Hello everyone. We are going to start in a couple of minutes.
And welcome. Great to see you too, Danielle. It's been a while. Great. Hi, T.J..
Hello. Okay, so we'll kick things off as soon as I have the presentation on there. And I hope that we have many people joining. It will be an exciting webinar.
There we go. Okay, let's get into it. So first of all, once again, hi, everyone. Welcome. Thank you for joining us for this webinar today. What the context is here?
A little while ago, our own marketing team faced a challenge that we thought a lot of you will recognize. We needed to rebuild Acolad multilingual websites, and pretty quickly, we realized this wasn't just a redesign project or even just a translation project. It became something much bigger. It was a real opportunity for us to rethink how we create and manage and publish those multilingual contents in today's digital landscape.
And that's, of course, because the landscape has genuinely shifted. Marketing teams are being asked to produce more content, publish more often, and now also to adapt AI driven search experiences all at the same time. And so what we recognized, not surprisingly, is that the traditional processes didn't work for us anymore. We had to adapt.
We needed to move faster without giving up the quality, the consistency, and the governance across our languages. And so what made this project especially interesting is that we weren't just building a solution for our customers to think about, we were actually living it ourselves, managing those hundreds of pages across 12 languages, coordinating the stakeholders and the technical teams while keeping the terminology and the brand voice consistent.
So these were exactly the challenges that I'm seeing as we are talking to our customers and them facing it every day. And so today, we wanted to share that experience that our team faced and how we dealt with it directly with you. I'm joined by T.J. He's our creative director, and he was right in the middle of this from a marketing perspective.
He's going to walk us through what the old process looked like, the challenges that they were facing, what has changed, and what surprised him along the way. The goal today is for this to feel like a conversation. We'll be going back and forth rather than presenting at you for the next however many minutes. And so let's get into it.
T.J., maybe a good place to start is by looking at why rebuilding the website was necessary just to kind of anchor this discussion since many of our customers go through this journey themselves of having to think about rebuilding the website. Sure. Yeah, absolutely. As you already mentioned, you know, the landscape has changed a lot over the past few years. Our business has certainly evolved.
Our marketing needs have evolved quite a bit. It's also worth mentioning for those maybe who aren't aware is that, you know, Acolad as we know it today is really, at least partly, the result of a lot of different mergers and acquisitions. So we've had a lot of different brands. We've had a lot of different websites, a lot of different CMSs come through into the marketing team.
And so one of those legacy components was Adobe Experience Manager for CMS. So we were using that for a while. And AEM is really great CMS. It was a good fit for us at one point. But as things began to change, it became less of a good fit for our unique needs.
We found ourselves relying heavily on external dev support. So there was time and cost involved with that. We also found ourselves that we had the need to just be a little bit more agile, you know, even when it came to just basic page authoring, template creation, things that might kinda sound simple like that. We just found ourselves getting a little more slowed down, and we wanted to find something that could allow us a little bit more speed and agility.
So we took a moment, did some research. We, you know, looked into is AEM the best fit for us right now? We checked out a bunch of different CMSs, and ultimately weighing the pros and cons, we actually landed on HubSpot for a new CMS for us. We found that in HubSpot, there is gonna be some opportunities for us to potentially address some of those agility challenges that we had before related to page authoring, translation workflows, etcetera.
Also, there's a little bit more autonomy just in the control of the site. Additionally, and this is a big part too, is that the marketing team was already using HubSpot as a CRM. So we were using it to manage campaigns, email marketing, things like that. And so us making the decision to rebuild using a different CMS using HubSpot actually helped centralize our tech stack and really just create a lot of opportunity for us.
Do you feel like there was anything that you lost along the way? So you were switching from AEM to HubSpot. Was there anything that you felt like was missing as part of that transition or or not so much? It was just purely benefit for you to move to HubSpot?
You know, it was it was more so benefit than us losing something. I think that in in reality, if we're if we're being honest about it, HubSpot or excuse me. AEM is a very, very strong, big marketing platform, and we were underutilizing that. So there's a lot there that we just simply couldn't take advantage of because we're a very small marketing team.
We need to be lean, and bandwidth was literally an issue. So I think that, you know, it would be nice if we could be using AEM still, but that would also mean we need a much bigger team, we need to have a much bigger production setup. There's a lot more to that. So, you know, maybe we're missing out on some opportunity to what AEM has as this very massive CMS for us.
But ultimately, there were so many more payoffs for us to go the HubSpot route. That's just, like I said, it's a better fit for us. It might not be for everybody, but it just works better for our team. Clear.
And out of curiosity, because you described a number of factors which made you change, but when exactly did you realize that that previous process was not scalable? What was the trigger if there was? I think, you know, probably the trigger was the translation workflow. You know, being a company that specializes in translation, we wanted to make sure that we have, you know, a modern, fast, high quality workflow.
And we found ourselves it was just taking a long time. Maybe I can speak to that workflow specifically. So to back up, there'd be three reasons why we would even trigger a workflow in the first place. One would be we have a new page that we're creating, like a new solution offering.
Another option would be maybe we just need to edit an existing page, whether it's a big or small edit. And the third option would be because maybe we have a translation quality issue, and we need to address that on a language variant. So depending on which of those three reasons we would have to trigger the workflow, the flow would look like this. Inside of AEM, we would initiate the translation workflow.
This then gets sent over to our project manager, our dedicated project manager, who then farms out that page to all the different linguists who are involved. In our case, we are and we have been always translating into 12 different languages. So the PM would farm it out to 12 different linguists. Those linguists would do their job.
They would translate the content, send it back to the PM. Once the PM got all those pages complete, then it would send it back over to AEM. The marketing team would get an alert to say that these pages are ready. We would then go into AEM, and we would publish the pages.
So it's a very linear kind of out and in flow, and we knew that that was slowing us down. Ultimately, that took about two weeks for that process to go play out. Oh, wow. Yeah. That is definitely a cumbersome process.
What were the so you described that whole process, but I'm sure that there were particular challenges that were created by having this process. What were some of those challenges? Yeah. Absolutely. So, I mean, two weeks alone is one challenge.
You know, there's a lot of a lot of pressure for us to move faster. You know, one factor that contributed to that two week timeline is if we were to send a page out for translation into the 12 languages, like I said, maybe the French translation was available within two days, but the Portuguese translation maybe took thirteen days. Well, we didn't receive that French translation two days later. We still needed to wait until the Portuguese translation and all other languages are complete in order to get sent back.
So we're always like just very dependent on this just slow flow, all depending on each step along the way. So that big challenge was just the two weeks period. I think that there was several other maybe more hidden or kind of subtle challenges in there too that that we definitely faced as we're as we're working in the CMS. So one, for example, is while these pages are out for translation and in process for, like I said, for us, it was like that two week flow.
Those pages would essentially go into a content freeze. So we, once in a while, found ourselves in the situation where, you know, we have the pressure to launch a page as soon as possible. We get the stakeholder sign off. It's all good. We're happy about it.
We publish it, and we send it off for translations. But maybe two days later, somebody says, well, actually, we need to make another edit. Well, we can't. So we need to hold for now. So it's that content freeze I was referring to.
So it has to go through that whole translation process. Takes about two weeks to come back, then we would make the edit, and then we would do the whole flow again for another two weeks. So again, I mean, I can't emphasize time enough. That was a very big factor here.
You know, there there's other challenges too even when you look at the math. So if we send one page out, comes back, that's 12 pages that come back. But oftentimes we would send packages of 10 pages out, And so what would come back is 10 times 12 languages. That's a 120 pages now that come back into AEM.
So, you know, that's on the marketing team to make sure that we're making we're keep we're tracking this well make sure that those 120 pages are all in sync with that English page. So, you know, there's a little bit of risk that's involved there. It's not only that they are in sync, but you also really, based on what you just described, have to pay immense attention to make sure that that source is correct and that you're not rushing it. So then you end up having bad content not only locked in for two weeks, but also translated incorrectly Yeah.
Throughout that. So that's a that's a that's a big challenge, it sounds like. Absolutely. And, you know, we found that if we were sending, you know, one or two pages or 10 pages out for translation, that always took about two weeks.
So that kind of fourteen day was kind of our sweet spot. So we actually found that, you know, the most optimal way for us to use this this flow was to actually package 10 pages at once because that would still take fourteen days to come back. If we had even more pages than that, this the process would be even slower. And then again, if you're doing the math, you know, if it's 20 pages times 12 languages, that's just so much more like manual work when it comes back to CMS.
Excuse me, when it comes back to AEM for us to manage in the CMS. So there's a lot to manage for both us as the marketing team, and I think probably also the PMs that were helping us would say the same thing. That's for sure. What about any other challenges? Were you able to leverage context, any kind of style guides?
Did you have any issues with how you could stay on brand as part of that workflow in the old process? To some extent, yes. You know, we had certain things set up, like we had TM. We did have a glossary layer that we were able to add, so we were able to contribute to that to help, you know, continuously improve certain translation quality issues over time.
So there was some of that there. And honestly, our quality was good. I think that overall, when we're looking at this old workflow, we would be say that we were happy with the quality. I think that if I'm being picky, I could say that there was, you know, some areas maybe where we could have added a little bit more finesse or maybe it's transcreation that we would need.
And I'm thinking specifically right now in terms of like SEO or GEO. Because we spend a lot of time on the marketing side to research, you know, what is the best keyword for us to rank for or how do we want to position ourselves for certain phrases. And we're doing this on the English side, and when we publish that on the English page, maybe that's working and that's great. But when the translation comes in place, that's where maybe it falls short.
And I'm not saying that the translation quality isn't good, but I think what I'm really trying to say is that the translation quality and the local search intent are not always necessarily the same thing. There's not that clear overlap. So like I said, while that translation quality was good, there was still some room for improvement, maybe most specifically related to, like, SEO and certainly the brand voice. Yeah.
And that's definitely also an important area because when we talk about SEO, geo, it's definitely not just the English. It's definitely not just the way that that you are found. It's actually very, very important that in the translated language, you have a way to connect with your audience and that you are found easily because most local websites, most local languages is the way that people search. So if you are not found, however good your English is, it will not get you there.
So that's definitely something that is worth mentioning. Okay. Any other challenges that might be worth highlighting before we move forward? I mean, I think that that's probably sums up the main challenges that we had.
Okay. So I'm curious when it comes to the audience, because as I described, this is definitely a scenario that a lot of our customers are dealing with, whether it comes to a website translation or optimizing workflows when it comes to the translation process. We have a little poll, and I would like to have that pop up there. What is the biggest challenge in managing your multilingual content today?
Do you mainly see it being the translation quality? Is it the number of languages or the volume that you have to be dealing with? Or is it keeping the consistency and the brand voice? SEO perhaps that we discussed.
Or is it time to publish? What is the biggest challenge that you are facing usually? I see we have translation quality pop up. Why significantly? Which is an important one.
But it's also curious to me because it was one of the subjects that you were mentioning, TJ, even in the old process, you didn't necessarily have problems with, but you rather had problems with the speed as part of it. Of course, it was heavily human involved, and we will talk about the transition to the future workflow as part of this, but that's an interesting one. 50% of the audience has the translation quality as the biggest issue. Then we also see consistency.
Once again, that's a nice one because I think as part of your journey as well, it was the recognition of how important it actually is to build that in upfront. And we'll talk about that a little bit more as well. But certainly an interesting split. Thank you. That's great. So moving on, let's look at how did we actually change the process.
So what were the things that you ended up changing? What was the new workflow like? Sure. So like I said, in our experience, the quality was already pretty good.
We wanted to make sure that we kept that, But we wanted to add, you know, a new element of speed and agility to the process. So ultimately, we moved from that single linear workflow to two new scenarios that we could play out. So option one was just an AI only translation workflow. And the second option was two layers.
So it would first go through AI translation. Second layer was human review. So now, today, if we are publishing a page or making an edit, we ultimately choose which path do we want to take for this content. And we're, you know, most often thinking about this or making the decision, I should say, based on, you know, what type of content is it that we're translating, what languages are we focusing on, you know, and what is the quality level that we expect to return to us.
So we have these two different workflows that are great for, like I said, customized for our fit, and we're able to choose the routes based on that. But how do you actually decide what still needs that full review cycle versus what can move faster, what's the decision based off of? So a few different things. So if we need to be very, very fast and we are okay with accepting some, what I would say, some quality risk, we will maybe go the translate AI translation route only for today because we can get that so fast.
And then we will still have that second level of human review onto it, but maybe we accept that this page is live for a few days in that AI translation only mode. But once that human review comes in, it comes it comes in and overlaps to correct itself. So sometimes we will take that route depending on, you know, the nature of the content. If we are dealing with something that, you know, absolutely needs to be perfect, we will 100% always take the second workflow that involves the human work human in the loop because that is that important for us.
And we know that there is a significant difference when there is still a human involved. So the content, the type definitely matters. I would say it also kind of depends on, like, the pressure that we have to go live, you know, what time frame are we working with. And also, I I hear feedback from certain languages that there's more quality issues in some languages than others.
From what we hear, maybe French is working really well for AI translation only, but, you know, Finnish, for example, or Swedish, know, are historically more challenged languages for us when we do that AI translation only route. So if an area is if we're publishing a page and, you know, we're absolutely focusing on a Finnish market, we wanna make sure that that is a quality layer to it. So we will 100% make sure that there's a human in the loop for that.
Okay. So I see that there are a couple of decision points there. One might be that you are aware that the language doesn't perform as well. And then for those, you might even just suppress having the AI only route and always have the human as part of the process.
And it sounds like there are also the scenarios where you know that that page just cannot tolerate risk. I wonder, what are the kind of pages that you're talking about when you're saying, you know, you know upfront that this cannot have any errors in it? What kind of pages would that be? Sure.
So, you know, I should also say that, you know, because of the industry that we're in, we care a lot about quality. So I believe 100% of our pages ultimately eventually have that human review layer to it so that we make sure that we do have perfect quality. It's just maybe that sometimes there's that window where we live with the AI temporarily before that human review can come into place. But to be more specific and an answer to your question, certain areas that are more regulated will certainly not take that risk to publish an AI only content.
So something like life sciences, industry related topics, legal topics, you know, certain press releases, things like that. We will absolutely make sure that it has that level of human review to ensure that it is 100% correct because we can't take that risk to go live with AI only. Okay. And that actually makes me curious, and you might not have an answer for this.
So it's just, out of curiosity. Since we're talking about pure AI translation and even if it's for a short window of time, it might be the only thing that's live for a couple of days. Did you ever see any kind of translation that, you know, was kind of a funny mistranslation, which wasn't really appropriate? Because we know that AI can hallucinate.
Is that something that the team faced or something that you have recognized at some point? You know, I think that more often than not in the situation where it's like a transition quality issue based on the AI, I think the feedback that we're getting is more grammatical versus, you know, a pure hallucination. And luckily, we don't have any, you know, good examples that went viral as you see some brands. But one maybe that I can think of, and it's not even necessarily that funny, but it's kind of, you know, where did this even come from, is the we had a headline for an interpreting page.
It said something along the lines of, like, interpreting for all of your needs. But in Finnish, I was told the literal translation was interpreting for all modes of transportation. So I don't know where the AI came up with that one. But luckily, we had a reviewer in place to catch that.
They thought it was funny. They told us about this translation. Of course, that was corrected, and so that wasn't an issue for us, thankfully. But, yeah, you know, that that does happen, but it's very infrequent for us. Yeah.
And I think that's maybe the challenging part as well when we talk about relying on AI translation, especially as the content volume is exploding. AI typically reads much more fluent than neural machine translation, as we know. So it would not be uncommon that as you skim through a website and you're just doing like a quick check, oh, yeah, this this reads very well. And then there might be a really strange translation.
Like, maybe the one that you mentioned is easier to spot because it's kind of crazy. But Or maybe or maybe that's a niche market that we should get into in AI know something we don't. Yep. It could be. But definitely, it's just something to be mindful of that because it's more fluent, it's a lot more difficult to just quickly spot, oh yes, this is just completely wrong.
So you actually have to review it a bit more in-depth and you have to pay attention to what you're trying to spot. And I think it will take time for people to really adapt to this kind of content and this kind of a review process. And that's also why for us too, what we are seeing is although AI translation can make processes much faster, as we see in your case as well. Indeed, the human step is still a very important one, just to make sure that nothing strange slips through.
We do see many scenarios in the world where overreliance on AI has caused some pretty bad outcomes. So it's just something to be paying attention to. But talking about content volume. So your volume has also increased over time.
And so I'm wondering, did you end up also growing the team to manage this, or did that workflow change really absorb that growth? 100% the workflow absorbed the growth. If anything, the team has become more lean. We're able to do more than ever before.
The content volume has absolutely increased tremendously since we made the switch. Okay. That's quite positive. So incorporating that AI stuff has really made a big difference, it sounds like. Yeah.
Can you describe because in the old process, you kind of went through the workflow itself. Can you describe what the new work flow actually looks like step by step when you consider the different tools? Yeah. Yeah. So like I said before, there's we have two options, two routes that we can take.
So if we were to take that AI only route, the way we have it set up is our pages are under a parent child relationship, like hierarchy setup. So English being the primary source of truth, that's the parent page, And then all language variants are child pages or like children. So they're essentially language copies of the English page. So today, if we were to make a new page, we would draft it in English first.
And if we were to then send it for translation in that AI only route, we send it out. And about one minute later, it comes back into all 12 languages for us ready to go when we could publish. So, I mean, we're talking one minute compared to what used to be two weeks. So that already was just, you know, a crazy change.
But you know, as we've already covered, sometimes there's a quality risk there. So you know, do we want to take that route and go that fast? Sometimes yes. And that's great because that's I mean, that's lightning speed compared to what we were at before.
But then the other option would be would be would be would take the second workflow. So it would run through AI first. And when it runs through AI, I should also mention that it still references RTM from before. It still references glossaries, and now we also have additional context layers for it.
So it runs through AI, and then it goes out to the human, to the linguist for the human review. And then they send it back, comes back into HubSpot for us, and we can publish. That process now takes about four days. And so, again, four days compared to two weeks before is a huge win for us.
But if we're looking at it even closer, it's even it's even better than that. Because before we were stuck in this very linear step by step waterfall flow and we were dependent on each language, dependent on each linguist, etcetera. But now we have a continuous localization flow. So if our page is getting sent out through the AI and human in the loop workflow, if the French linguist finishes the job in one day, we get that back in one day.
And if the Portuguese page, like I said before, maybe takes a couple days longer, well, then that will come back three days later, four days later. So it's not like an all or nothing thing like it was before, where we would still be dependent on all the languages, and it would take x number of days. It just so happens that right now, on average, it takes about four days to complete all of those pages, all those language variants. But some of them are live even faster than the four days.
That's pretty impressive. That's a huge change, definitely, compared to that fourteen day period. And so with this, would you change anything else? Is there anything that you feel like, I suppose, was the most drastic change beyond the speed?
I would say that, you know, before I mentioned that we would go into this content freeze when pages were out for translation during that, you know, formerly two week period. But that's no longer the case for us at all. So now if we publish the page, we send out for translation, and it is taking that longer route, that four day route of the human interview, we can still make changes to that English level page and then send it out again, and it automatically catches up to itself.
So we're no longer locked out of the page while it's in process. We can just continually making updates to us. So that's a I think that that's a very significant win for us in in terms of efficiency. So not only is the whole process a faster speed, but we're able to just like I said before, like, we work more agile for sure.
Okay. Yeah. That's definitely a huge win as well. What about SEO and GEO? So that's kind of in the context and brand voice and SEO. Are you doing anything different now compared to previously when it comes to this?
Yeah. Yeah. We are. Definitely. So that's where it goes into the different context layers that we have, in which case when it's going through the AI part of it, you know, AI can reference these context layers. So we might have context related to keywords where we're saying certain keywords that we want to, you know, target in different languages similar to a glossary. But even further, we can now optimize it more to our brand voice so we have more consistency there.
And also, we're able to give kind of, like, guidelines to it. So maybe a more a concrete example would be, you know, every page has a page title. That's the title that you see that comes up in a search engine search engine result page, like if you're doing a Google search. You know, the rule of thumb is to keep that restricted to something about, 60 characters because anything longer than that, it gets truncated when it comes up in your display if you're if you're doing the Google search result.
And especially if you're doing a mobile search result, it gets truncated even further. So we're able to add this context layer to give extra guidelines to something as specific as this page title to say that when you translate our 60 character title into German, historically, would come back into, you know, a 120 characters because German's language is just so much longer than English. But we can now have this context layer set up so that when AI is doing the translating, and also the reviewer, they can see this as well, They know that we are, we're asking for that 60 character limit.
So that translation that comes back now fits more strategically with our marketing efforts. So there's a huge benefit of having context layers there. Indeed. And that's actually something that maybe I will add as context.
No pun intended. For the audience, so what we have done there in this process, we've actually incorporated our own solutions for this workflow. So, we have we have our relatively new we launched it at the beginning of the year, self-service LLM first translation product called Lia GO. And in fact, what we are doing here in this workflow in a new one is to incorporate Lia GO via API into this workflow.
So we are plugging it in effectively into the TMS that is being used in the background. And we can feed all this context that TJ was describing, including the character limitation, including the term base, and the voice, the style guide into the go, and that can be sent as part of that LLM translation to the LLM. So it's really using all of these elements for the core translation effort, which makes a huge difference for us. And of course, it's our own technology.
So it just made sense for us to leverage that in our in our own workflow as well. Yeah. So that's a lot of improvement. What I'm wondering about is, did anything surprise you during this project? Was there anything that's beyond this massive gain of speed?
Was there anything that surprised you? Yeah. So, you know, maybe this shouldn't have come as a surprise, but I you know, I'm a I'm a fan of technology and all the changes that are happening. I think there's just a lot of cool things, I get excited about that.
But, you know, one thing that became clear after we implemented this is, yes, we got the speed that we were trying to achieve, but the human review still remains absolutely essential for quality in in the workflow process. So, you know, that was a surprise. Maybe it shouldn't have been, but that's that is a truth that became very evident quickly. I think also, you know, the value of those context layers, as I was referring to before, it surprised me how much they actually did improve the translation quality and in our strategic efforts when it would come back to us.
I knew ahead of time that, you know, we all heard, you know, quality in is quality out, so that that's important. But I think that I underestimated that. We could have invested in that sooner and it could have been a better transition quality right away from the get go. That's something that continues to evolve for us and the those context layers continue to change and get better and better.
And I think that's natural. But if I'm being honest, that probably is something that we could have paid more attention to earlier on in the process so we could have had those quality gains sooner. Makes sense. Thank you. Is there anything about the new process that you're still iterating on or that hasn't fully settled yet?
You know, like I said, the context layers are constantly under iteration. As new as we offer new products, as new solutions come up, we are constantly changing strategies to market those. User patterns change, you know. I mean, even, you know, examples over the past five years is before it was all about SEO, but now, you know, everyone's using ChatGPT to search for things.
So the marketing strategy has had to change over time. So we're not just targeting keywords for a Google search. We're, you know, we're optimizing for generative search engines. We're up you know, for a lot of different things.
So that change that we have to fundamentally make as the marketing team, we're able to continually add that layer to context layers so that that can change with us as we go to make sure that we remain on strategy as we have to shift the time. Makes sense. And definitely, the search behaviors have changed so tremendously. I mean, nowadays, most people end up indeed using any of the LLMs, chat GPT. It could be thought.
It could be whatever Gemini to try and find information about products, companies, destinations. It could be anything, and no longer really just going to Google and doing a traditional search. It's not as much common. They might not even be visiting our site anymore because the whole experience might be in one of those, you know, chat platforms.
And so we wanna make sure that we have the right information for the platform to read. So, yeah, there's constantly change. That's correct. And actually, that's also why I find that multilingual websites are so critical nowadays.
Because if you if they are searching in their own local language, in a chatGPT, for instance, and you don't have your website localized, you will not find anything. You it will not it will simply not pop up. And so it's quite an important angle to bear in mind nowadays and having all that content GE already as well. Okay.
If another marketing team were at the beginning of a similar journey where maybe you were 18, twelve, a few months ago, what would be the first thing that you would tell them to do? How would you tell them to approach the situation? So, you know, specifically for the localization workflow, I would say that, you know, don't treat localization as a handoff. It's not we've published our page in English. We've got the content approved by stakeholders. Okay.
Now, linguist, do your job and pretend it's gonna come back and everything's gonna be perfect. It's not it's not as simple as that. There's much more nuance. There's much more just constraints, whether that's technical constraints, time constraints, budget constraints.
There's just a lot more to it. And so like I said, don't treat localization as a handoff. I think what we found really helpful for us is that we were able to have open communication with our solution specialists, with our PMs, and they were able to listen to us as we are, you know, complaining, saying, this is slow. How do we speed this up?
And they're able to help find a more customized solution for us. So our workflow is very customized around our constraints and around our specific needs. So, you know, keeping an open dialogue with the solution specialists, with the PMs, with the linguists, being clear about what we are trying to achieve has really helped us, and that's something that I would that I would recommend. Yeah. That makes sense.
And certainly, sometimes just being able to collaborate and brainstorm on a challenge that a certain team is having or even recognizing that the process might be subpar for each of the teams that are involved can trigger that change and 100%. And having those open conversations that are just collaborative and proactive to try and solve a challenge together, that is quite an important critical angle. Is there any other part? Is there any other mention around this?
Any lessons learned that you would want to mention before we wrap up? I guess I just can't, you know, say this enough that human in the loop makes a big difference. You know, I I see that when we're looking at that poll, the numbers were clear that people are having translation issues, quality issues. I think that we have this expectation today, or at least I do, that this is gonna be a really fast flow now because we've got AI. That's the magic answer. Right?
And it is in a lot of ways. It does speed things up, but human in the loop still remains vital in order to achieve the quality that we want to achieve. Maybe that's not for everybody, but that is essential for us. Yeah.
And I think that's also really just assessing what is the what is the content as you described. You were talking about the language. You were talking about the particular type of content and where risk can be tolerated. That is certainly true.
You need to look at what are you actually talking about. Because in some cases, you might have and here, I'm kind of moving away from the realm of a website. But let's just say you are translating some internal exchanges, perhaps in Jira tickets or any kind of support system where all you need is for your engineers to be able to comprehend each other. There, you might be okay with a rough AI translation, and you might not need that to be reviewed by humans.
That's okay. But then there might be content types where maybe like a marketing campaign where you might think it's not as high risk because it's a low time frame project, perhaps, like a sales campaign. But it can absolutely be high risk as soon as you put something out there that is offensive or inappropriate for your brand. So it's just about having that governance to really think through what are the risks for a given content type and how do you want to approach that.
From there, you can make the decisions and need to route content appropriately or set up workflows appropriately. But that's a good point. And, yes, human element and human touch can be vital for specific types of content. I fully agree.
Okay. So before opening things up for questions, I just want to pull together a few of the threads that we talked about today. And so if there is one thing that this project reinforced for us is that there is no single piece of technology that solves every localization challenge on its own. AI became an important part of our workflow, but indeed, the real transformation didn't come from just the technology itself.
It came from redesigning the process around it, recognising that HubSpot was a better fit for us compared to AEM, connecting Lia GO into our localization workflow, rethinking where the human expertise adds the most value and when exactly, and then incorporating the relevant context upfront, not only for the English content as well for SEO, for instance, and keyword search, but also for the translated content. And really, a few lessons that stood out that I think will apply beyond our own website is optimizing the source content can be valuable, but indeed looking at what is the context that that you need for the translation and how do you optimize that the best is really valuable as well.
And the AI really delivers its best value when you assess how and where it is appropriate and where do you want or need the human touch to be paired to have that professional review. It's not meant to fully replace it, to be honest. Okay. Let's go into some q and a. Let's see if there are any questions from the audience, and we really hope that this kind of journey is relevant to you.
It's something that you can relate to. Let's see if we have any questions that we might be able to answer. So we had one question, which is, are you using any translation memory prior to sending the contents to the LLM? And so we you might be able to take that, TJ, in fact.
Yes. And the answer is yes, we are. So we have translation memory built up over many years at this point. I mean, that same translation memory that we were utilizing back when we were using AEM was even brought over into this new HubSpot hosted website. So, yes, absolutely using TM.
Perfect. Thank you. We have another question there, which is where do you draw the line between fully automated LLM output and the mandatory human proofreading? For me, it's again speaking for, you know, as an Acolad marketer, I think that it depends on I said this before, but it depends on the content type. It depends on the language priority, and it depends on the speed that we need to achieve.
So it's kind of this this this balance that we need to assess all of those. If quality is of the utmost importance, then absolutely we take the route where it involves the AI, then the human in the loop review. If there was a content edit to a page, and maybe that page originally had already gone through the human in the loop review, but maybe we're changing just one single small line in there and we send that up for translation, maybe we're okay with risking that with AI only because it's, you know, for whatever reason, maybe a noncritical page for us.
But again, it really depends on what that content is, what the language is that we're what we're going into, and what kind of time frame that we're working with. And, again, there's certain languages that we know from our experience that have maybe less success on quick quality from AI versus others. Yeah. That's correct. And what I would also add to that question, it's not sometimes it's also about the tone.
So we found that with some language pairs, as we described, LLMs perform differently. Some languages, it performs better. Some languages, it performs worse. But what we've also recognised, and this was particular to Finnish, I believe, which you already mentioned, was generally not as good.
But for instance, there with the formality, having a more formal translation tone was a challenge for the LLM as well. So there are really these kind of smaller nuances that sometimes as a native speaker, you can really spot when you're when you're reading the translation and when you're when you're trying to evaluate how good it is or whether it's appropriate. But indeed, you really have to think about the language, the domain, and the intent of that particular content that you're translating.
And I think that answers some of the other questions there as well, where we are being asked, you know, when should a human step in, really? And that will depend on your content type, but you do need to look at it. And, frankly, if you are not a language speaker, it will be maybe a little bit difficult for you to be able to look at the translation and say, yeah, this is great because it will look good. It will look good, but it might not actually be good.
We have one more. Let's see. Well, we do have one question that is more particular to the product launches indeed. It is still ultimately the same question. But I suppose when we talk about product launches particularly, and here I will briefly answer, but of course, there is a website element to that as well.
You might think about, of course, a user interface. And there, as TJ described with some buttons or some of the text that is displayed, you have to look at what does fit. So a particular language like a German, that's a really good example with the text expansion, you really have to look at is the is the product designed in a way to accommodate for a very long string. And if it isn't, then you do need to have a human check.
You need to you need to at least validate, does it look good before getting it live ideally? Because otherwise, it can really degrade the user experience if they cannot see what they are supposed to interact with or what they are meant to be able to read. If there is any perspective on your side there, TJ, to add, feel free to. Yeah.
I mean, we have literally seen instances where, you know, a button you expect it to be, like I said before, like three words or so. But when the translation comes back, it would it would require line breaks. And so it doesn't even look like a button anymore at this point. So part of it is maybe more technical setup.
So it's making sure that we have our button component set up so that it's able to accommodate a longer phrase if that is absolutely required. But then there's also that layer of context where we can provide the AI engine or the human reviewer to make sure that we're also not going beyond, you know, x number of character limit that we might set for the button. So it's kind of when I look at it, I think of both of those scenarios. Okay.
Great. And I think we'll take one final question. I see one from Sarah over there. For an update to a page, will you save the previous contents to translation memory as an in context exact match and then just send the new content pieces to the TM and then to the LLM because he's trying to understand how much we can count on the translation memory especially in the thought process. So if I'm understanding the question correctly, the previous content so if we we've sent the page to translation at one point in time, our translation memory gets populated with that old translation.
And if we go back to the page and we make some changes to it, that old translation memory will hold and will stay, and it will the system will automatically identify what has changed and only change that new segment. So we will keep the TM in place. I'm not sure if that's answering the question, if I'm understanding the question. Here, at least from a solutions perspective, looking at this question, what I what I understand it as and maybe where she's leading with it.
If we're talking about an AI only workflow, especially where TM is part of the workflow, we often get this question, does the TM then get populated with the raw AI translations? And does that potentially create a bad match the next time around that might become an ice match? Or perhaps, does the TM only get updated after the human review stuff? This is something that we typically recommend or that we typically follow because then the content can be relied upon the next time around.
Understood. Misinterpret. That makes sense. And also again from, you know, speaking from our experience, we ultimately always do have that human in the loop in the process.
So like I said before, maybe there's a couple day window where we might live with the AI translation only, but the human in the loop is always involved. So in our unique experience, the TM does always stay up to date with the quality that's already been reviewed from Yes. So you never actually update the TM before the human review takes place, I understand. And for you, re-leveraging the same translations.
And, of course, we're talking about here in context exact. So that's typically pretty safe because it's not only matching the segment itself, but also the segment before and after it. It should be the right level of quality to be able to rely on that translation quality generally. Okay. Thank you everyone for joining us.
Thank you for all the questions. We hope that it was an interesting exchange for you. And yeah, thank you TJ for sharing all the expertise and all the journey that you have gone through with this. Yeah, thank you very much.
It was a great opportunity to speak. Thank you. Have a lovely day everybody and enjoy the rest of the week.
.png)