Picture three businesses. A two-wheeler financier has customers in Maharashtra, Karnataka and Tamil Nadu. A college in Pune hears from parents in Marathi and students in English on the same line. A D2C brand gets orders from every state, and the only thing it knows about a buyer's language is their pin code. Each of these has a different right answer.
AI Call Studio works in English, Hinglish and Indian languages, and only in India. You can set an agent up in two ways.
Option one: one agent per language
Each agent is set to a language, and stays there. You send the right people to the right agent, because you know something about them: a state, a branch, a language field on your list.
This suits you when you already know each person's language, and when the greeting or the offer really differs by region. It also keeps your review simple, because each agent's calls can be judged by whoever speaks that language. The downside is more agents to keep updated. A fee change has to reach each script unless your facts sit in a shared knowledge base, which is another reason to put them there.
Option two: Multilingual mode
With Multilingual mode on, one agent replies in the language the caller speaks, and the fixed language is then ignored. It is set agent by agent. The script stays in English, and the job carries on with the same fields and the same next step, so your reports stay comparable.
It suits callers who mix languages, and lists where you don't know who prefers what. Mixed Hindi and English is handled best, so test other mixes on your own calls. Start in Hinglish and let the caller move.
How to decide
- You know each contact's language: one agent per language gives you more control.
- You do not know, or your callers mix: Multilingual.
- Inbound lines, where anyone can ring: Multilingual is often easier, and worth testing first.
- A legal or regulated line where the wording must be exact: separate agents, each checked by a native speaker.
Tiers come first
Each voice tier speaks a different set of languages, and which languages an agent can move between depends on the tier. Inside a tier, not every voice speaks all of them. So pick your languages, then find which tier covers them. Some languages are marked early or beta on the languages page, and those still await a real-call and native-speaker check. Try them on your own callers before a full run.
Who checks the words?
Nobody on your team may be able to judge a Kannada greeting by ear, and that is a real risk. Ready-made greetings in Indian languages show a warning until a native speaker has checked them. Have someone who speaks the language read yours, and edit it on the agent's page. Then have them make a test call, and ask them two things: does it sound natural, and would they stay on the line?
Accents matter too. An agent listens in the language you set, and no system hears every accent perfectly. Test with people who sound like your callers. The Live Call Console shows what the agent heard, which is usually the clue when something goes wrong. Some voices can also take custom vocabulary for names and words it mishears.
What about hand-over?
A Tamil caller who asks for a person should not reach someone who speaks only Hindi. A desk can be set to work in a language. So a floor with Tamil, Marathi and Hindi speakers can have one desk per language, with each agent's hand-over list pointing to the right one. With Multilingual, think about what happens when the language turns out to be one your staff don't speak. Offer a callback, and tell the caller honestly when to expect it.
Reviewing what you cannot read
Your QA lead may read only two of these languages. Call translation puts a reading copy next to the original, turn by turn. Call QA grades the original, in the language the call was held in. Reading a translation can tell you what was said. It can't tell you how it sounded, so keep one native-speaking reviewer for each language you run at volume.
A sensible rollout
- Start with your largest language or Hinglish. Get one agent working well.
- Add your next language on a small list, with a native speaker listening to the first calls.
- Compare the two in Analytics, using the same outcomes, and don't forget that fewer calls means a noisier number.
- Only then decide whether to keep separate agents or move to Multilingual.
The recording notice, the calling hours and your Do-Not-Call list apply the same way whichever language the call is in. Opt-out capture, where you have it, works from what callers say in English, Hinglish and Hindi, so in other languages a person may need to add a number by hand. Take that seriously: a caller in Tamil who says stop still means stop. Read how we think about mixed speech in why Hinglish is the hard part.