust the tool and keeping to brand, keeping to the brand tone, the brand everything, what do they need to do to ensure that that tool knows their brand? Yeah, really great question. And it's one we're really familiar with on the localization side.

One of the first things we do when we have a new client is we actually get our hands on their existing brand style guide, and we convert it into actual, you know, AI instructions. And we ensure that that's then enforced on every single translation that moves through the system. And obviously, there's varying degrees of that, just as you would for any source content. You are very strict about those rules when it comes to specific marketing content, But you still even want brand guidelines to be adhered to for other types of content that a company might produce. You know, your help content, your product documentation. But it won't be quite as stringent, right? And it won't be quite as heavily enforced because that's not the content's intent. And those other groups might have other rules that they need followed. So the reality is you might have a really good handle on that for your own language. But as soon as you hand it over for translation, people just seem to forget that, no, in fact, you need to translate, quite literally, these types of instructions and these terms and

these brands for all these other languages to ensure that when you take advantage of AI translation, it doesn't inadvertently introduce these types of mistakes. And I go a step further and say that, and this is all about AI governance, I go a step further and say it's really important to provide these types of instructions, even if the AI models produce quality translation. If you don't give it your specific company's brand terminology in those languages, what's going to happen is the content that it produces is going to be reduced to whatever the average is. You know, the way that LLMs work, they reproduce the most obvious patterns, which means by default you're always going to get a generic response or a generic generation if you do not provide it the right instructions initially. And a lot of end users in their processes don't think about that

when they click the magic translate button in whatever app they have. So if you haven't wired it to your specific instructions, you're going to get a very generic translation, and it's going to sound like everybody else in the market. So the same way that marketers think about sounding unique in their own language, it's just as important in all the other languages that you're serving content in. And if you don't take the time to translate those style guides and those kinds of instructions and then embed that into your AI translation process, that's the result that you will get. And it's likely one of the reasons why the content isn't performing well. So it sounds like the results in the output is very much dependent on the instruction. But even with that, I mean, I'm sure that we still have instances. And you mentioned in the conversation that sometimes a human is needed in the loop. So where does AI still miss that cultural side so that a person needs to step in and get involved? At Eclero, we like to think about it as, you know, getting involved in the beginning of the process with the prompt and the data setup that we've just talked about,

but then also getting involved at that last mile when, you know, AI has taken you 80% of the way there or 90% of the way there, and you just need to polish it that much further to really land in market. And usually what that involves, particularly with marketing, is what we call cultural nuance. And it's really, truly understanding what the culture values in that particular market. And the reason why AI is never going to be perfect at this, at least in my opinion, you know, maybe we'll get there one day, but I think it'll be quite some time before we do, is quite simply that culture continually evolves, continually shifts. It's a moving target. It's never the same. And, you know, these models are trained on preexisting data. So by their inherent nature, they're always going to be generating content for a culture from the past.

And, you know, this isn't the most fun example, but it is a great example. Google Translate prior to COVID, if you were to give it a translation, it would about that use the word mask in that translation. It would translate the word mask as if you were talking about a beauty product. And that's because the volume of content that it was trained on, that was the obvious choice for that word. But then COVID happened and suddenly everybody's world changed overnight. the culture shifted dramatically and we weren't talking about beauty products when we said the word mask. We were literally talking about medical masks, you know, to protect ourselves. And it got the translation wrong. That's not the model's fault. It's the fact that the world changed overnight and without updating the models, it was producing wrong output. If you have a human in that process, they easily spotted that issue and could correct that issue. But even better, if you had a human involved in the prompt in the beginning that could help you fine-tune and set up the data in the beginning,

that problem never would have happened in the first place. So long story short, if you have professional linguists and people that understand the culture involved in the beginning, helping you set up your linguistic data for your AI translation process, and you can easily bring in humans in the loop at the end where more cultural nuance is needed. And by the way you can determine that based on that content performance metrics as long as you got those teams plugged in with your MOPs team and they can talk to each other and coordinate data then you can ensure that the content is always performant and is actually delivering what you want and really what the customers want. Yeah, I like it. I see the picture in my mind so clearly of, you know, the human feeds the AI, the context, the style guides, the general feel and tone of the organization. The AI cooks up whatever needs to be put out in different languages, but taking into consideration the different culture and the requirements in terms of how formal

and how straight up it should be, and then having a human again at the end to fine-tune it when needed. Andrew, we've got time for one more here. So I wanted you to think of a head of global content who's been told two things this quarter. One, get more out of AI, but two, don't embarrass us anyway. And one path is to put everything behind review and give up speed, which is obviously not ideal. And the other is to let it run and just hope that the regional numbers will hold. And I think this is hoping and not strategizing. So for a leader in that seat, what practical steps built a multilingual workflow that's scalable, reliable, and ultimately sustainable over time? I'm not trying to sell a Claro, but I think this is a time where people should be really looking at who they can partner with that has the technical competency to help them with setting up their AI translation correctly so that they can reasonably let it go and let it do everything while also providing them with the right gates in the process.

to determine when to bring humans in the loop. And those humans can either come from the company or they can come from agencies like ourselves. It doesn't really matter. It's more important, though, that you have it sorted out in the process so that you have governance baked in. Because you have to understand that the way AI works, you have to make the assumption that at some point there will be a failure. If you do not plan for that failure, it's going to eventually bite you. But if you do plan for it and can understand how to quickly resolve it, then you can basically let the AI run free and have the peace of mind knowing that if an issue should arise, you can handle it correctly. One of the things we do at Aclaro is we actually use AI to gatekeep AI. We can actually check the output of an AI translation and run it through our own automated quality checks using other AI models and probabilistically determine when we need to have a human in the loop and when we don't.

And that gives companies a lot of reassurance that they can go full throttle with the AI translation, knowing that we will catch whenever we need to bring a human in the loop and ensure that that content doesn't embarrass them, cause some sort of PR nightmare, that sort of thing. That's exactly why you want to partner with companies that are good at that exact skill. Don't try to do it yourself. You can reach that state eventually, and we can help companies get there. But if you don't already have those skills natively in-house, your best bet is to find the right partner to make sure that it happens. Yeah, it's no longer just about translating. It's about making sure that the partner has some accountability in what that output looks like and what drive you get from it. Andrew, before I let you go, VP of Global Marketing takes two things from this into their next planning cycle. Here is what I heard. I heard that a landing page that is accurate in 12 languages can still underperform in eight of them. And we need to judge each market's content by engagement and conversion,

and not just the general ones, but what it was meant to drive to, what was the goal from the beginning. And starting with an assessment of what each content type must achieve in each market. Then let the content owners set the quality thresholds and decide what strips on AI alone and what goes to an in-country expert and take it from there. Does that sound right? Yeah, that's exactly right. And I like to call it quality of outcome. For too long, our industry has frankly been focused on broad linguistic grammatical quality. And frankly, that's not what matters. What matters at the end of the day is that the content that you spend time and money creating generates the right return that you are looking for as a company. and that return on that investment is completely different based on the type of content that you're creating and the markets and the audiences that you're delivering it to but at the end of the day you want a return and so our job in this industry has shifted to in helping ensure you that quality of outcome that's how i like to summarize it perfect andrew thank you so much for your time

today it's been a great conversation great thank you for having me Wrapping up today's episode, let's look at the three key takeaways from our conversation with Andrew. First, content that is translated accurately can still underperform, so leaders should judge each market's content by the engagement and conversion it was created to drive. Second, AI translation defaults to generic average sounding output unless brand style diets, terminology, and past translations are built into the tool every team uses. And finally, people who understand the local culture are still needed at the start to set up the instructions and data that the AI works from, and also at the end to refine contents that automated quality checks flag for review. In this episode, we covered why marketing, localization, and marketing operation teams need to share data and work together for AI translation.

To go deeper on this topic and learn how leading organizations approach AI investment more like a venture portfolio, and why interdisciplinary collaboration is critical to defining the right data for AI success, download our free PDF report, beginningwithai at emerge.com slash aik1. That's emerj.com slash aik1 to download your copy. For further executive level analysis and to join our network of leaders delivering workflow impact with AI, visit emerge.com. On behalf of the team at Emerge, we'll see you on the next episode. Thank you.