Why your support bot fails at Hinglish, and what to do about it

The message that breaks most bots

"Bhaiya order kahan hai" is a completely ordinary customer message on any Indian D2C WhatsApp line. It's also enough to break intent classification on a support bot trained primarily on English text, because the system doesn't recognize "kahan hai" as "where is" and falls back to a generic "I didn't understand that" response, exactly the moment a customer is already frustrated about a late order.

Why this happens

Most conversational AI systems, even ones marketed as "multilingual," are trained to detect a single language per message and route accordingly. Hinglish doesn't work that way: it code-switches within a sentence, sometimes within a clause, following its own consistent internal grammar rather than random mixing. "Order kahan hai" isn't broken English or broken Hindi; it's a third, stable pattern with its own rules, and a system looking for "English" or "Hindi" as discrete categories misses it entirely.

What breaks specifically

Three things tend to fail first: intent detection (the system can't map the phrase to "order status" because the training data didn't include this exact code-switch pattern), entity extraction (pulling out order numbers or product names embedded in a Hinglish sentence), and tone matching (a bot that replies in stiff formal English to a casual "bhaiya" message reads as robotic and off-putting, even when the answer itself is correct).

What actually works

Systems trained specifically on code-switched conversational data (not formal Hindi text, not formal English text, but actual WhatsApp-style Hinglish chat) handle this materially better, because they've seen the code-switch patterns directly rather than trying to decompose them into two "pure" languages. Catalog-native agents that stay anchored to your actual product data also do better, because "order kahan hai" maps directly to an order-lookup action, not to open-ended text generation that has to first correctly parse the sentence.

What to check in a vendor demo

Don't test with clean, formal-language example messages picked to work well. Test with real, messy customer chat pulled straight from your own support logs (abbreviations, spelling variants, code-switching mid-sentence), and see whether the bot actually understands or just pattern-matches its way to a plausible-sounding non-answer.