Your AI speaks multiple languages. But does it speak like your users?

Last month, I participated in the European Chatbot Conference. There were plenty of discussions around GenAI, automation, and the future of customer interactions, but one comment, almost casual, stayed with me.

Oyikansola, a PhD researcher in AI and Data Science, with whom I had the pleasure of sharing a panel, mentioned that in some parts of Africa, people don’t speak just one language and instead, they mix them. English, local dialects, sometimes multiple languages within the same sentence, switching naturally without even thinking about it.

This made me think about something I have not thought too much about before. LLMs can understand multiple languages, even if we mix them. However, can they also answer in the same way, mixing multiple languages and, most importantly, in a way that feels right?

Across the world, millions of people communicate by blending languages in ways that feel completely natural to them. In Africa, some countries mix multiple languages or dialects in one sentence. In Mexico, conversations often move fluidly between Spanish and English, especially in professional or urban environments. In the Philippines, Taglish, and in Singapore, Singlish are part of everyday communication.

Even in more structured environments like Japan, particularly in business or tech contexts, it’s common to hear sentences that incorporate English terms seamlessly into Japanese conversations.

Now compare that to how most conversational AI systems are designed. They tend to assume a much simpler model of communication, one user interacting in one language at a time.

Users are often expected to choose a language up front, and from that moment on, the system assumes consistency. English, Spanish, and Japanese are each treated as separate and self-contained.

But real conversations do not follow those boundaries and sometimes languages are mixed.

For a long time, conversational systems struggled with this reality because they were built on rigid structures such as rules, intent classification, and predefined flows. Now with LLMs mixing languages is not a foreign thing anymore. However, is not as easy at seems.

This behavior is not new. In fact, it has a name: code-switching.

Code-switching describes how people naturally move between languages within a conversation, often within the same sentence. Not as an exception, but as a normal way of communicating.

For Conversational AI, it is not the same to understand code-switching as it is to respond to it, and even less to design an experience around it.

Models like ChatGPT or Gemini can already understand mixed-language inputs quite well. But when it comes to responding, if languages are mixed, something still feels slightly off.

This leads to a more fundamental question: should conversational AI aim for accuracy, or authenticity?

If the priority is accuracy, the path is relatively clear. Responses are clean, grammatically correct, and consistent, typically within a single language. This approach makes systems easier to evaluate, scale, and control, particularly across multiple markets.

But accuracy alone can create interactions that feel distant. The system works, but it does not fully resonate with the user.

On the other hand, aiming for authenticity means trying to mirror how users actually communicate. This includes allowing mixed languages, adapting to tone, and reflecting conversational patterns more closely.

The challenge is that current systems are not yet fully capable of doing this in a reliable way. While they can imitate mixed-language communication, the result is not always convincing. When authenticity is attempted but not achieved, it can feel forced or artificial, which can be just as problematic as being overly rigid.

The trade-off is real. Choosing a single language is often the safest and most consistent option, but safety comes at a cost. The moment users are forced into a single, “correct” way of speaking, the system stops adapting to them, and instead asks them to adapt to the system.

This is where the discussion moves beyond language and into customer experience.

From a CX perspective, especially when operating across multiple markets, supporting a language is not the same as understanding how that language is used in practice. The differences are often subtle, but they have a meaningful impact on how users express intent, emotion, urgency, and even politeness.

For example, Mexico and Spain may share the same language on paper, but the way it is used in conversation can differ significantly. Similarly, communication styles in Philippines differ from standard English in ways that go beyond vocabulary.

If conversational AI systems ignore these nuances, they risk delivering experiences that are technically correct but fail to feel natural or inclusive to the user.

This does not mean that every chatbot should immediately start mixing languages. Instead, it highlights the need for more intentional design decisions.

Companies should begin by designing for how users actually speak, rather than how systems expect them to speak. This includes testing with real user inputs, which often involve mixed-language scenarios, rather than relying on idealized or standardized language.

It also requires making conscious decisions about when clarity should take priority over naturalness, and when a more conversational approach adds value to the experience.

Perhaps most importantly, this should be treated as a deliberate design choice. In many cases today, it is not a choice at all, but rather a side effect of how systems have traditionally been built.

We often frame inclusivity in conversational AI in terms of how many languages a system supports. While this is important, it only addresses part of the challenge.

Inclusivity is not just about language coverage. It is about communication, and more specifically, about aligning with how people naturally express themselves.

The reality is that there is no perfect answer. Whether you decide to prioritize accuracy, authenticity, or a balance between the two, there will always be trade-offs. In some cases, consistency and clarity will matter more. In others, sounding natural and familiar will have a bigger impact on the experience.

What matters most is not the specific choice you make, but the mindset behind it.

Designing conversational AI for multiple markets requires acknowledging that users are not uniform, even when they share the same language. Each market has its own nuances, habits, and ways of communicating, and those differences should be considered from the beginning.

When those scenarios are explored upfront, decisions around language, tone, and behavior become intentional. Because instead of taking the safest or most convenient path by default, you are making conscious trade-offs based on the experience you want to create.

In the end, inclusivity is not about trying to perfectly replicate how everyone speaks. It is about recognizing those differences, designing with them in mind, and making deliberate choices rather than one-sided assumptions.

Anything else is not really inclusive. It is just standardized.