Localization Leaders: Meet Booking.com’s Paula Manzur

Season 3, Episode 4: Why Linguists Are Essential to the AI Language Revolution

In this episode of the Localization Leaders Podcast, we’re joined by Paula Manzur, the Language AI Strategy Lead at Booking.com. Paula takes us through her unique journey from linguist to AI strategist, sharing how she transitioned into the AI field and her current role in shaping Booking.com’s language strategy with AI.

From her early career in translation to embracing machine translation and large language models, Paula offers valuable insights into the intersection of language and technology. Tune in as we explore the future of localization, AI-driven language tech, and practical advice for businesses looking to scale their global operations.

Watch the full interview below:

Can you tell us a little bit about where you’re from, how you first became involved with the language industry, and what led you to where you are today at Booking.com? 

Well, I actually started in the language industry as a translator myself. I’m an Argentinian Spanish speaker and I always had a passion for the English language. One day at home, I decided to translate a short story. That’s when I realized, “This is what I want to do.”  

I studied translation, got my degree, and not many people know this, but I’m also an interpreter. After a few years of translating user manuals for home appliances or being pressured by deadlines, I — like many other folks in the industry — got very curious about the other side of those tasks and all the mechanisms behind the deadlines. 

My localization career really kicked off as a localization lead and engineer on the agency side at a big LSP. That’s where I got closer and closer to language technology, like CAT tools, TMS, quality assurance tools, and machine translation. 

Eventually, I discovered that machine translation was one of my passions, thanks to a former colleague, Jon Ritzdorf, who is also a professor. Now, of course, we also have neural machine translation and LLMs for the task of translation. 

And that’s what my role is about at Booking.com. I’m part of a really big localization team, and I’m responsible for designing the strategy for how we use automated translations and for which content types. So I focus mainly on my passion, language technology. 

When you started out as a linguist, did you ever expect to end up really falling in love with the tech side of localization and actually making that your career? 

Not really, to be honest. When I started as a translator, I just did it because I love the English language but I really didn’t want to change the world. But now, I feel like I can contribute with something small. For example, I like working towards the bias of machine translation; LLMs are full of biases. But honestly, no, I didn’t expect to have a role so close to AI. Mainly because I’m not a machine learning engineer or a computational linguist, which is maybe what I should have studied. 

I think that the roles that I have are how we bridge the gap, really. There’s the application world, and then there’s the research world – and they’re not very well connected sometimes.  

Yes, it’s so important to have creative thinkers be a part of AI language tech moving forward. All the biases aren’t going away and AI is not going away. So, let’s at least improve it. 

Honestly, it wasn’t a smooth transition to AI because you need to prove yourself. I might not be a computational linguist or an engineer, but at the same time, I have an understanding of how these systems work because I study them and work with them all the time. And I work with both engineers and linguists. 

You need to make a little extra effort to explain the why to engineers. Why is this important? Why do we need this? Why is consistent, reliable data important? It’s rewarding to work in this field. 

Can you share some more insights into how your team at Booking.com is leveraging AI in its language strategy? Have you had any specific tools or initiatives that you feel have really made an impact? 

Booking.com has been leveraging AI in its language strategy for many years now because we trained our own MT engines in the past. We don’t maintain them anymore, we use an external MT provider, but we built a very big MT cache. We reuse a lot of machine translations, and that’s very good for savings.  

But in relation to LLMs and what I like to call the “AI agentic hype,” we have an objective this year to embed AI in localization. 

A very important aspect for Booking is experimentation. We’re putting a lot of experiments out there with Gen AI.  Even before an experiment goes live, it goes through an AI council. 

We need to go through a lot of approvals, questions. The AI council reviews the idea for the use case, and then the idea is reviewed from a legal and security perspective. Once all that is approved, the experiment goes live. But first, it needs to be proven valid or impactful in English. 

So even before it becomes localized or multilingual, it needs to work live in English. 

Our source language is British English. If it works in English, then we expand to other languages—but not by generating multilingual content, we localize using either automated or hybrid translation. 

The experiments are very important, but we always evaluate large language models (LLMs) using linguists.

We actually send a human evaluation; they are still the golden reference for us. Before anything goes live for localization, it needs to go through human evaluation. 

So that leads me to my next question. How do you see the relationship between language and AI evolving? And where do you see the role of humans fitting into it? 

Yeah, that’s a very good question actually. I’ve been hearing a lot about keeping humans at the core or humans in the loop. 

But really, if you’re going to keep linguists just doing the same types of edits—changing a comma for a full stop or straight quotes for curly quotes—you’re not utilizing the skills that linguists have, their creativity. There’s a way we can automate all that. 

If you’re a company that has the capacity to build an LLM or train one—or you’re a company that wants to buy from providers like TAUS or ModelFront—you can automate and use quality estimation and automated post-editing so linguists can then focus on more nuanced elements like cultural nuances, or quality and consistency. 

That’s where I really see the future: with linguists and AI focusing on what makes sense. And I think it will be up to all of us to determine what that “makes sense” really means, right? Because I don’t think it’s decided or set in stone. We still need to define that. 

I know you mentioned that what you’re doing now is still very much localizing, even with AI. Do you see things shifting more towards multilingual content creation, or do you think localization will always remain relevant? 

Yeah, that’s a great question that I also have at the moment. It’s an open question. I think for multilingual creation to happen, you really need to take legal aspects into consideration because you might have different content. 

If you publish raw AI or multilingual AI directly without any human review, then you might end up having different content across languages, and it can lead to brand inconsistencies. 

I don’t think we’re quite there yet. We still have a source, and like I said, first experiments go live in English. If they’re proven successful, then we localize. 

I’m really eager to start experimenting with multilingual generation instead of asking an LLM to provide a translation. But we’re still not quite there, I would say. It might be the future, but only after we go through all the legal aspects of having potentially different content across languages. At the same time, content is already more personalized than before and AI helps a lot with that. So maybe we already have different content out there. 

We know that many companies are using AI to help them scale their localization efforts. What do you think is the biggest mistake companies make when trying to leverage AI, and how can they avoid it? 

Honestly, I think the first mistake that always comes to my mind is treating localization as an afterthought. I think we still haven’t learned that lesson yet. Whenever content is created, it has to be created with the consideration that it might be, or will be, localized or become global. 

We need to get closer to the content creation process—actually, not just close; we need to be part of it. That’s something we need to change. 

Now, with Gen AI, you can create content together with humans or copywriters. For example, in UX, it could be a little riskier, but creating content with Gen AI will result in a massive explosion of content. And we also need to understand that it’s not a human creating that content. So how will that affect translation? Will it be done by AI or a human? 

Another mistake is always seeing localization as a cost center instead of a growth driver. That mindset needs to change. We need to report more on the revenue we bring or how we contribute to business success, rather than just reporting how many words we translate per year. It’s time to move away from that. 

One way to do that is to track the revenue we bring, if possible—probably through experimentation. For example, there was one experiment done here at Booking a few years ago that was a blackout experiment. 

We turned off a few languages on the platform for a certain period. I can’t share specifics because it’s confidential, but it was mind-blowing. 

You could see the impact of not having those languages live, it was very significant. The Holy Grail is proving the ROI of localization—the value it brings—regardless of whether you use automated translation or a hybrid model.

That’s what we need to report more on: progress through the growth we bring, user feedback, and how the customer experiences the platform. Not just minor accuracy errors. 

I appreciate that traditional QA is valuable, but we don’t currently have a way to assess quality based on the customer experience. We don’t have typologies for that, and I think that’s a big mistake. As an industry, we’re still not there. It’s up to each professional or company to decide and actually do it. 

Everywhere you go, you hear about AI—on LinkedIn, at conferences, on this podcast. It can be overwhelming. Do you have any favorite resources or recommendations for professionals to stay up to date with AI and their craft? 

I can share what works for me. If you’re a professional in a company, it’s hard to stay up to date because you’re in such a practical mode—working, experimenting, implementing strategy. 

The best way I’ve found is to attend at least one or two conferences a year. These are events where we all gather in localization. You can really immerse yourself, discuss topics with colleagues, and network. Technology changes fast. In six months, we might be discussing the same topics with different tech or angles. 

If you can’t attend in person, find something online—a webinar. What I’ve noticed is that now you often need to pay for those, which isn’t great. That’s why resources like this podcast or other free webinars are really useful. You learn from professionals, from people. 

That’s where I learn the most—through people, not papers. 

Of course, I read research, but discussions are where I get the most value. 

Great, Paula. Let’s wrap up with a few rapid-fire questions, just for fun. What’s your favorite language? 

It used to be English, but now it’s Argentinian Spanish. I miss speaking my own language. I live abroad, so now it’s Argentinian Spanish. 

What’s your favorite localization tool? 

I think it’s Google Workspace—Sheets, Docs. I’ve tried many tools, but nothing beats a good old-fashioned Google Sheet for staying organized. 

Where’s your favorite place you’ve ever traveled to? 

I’d say it’s Pipa, a city in the north of Brazil. It’s really beautiful, a paradise. 

What’s the best localization advice you’ve ever received? 

Not everything needs to be Shakespeare. I always remember that when discussing quality—what’s good enough? 

What’s your localization nightmare? 

The definition of quality and the thresholds. Also, having to speak the language of the business. If you’re not speaking with another localization peer, you need to translate your own thoughts into business terms. It’s ironic but necessary. 

Who is your localization role model? 

My mentor, Eva Gross. She’s the Director of Technical Program Management for Localization at Toast in the US. I really admire her—she’s knowledgeable and always finds time to help people. 

Lastly, which brand is your localization crush? 

“Crush” is a bit of a stretch. But as an expat in the Netherlands, I’d say ABN Amro. Their product localization from Dutch to English is very good. I also admire Uber—the experience is seamless in both English and Spanish. I don’t know how ABN Amro localizes, but I do know a bit about Uber. Both are good—and as a user, that’s what matters. 

Thank you so much, Paula.

Explore other episodes of the Localization Leaders Podcast for more valuable insights and conversations.