Your Customer Opens in Tamil and Finishes in English. Your Phone System Should Not Care.
Indian jewellery customers switch language mid-sentence and send voice notes rather than typing. What multilingual actually has to mean for a showroom's phone line and WhatsApp number.
Language is not a setting. It is how the sale happens.
A jewellery purchase in India is rarely a solo decision and rarely a single-language conversation.
A daughter messages in English. Her mother rings and speaks Tamil. The father asks about the making charge in a mix of both. A customer in Ludhiana opens in Hindi, drops into English for the numbers, and goes back to Hindi for the negotiation. Somebody sends a thirty-second voice note in Malayalam because typing it would take five minutes.
None of that is unusual. It is the normal texture of Indian retail, and it is the thing most customer-facing automation handles worst.
What "multilingual" usually means, and why it is not enough
Plenty of systems claim multilingual support. In practice it usually means one of three weaker things.
Language selection at the start. "Press 1 for English, press 2 for Hindi." This asks the customer to declare a language before they have said anything, and then locks them into it. Real conversations do not stay locked. Separate agents per language. A Tamil line, a Hindi line, an English line. Now you are maintaining three sets of information and three sets of answers, which drift apart within a month. When the gold rate changes, it changes in three places or it is wrong in two. Translated, not fluent. Answers written in English and machine-translated on the way out. It is technically the right language and unmistakably not a person, and in a category built on trust that reads badly.The requirement is harder than any of these: one agent, one set of approved information, that handles whatever the customer actually does.
The three things that have to work
Mid-conversation switching. A caller who starts in Hindi and says "what's the making charge on that" in English has not changed language — they are speaking normally. The agent needs to follow without the customer noticing there was anything to follow. Hinglish as a first-class language. "Kitna hua mera scheme" is not broken Hindi or broken English. It is how a very large number of Indians type, and an agent that handles Hindi and English separately handles this badly. Voice notes. A substantial share of Indian WhatsApp enquiries arrive as spoken audio rather than text. An agent that reads text only is an agent that ignores a large slice of the inbox — and disproportionately the slice from older, higher-value customers who do not enjoy typing.Where it actually matters in a jewellery business
It is not evenly distributed. A few moments carry most of the weight.
The rate call. The highest-volume inbound call in Indian jewellery retail, and one that arrives in every language your catchment speaks. Getting this wrong is not an edge case; it is most of your phone traffic. This is covered in detail on the showroom voice agent page. The bridal conversation. Multi-visit, multi-person, and frequently multi-language within the same family. The follow-up that reaches the mother has to work in her language; the confirmation that reaches the daughter may not. Scheme enquiries. Long-running, often with older customers, often spoken rather than typed, and about money — which is exactly where a customer will not tolerate being misunderstood. Festival and wedding outreach. A campaign that goes out in one language to a list that speaks four is a campaign with a built-in ceiling.The regional reality
A jewellery chain with showrooms in Chennai, Coimbatore, Hyderabad and Bengaluru is running across at least four language environments before anyone has said a word about Hindi or English.
Staffing for this the traditional way means language-capable people at each branch, at every hour the phone is on. That works for the floor, where people are physically present anyway. It does not work for the phone line at 9pm, or for a WhatsApp inbox that never closes.
One agent covering Hindi, English, Hinglish, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati and Bengali removes the staffing problem from the channel where staffing was never realistic — without removing anyone from the floor, where language skills matter most.
What it must still refuse to do
Multilingual capability raises a risk worth being explicit about: confidence in the wrong language is worse than no answer at all.
An agent that half-understands a Malayalam voice note about a repair and confidently confirms a collection date it invented has done real damage — to a customer who will now arrive at your counter expecting something that is not there.
So the same rule applies as everywhere else: the refusal path is designed before the answer path. If the agent cannot resolve what was asked with confidence, in any language, it says so and hands to a person with the audio and the transcript attached. An advisor who listens to twenty seconds of a voice note and calls back is a good outcome. A wrong answer delivered fluently is not.
What to test before you trust it
Do not evaluate this on a demo script. Evaluate it on your own traffic.
Take twenty real enquiries from your WhatsApp number — a mix of typed, spoken, Hinglish and regional — with customer details removed. Run them through. Look at three things:
- Did it understand what was asked? Not whether the reply was fluent. Whether it identified the right intent.
- Did it answer from your information, or from something plausible it generated?
- Did it refuse the ones it should have refused? This is the one people forget to check, and it is the one that matters at scale.
Related reading: WhatsApp automation for jewellery stores, the showroom phone line, and the overview of AI agents for jewellery retail.