Just because you can doesn’t mean you should: Setting up an AI Localization Strategy

You are too academic…

I worked at a company where my American teammates told me that I was too “academic.” At first, I was a tad offended as it was said in a way that implied I was being too theoretical, or asking why too much, which was supposed to be bad. The other day, I was revisiting the Culture Map book by Erin Meyers, and came across this:

Source: Graphic summary by Reading Graphics.

This explained why my American counterparts perceived me as too “academic.” Europeans tend to be principles-first and deductive thinkers, while Americans and other Anglo-Saxon cultures may look for application before theory, and focus on the how rather than the why.

Honestly, I felt better after revisiting this, as I thought there was something wrong with my thinking when it was only a cultural difference. But like it or not, I tend to think about the why of what I do, or what our industry does.

Lately, I’ve been reflecting a lot on AI strategies for localization, and I’ve realized that we often don’t ask why nearly **as frequently as we should. We focus on the how because we want quick results to show progress and signal innovation. Sometimes, it is better to spend more time on the why and set the foundations for a strong long-term strategy than jump into action without a clear path or set of objectives.

Which brings me to how this article came to be, inspired by a New Yorker Cartoon by the artist Lars Kenseth:

Source: Condé Nast Store and archive. Cartoon by Lars Kenseth.

As funny as it is, this cartoon is also very real. It encapsulates one of the main challenges of recent years in the localization industry. Many localization programs are forced to prioritize cost-saving strategies over other initiatives, highly motivated by the AI hype. With the new capabilities brought by AI, it is only expected that executives and stakeholders outside of the localization circles (especially those working in procurement) show curiosity about how AI can be leveraged for cost optimization and reduced time to market, among other potential benefits.

In my experience, localization teams are at the forefront of AI adoption in enterprises of all sizes, as translation is one of the most widespread use cases for Large Language Models (LLMs), and because localization teams have been using AI in one shape or another since Neural Machine Translation (NMT) engines became mainstream in our industry almost a decade ago. So localization has become the playground for AI experimentation for some enterprises, to see if AI can produce ROI. But just because we can, does it mean we should… every time?

The why

When setting up an AI strategy, it’s essential to define a very clear vision, objectives, and a realistic timeline. To draft a strong vision for your AI strategy, you can apply the Socratic method by using questioning and critical thinking to define a solid why. You can also use this method to negotiate with your stakeholders and reach a consensus.

If your goal for an AI strategy is to reduce costs, ask yourself why you want to reduce costs:

• Why do I want to prove my value in my company through cost reduction? Are there other ways to demonstrate value and impact with my localization program beyond cost reduction?

• Why is cost reduction being asked of me, and what expectations are tied to this request?

© Alfredo Dagli Orti/Shutterstock.com

• Why do we focus on increasing company margins — who benefits from that, and who faces the consequences? Could there be a different approach that creates value for everyone?

• Why is the company struggling financially, and how could cost reduction keep localization efforts afloat during recovery?

• Why might cost reduction allow us to localize more assets and more languages, making content more accessible? Could some of the savings be reinvested into formats we currently overlook, such as video or image content?

• Why does the hype and stakeholder pressure push us toward cost reduction, and how can I push back by asking questions and showing better alternatives?

• Why do we assume AI is the best or only way to reduce costs? What other options could be more efficient?

Your localization why should also be aligned with your company’s mission and vision. You want to make sure that you’re supporting the overall business goals of your company by providing support through the localization function. Once you have clarity on the why, and if you’re aligned with the values that drive that decision, you can move on and think about how to be strategic about implementing AI in your workflows.

The how

Now that you know why you want to implement AI, let’s get tactical. It is vital to approach AI implementation as a project to be carried out to completion, rather than a vague strategy for which no one is accountable. To prepare your program for success, you can start with three easy steps:

• Create a dedicated cross-functional team of experts to carry out the plan and make them accountable for progress.

• Set clear goals assigned to measurable KPIs to track progress and success over time.

• Set a clear timeline and schedule monthly or quarterly check-ins to ensure the project is progressing as expected. Remember to celebrate each milestone you reach (for motivation and job satisfaction) and then assess what went well, and what could be improved to keep iterating your process over time.

  
  

Now, there are certain aspects to consider during your AI strategy journey:

• Your data

Data governance is essential for an effective AI strategy. If you’re planning to customize NMT engines or use Retrieval Augmented Generation (RAG) combined with an LLM, you need to have your data clean and shiny. Before embarking on a customization journey, make sure you have all your data in good shape. Your TMX files can be checked to make sure that the terminology is properly implemented, tagging is consistent, style guide choices are consistent (for example, formal vs informal register is consistently applied), and inclusive language guidelines are followed, among others. Also, make sure you have legal clearance to leverage that data for AI training.

• Your source content

Too often, we forget about the immense impact of the source text during the translation process. If your source content is poorly created or completely disregards localization best practices, you might encounter issues when trying to automate your workflows. Also, some content types are more suitable for AI workflows than others. Always think about the source text when planning your strategy.

• Your content prioritization strategy

At the same time, a deep understanding of your source content will guide your content prioritization strategy. You can create a classification system for your source content based on your requirements and needs. Some criteria that could go into a content prioritization matrix are: a) level of visibility of the content; b) level of complexity (general content vs specialized content); c) impact on the end users; d) potential legal, political or health-related implications of the correct understanding of the content; e) expected outcome of the content; f) level of creativity and craft involved; g) level of contextual implications; h) media where it goes (text, video, audio, etc.), and many more. The key here is to identify how much of a cost/time benefit you will achieve by implementing AI at the right/expected quality level for a particular piece of content; that is, when and where it is worth it. There is no such thing as one-size-fits-all for AI implementation, and that approach can actually bring undesirable results.

• Your language prioritization

Not all markets have the same weight for a brand, so understanding where and how to invest wisely is key. Also, not all languages perform equally well with the different AI technologies out there, and performance can vary depending on the content type and the task that the AI is expected to execute (translation, quality evaluation, post-editing, and more). Understanding how all these factors align and come together will help you guide your strategy more efficiently, minimizing risks and hiccups.

• Your definition of quality

To have a clear understanding of how your AI models perform on your content type and your language combinations, it is important to conduct a quality evaluation assessment first. Quality is always a slippery topic in our industry, and much has been said and debated. AI has further disrupted the concept of translation quality, with conversations around good enough or fit-for-purpose quality, or quality measured on the outcome of the content (conversion rates, click-through rates, etc.), and not only the linguistic quality per se. On top of that, there’s the topic of quality metrics and frameworks, which adds to the complexity of this matter. Our Think Tank Member Marta has extensively tackled quality metrics in her previous blog posts (you can find them here).

I have created a quality framework that includes three elements that, in my opinion, are part of the current quality equation:

1. LQA for human translations: Having a good understanding of the quality delivered by your vendors is a key aspect of your vendor management strategy, and most mature localization programs have it implemented and use it as a KPI during the QBRs. These metrics are internal and only relevant for localization teams and their vendors.

   
   

   

2. AI QE for machine-translated or MTPE outputs: If you’re using MT or AI for localization, you should also be familiar with automated metrics such as BLEU, COMET, chrF, TER, BERT, and all the other usual suspects. Automated metrics are useful to compare the quality of different engines or quality improvements over time if you’re conducting model customization and fine-tuning. More recently, AI QE has also been used to filter segments that are considered good by the machine and send only segments under a certain threshold for post-editing. This is the next step of automation, and still not widely adopted across the board.

3. International User Experience: This is where things get interesting. To break the silo of localization teams, it is important to correlate what we do with other metrics that are relevant for other teams, such as marketing or product. Collaborating with user research, product, and marketing teams to understand how language and cultural subtleties impact the overall user experience and content performance is a great way to break that silo and start creating impact and value for your company. Leverage your communication and cultural knowledge to take your company’s globalization strategy to the next level.

   
   

   

With these elements and data in mind, you can start painting a quality picture that makes sense for your program, and see what role language specialists and AI play in it.

💡💡💡One lesson I learned over the years is that you must listen to your language specialists and UX experts. If they say that an NMT engine or AI solution is not providing good results for a specific language pair and content type, they mean it. So carefully listen to them and act accordingly.💡💡💡

Forcing bad MT/AI into localization workflows can have very negative consequences on your brand, but also ethical consequences, as your post-editors will be under-compensated for their real effort, and demoralized to have to deal with poor language production.

• Your content pipelines

  Once you clearly understand your content, your language combinations, and your definition of quality, you can start being strategic about your content pipelines and workflows. You don’t need to have a content pipeline for all your content (meaning, everything has to be TEP or everything has to be MTPE). In fact, mature localization programs leverage the different options and make the most out of each approach to create a comprehensive, efficient, and tailored strategy for their content. Below, you can find an example of different workflows that you can implement in your program:

  
  

  • Translation, Editing, and Proofreading (TEP): This is the most premium translation quality that can be applied to high-stakes (like legal documentation, or content that can impact the safety of the end users), high-visibility, or extremely creative content.

  • Translation and Editing (TE): This is the most common human workflow in the industry right now, applied to a wide variety of content where human quality, accuracy, and fluency are prioritized.

  • Translation (T): A one-step, human workflow is applied when there are time-to-market or budget constraints. The output will still be of human quality, but with room for potential oversights that can be, to some extent, mitigated with automated checks, such as spellchecker or terminology adherence checks.

  • Machine Translation Post Editing (MTPE): This method has become mainstream in the past few years (and has already surpassed the demand for TEP services, according to research companies such as CSA), and it combines machine translation outputs with human post-editing. MTPE is usually applied for cost reduction purposes and quicker turnaround times. The final quality will inevitably differ from the fully human workflows, as the cognitive process to generate the translations is different. In MTPE workflows, the machine leads, and the human supervises, while in a fully human workflow, it’s the other way round.

  • MTPE with AI QE: This approach is a combination of MTPE with further automation. With a quality estimation model, the segments are automatically evaluated, and only those under a certain threshold will be sent to post-editing, while others will be published without human oversight.

  • Fully Automated MT: Here, the output is fully generated by an NMT engine or LLM-based translation model without human oversight. This approach has been applied for user-generated content, such as reviews in digital stores, or real-time communication between customer support agents and consumers.

    
    

• Your humans

Last but definitely not least, you can never forget about the humans who are impacted by your AI strategy. There are, at least, three core groups that come to mind:

• End users: Will your AI strategy enhance the user experience, or impoverish it with lower-quality multilingual content? Will it make the experience more accessible and inclusive, or not? These are questions to be asked, thoroughly considered, and potentially measured before and after an AI strategy is implemented. We’ve already seen examples of how an AI-first strategy negatively impacted the user experience; I am thinking about Klarna’s lesson learned.

• Stakeholders: Your AI strategy doesn’t happen in a vacuum. Change management is one of the hardest parts of any transformation. Maintaining your team informed at all times, providing them with a clear roadmap and set of goals, listening to their feedback, and understanding the needs and expectations of everyone involved is key. Clear communication and training are an intrinsic part of this journey. Make sure all your team members and other stakeholders have a foundational understanding of what AI means for localization, the pros and cons, and the benefits and limitations. Create a shared and realistic vision that inspires the team, and does not create anxiety and demotivation.

• Experts: Think about how AI will impact your experts, including translators, editors, transcreators, testers, SEO specialists, culture specialists, localization engineers, accessibility experts, DEI experts, project managers, and subject matter experts of any kind. Implementing AI in your workflows can have both positive and negative consequences for your talent, so you need to analyze the risks and mitigate any potential damage. For example, if your AI strategy pushes rates down for your talent, they might stop working with you, or get stuck with poor working conditions, which will dissuade them from caring about your brand and your users. Every action has its consequences; make sure you don’t have any blind spots.

  
  

The what (you want to do with AI)

Now that you have clarity on your AI strategy and the steps that you need to follow to be successful in this journey, it’s time to think about what you want to do with AI. There is no right or wrong, as every program and every team has its own unique set of circumstances. However, how you decide to leverage AI will tell a story about your program and the value that you offer.

Strategy vs Commoditization

At the beginning of the year, I wrote an article about the localization value spectrum. Different localization programs show their value differently. As I see it, there are three ways to look at it:

• Localization is positioned as a cost center, and the ultimate goal is to find ways to reduce costs to a minimum, leveraging technology and process engineering. In this case, localization is seen as a commodity.

• Localization is positioned as a strategic function to grow the company, where a hybrid approach is applied to leverage technology for specific content types and use cases where it adds value and makes sense. This is a hybrid approach between the cost center and the revenue enabler options.

• Localization is positioned as a revenue enabler function, and it is considered an integral part of the go-to-market strategy of an enterprise. Here, localization is seen as a strategic function whose ultimate goal is to grow the company and focus on providing an excellent user experience for everyone.

Source: LinkedIn article by the author: New Year, New Motivation, Renewed Energy!

You can leverage AI in any of these points of the spectrum, but where you put the emphasis will dictate how your program is perceived. In my experience, most mature localization programs are leaning towards the hybrid approach to show expertise and innovation, but without diminishing the strategic value of localization.

Automation vs Augmentation

Do you want to use AI to only automate, or to augment? This is a question that you have to ask yourself when creating an AI strategy. The current wave of GenAI technology is stochastic by default, so these solutions are not necessarily the best to automate workflows, as compared to other, more deterministic technologies like Python scripts or regular expressions. Instead of putting your efforts into automating, you can think about more creative ways to augment your localization workflows with AI.

For example, providing tools to your experts to work more efficiently, and not necessarily through post-editing. You can create chatbots that contain all your answers to questions related to your style guides or terminology so that your translators don’t have to waste time searching for specific answers in that +100-page style guide multiple times. You can also use AI to detect potential risks related to sensitive topics (political, cultural, etc.) and inclusive language. Not to replace the humans, but to augment what they can do and increase quality.

The sky’s the limit when it comes to potential use cases for AI that augment and not only automate. Think strategically. Don’t do the same with less; do more with the same. Reimagine your program to become more strategic. Now you have the tools to make it happen.

Moving forward

The ball is on your court now. Setting up an AI strategy comes with great responsibility. It is up to you to think about how you want to use that power. Perhaps you will decide to become more strategic and use your cultural and communication intelligence, paired with your technology expertise, to expand your program’s reach, and not make it shrink. Perhaps you can use AI to create a positive impact in the industry and on the people around the world who are using your products, watching your content, and leveraging your services. And next time you ask why you’re doing this, you will have an answer that makes you feel proud. Good luck on this journey!