Code-switching is when a person moves between two languages in the same conversation or sentence — "Mujhe pricing samajhni thi, actually I need it for my sales team." In India it's the normal way customers write and speak. Customer service AI built mainly on English, or on textbook Hindi, misreads these messages, replies in the wrong language or asks people to repeat themselves. Good AI understands the mix and answers in the same register.
What does code-switching look like in real customer messages?
These are the kinds of messages Indian businesses receive every day on WhatsApp and Instagram:
- "3BHK chahiye, Gurgaon mein."
- "Price thoda high hai."
- "Bhaiya ye wala package kitne ka hai?"
- "Hi, mujhe pricing samajhni thi. Actually, I need it for my sales team."
Two languages, one message, no warning. Linguists distinguish code-switching (changing language between sentences) from code-mixing (inside a sentence). For a business, the difference doesn't matter. What matters is whether your AI understands both.
"In NimbleBiz conversations, 1 in 4 messages mixed Hindi and English."
What does research on Hindi language AI show?
The gap starts with data. Research on emotion detection in Hindi found that English emotion datasets have tens of thousands of labelled examples; Hindi had barely a handful. Models trained on so little failed more often than they worked — not because they were weak, but because they had almost nothing to learn from.
Generating balanced Hindi training data with GPT and fine-tuning on it raised accuracy from 25% to 85%. The model stayed the same; the context changed. The research was presented at the SGAI 2024 International Conference on Artificial Intelligence in Cambridge.
The same pattern shows up in real customer conversations: AI that doesn't understand how people actually speak will miss what they mean.
Why do English-first chatbots fail at Hinglish?
They fail in four predictable ways:
- They reply in the wrong language. A customer writes in Hinglish and gets formal English back. It reads as cold, and the conversation often stops there.
- They miss meaning inside the mix. "Thoda high hai" isn't a complaint, it's an opening to negotiate. Keyword matching can't hear that.
- They ask for what was already said. A customer gives a budget, location and timeline in one Hinglish line and then gets a form asking for all three.
- They treat an unexpected language as an error. When a prospect on our own WhatsApp typed one word, "español," the conversation simply continued in Spanish, and we learned they ran a travel agency that wanted to sell on WhatsApp. That's what language support looks like when it works: invisible.
How can you tell if your AI handles code-switching?
Run these five tests before you trust any chatbot or voice agent with Indian customers:
- Switch mid-sentence. Type "Price kya hai, and is delivery free?" and check that it answers both parts in the same mix.
- Use transliterated Hindi. Most Indians type Hindi in Roman script, not Devanagari. "Kal tak chahiye" must work as well as "कल तक चाहिए".
- Mix numbers and units. "Budget 1 crore ke aas paas" should be captured as ₹1 crore, not ignored.
- Check the tone. A reply in formal "shuddh" Hindi to a casual Hinglish message is still the wrong register.
- Read last month's chats. Count how many conversations ended after the bot's first reply. Language mismatch rarely triggers a complaint. It triggers silence.
What should a business do about it?
Start by measuring. Look at a week of WhatsApp and Instagram enquiries and note how many mix languages. If it's more than a handful, your customers are telling you how they want to be spoken to.
Then fix the reply language before anything else. As our language-routing guide puts it: reply wide, staff narrow. Let the AI answer in whatever mix the customer uses; keep human escalation to the languages your team actually speaks.
At NimbleBiz, we built our agents to detect the language of every message and reply in the same mix on chat and voice calls because this was the gap we kept seeing in Indian businesses.
Frequently asked questions
Is Hinglish a language? Not officially. Hinglish is a mix of Hindi and English used in everyday speech and text across India. For customer service it should be treated as its own case, not as broken Hindi or broken English.
What's the difference between code-switching and code-mixing? Code-switching usually means changing languages between sentences; code-mixing means mixing within a sentence. Indian customers do both, often in the same chat.
Can AI understand Hindi written in English letters? Good AI can. Most Indian customers type Hindi in Roman script, so test any tool with transliterated messages, not only Devanagari.
Does code-switching happen in other Indian languages? Yes Tamil-English (Tanglish), Bengali-English and Marathi-English mixes are common too. The same tests apply.
Should a chatbot reply in Hinglish if the customer writes in Hinglish? Usually yes. Matching the customer's register feels natural; switching to formal Hindi or English reads as a change of tone.
Want to see how NimbleBiz handles Hinglish on WhatsApp and calls? Book a 15-minute demo.