How to build a culturally aware multilingual AI strategy
How to build a culturally aware multilingual AI strategy
By Enrique Dans | Published: 2025-12-03 15:18:00 | Source: Fast Company – technology
In the race to deploy big language models and generative AI across global markets, many companies assume that the “English model it translates” is sufficient. But if you’re a US executive preparing to expand into Asia, Europe, the Middle East, or Africa, that assumption may be your biggest blind spot. In those regions, language isn’t just packaging detail: it’s culture, norms, values, and business logic all rolled into one. If your AI doesn’t code-switch, it won’t. It merely performs poorly; it may misinterpret, misalign or serve the new market
The multilingual and cultural gap in LLMs
The main models are still mostly trained on English language corpora, and this creates a double disadvantage when deployed in other languages. For example, A He studies found that non-English, morphologically complex languages often require 3 to 5 times more symbols (and thus cost and computation) per unit of text than English.
Another research paper is about it 1.5 billion people speak low-resource languages at higher cost and worse performance When using dominant English-focused models
The result is that a model that works well for American users may falter in India, the Gulf, or Southeast Asia, not because the business problem is more difficult, but because the system lacks the cultural and linguistic infrastructure needed to deal with it.
A noteworthy regional example
takes Mistral SabaLaunched by a French company Mistral AI As model 24B parameter Specially designed for Arabic and South Asian languages (Tamil, Malayalam, etc.) Mistral touts that SABA “provides more accurate and relevant responses than models five times its size” when used in those areas. But it also performs poorly on English standards. Here’s the point: context matters more than size. The model may be smaller in size but it is more nimble for its location
For a US company entering the MENA (Middle East and North Africa) or South Asian market, this means that your “global” AI strategy will not be global unless it respects local languages, idioms, regulations and context.
Token costs, language bias, and global ROI
From a business perspective, the technical details of coding are important. Recent article He points out that inference costs for Chinese can be twice as much as for English, while for languages like Shan or Burmese, token inflation can be as much as 15 times.
This means that if your model uses an English-based encoding and you publish in non-English markets, the cost of use will go up dramatically or the quality will decrease because you’re shrinking tokens. Because your training group was largely English-centric, your “base model” may lack semantic depth in other languages.
If we add cultural and normative differences to the mix: dialect, references, business practices, cultural assumptions, and so on, we arrive at a completely different competitive set: not “were we accurate” but “were we relevant.”
Why does it matter to executives about expanding abroad?
If you lead a US company or scale a startup into international markets, here are three implications:
- Model selection is not one-size-fits-all: You may need a regional model or specialized fine-tuning layer, not just the largest English model you can license
- The cost structure will vary by language and regionToken inflation and crypto inefficiency mean that your unit cost in non-English markets is likely to be higher, unless you plan for it.
- Brand risk and user experience are cultural: A chatbot that misunderstands basic local context (e.g., religious calendar, local expressions, regulatory norms) will erode trust faster than a slower response.
How to build a culturally aware multilingual AI strategy
For executives ready to sell, serve, and operate in global markets, here are the practical steps:
- Map languages and markets as top-tier features. Before you choose your largest model, list your markets, languages, local standards, and business priorities. If Arabic, Hindi, Malay, or Thai are important, they should not be treated as “translations”, but rather as first-class use cases.
- Consider regional models or co-deployment. A model like the Mistral Saba may handle Arabic content less cheaply, more accurately, and more authentically than a more finely tuned generic English model.
- Plan for nominal cost inflation. is used Price comparison tools. The US model might cost $X per million tokens, but if the deployment is Turkish or Thai, the actual cost could be 2X or more.
- It’s not just about fine-tuning the language, but also the culture and business logic. Local datasets should not only include language, but should capture regional context: regulations, trade customs, idioms, and risk frameworks.
- Design for active switching and evaluation. Don’t assume that your global model will behave locally. Deploy beta tests, evaluate local standards, test user acceptance, and include local management in your rollout.
The larger ethical and strategic lens
When AI models privilege English and Anglophone norms, we risk reinforcing cultural hegemony. Technical shortcomings (token cost, performance gap) are symptoms of a deeper bias: which voices, languages, and economies are considered “core” versus “edge.”
As executives, it’s tempting to think that we’ll translate later. But translation alone fails to address symbolic inflation, semantic mismatch, and cultural irrelevance. The real challenge is to make AI institutionalized locally and globally
If you’re betting on generative AI to fuel your expansion into new markets, don’t treat language as a footnote. Language is infrastructure. Cultural fluency is a competitive advantage. Token costs and performance differences aren’t just technical: they’re strategic
In the world of artificial intelligence, English has been the path of least resistance. But your next growth frontier? It may require language, culture and cost structures that act as differentiators rather than obstacles
Choose the model, languages and rollout strategy not based on the size of the number of parameters, but on how well you understand your market. If you don’t, you will not only fall behind in performance, but you will also fall behind in credibility and relevance.
The final deadline for Fast Company’s World Changing Ideas Awards is Friday, December 12 at 11:59 PM PT. Apply today.
(Tags for translation) Artificial Intelligence
ــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــــ
