Multilingual AI Agents for Business and Government
By Lal Antony
A multilingual AI phone agent answers one business number and holds the conversation in whichever language the caller speaks. Frontly runs an agent in one of three modes: fixed English, adaptive detection that matches the caller turn by turn, or pinned to a single language. Full conversations are supported in more than 50 languages, with coverage varying by language. There is no second number and no press-one menu.
That is the short version. The longer version is about why almost nobody solves this with people.
Why is multilingual phone coverage so hard to hire for?
Because it is not one hiring problem. It is three, stacked.
Fluency is a narrow filter. A receptionist who is genuinely fluent in English and a second language, to the standard where a patient can describe a symptom or a tenant can describe a fault and be understood correctly, is a far smaller pool than a receptionist. In most local markets it is a pool of a handful of people, and they are usually already employed.
Coverage is a multiplier. One person covers an ordinary full-time week. They do not cover lunch, vacation, sick days, holidays, or the 5pm to 9am window where a large share of inquiry calls actually land. If the second language has to be available whenever the phone rings, one bilingual hire is not the answer. Two or three are.
Demand is lumpy. A practice with a largely Spanish-speaking patient base may take thirty Spanish calls one week and six the next. You cannot staff to that. You either overstaff and carry the cost, or understaff and lose the call.
Most businesses resolve this by quietly not solving it. The caller who is not confident in English rings, struggles, and hangs up. Nothing appears in the system, because a call that ends in silence leaves no record. It is the cheapest problem in the business to ignore and one of the more expensive to actually have.
What does a multilingual AI agent actually do on a call?
It runs the whole conversation, not a translation layer bolted onto an English script. The agent listens, understands, answers, captures the details and books the appointment in the caller's language.
There are three modes, set per agent rather than for the whole account:
| Mode | What happens | When to use it |
|---|---|---|
| Fixed English | Every call runs in English | Your customer base is English-speaking and you want no ambiguity |
| Adaptive | The agent detects the caller's language and matches it, turn by turn | A mixed customer base on one public number |
| Pinned | Every call on this agent runs in one nominated language | A dedicated community line, or a second number for a specific language |
Voice is chosen separately from language. There are more than 50 voices spanning American, British and European accents, and a custom voice can be cloned from a short sample. That matters more than it sounds: an American-accented voice on an agent in adaptive mode still switches languages mid-call. Accent and language are not the same setting.
The full detail on modes, voices and persona is on the Voices and Languages page.
What does this look like in a real business?
A four-person physical therapy and massage clinic runs Frontly in English and Mandarin today. They had been trying to hire a receptionist fluent in both, at a wage the clinic could justify, and had not been able to. Instead of hiring, they went live with Frontly connected to their practice software.
The shape of that outcome is the interesting part. It was never really a technology decision. It was a hiring decision they could not close, and the phone kept ringing while they could not close it.
The same shape appears across other industries. An immigration or family law practice takes inquiries from people who will explain their situation far more accurately in their first language. A property manager gets a maintenance fault described correctly the first time, which means the right trade is dispatched the first time. A hotel takes a direct booking from an international guest instead of losing it to an agency.
If you want the money side of the comparison rather than the staffing side, the worked numbers are in AI vs human receptionist: the real ROI and in the ROI calculator.
Where does this matter for government and public services?
Differently, and more sharply.
For a business, multilingual coverage is a commercial advantage. For a public-facing service, language access is usually a service standard rather than an optional extra, and for agencies receiving federal funding it is a documented obligation. The department, county or city is expected to be reachable by the community it serves, not only by the part of it that is comfortable in English. That expectation does not scale down on a quiet Tuesday, and it does not pause after hours.
The usual operational answer is an interpreter line. It works, and for complex matters it is the right answer. But it carries a specific cost shape that anyone who has run a public hotline knows well:
- The caller waits while an interpreter is brought onto the line
- Availability is thinnest in exactly the languages with the smallest speaker pools
- Every routine call is priced the same as a complex one
- Out-of-hours coverage is either expensive or absent
A large share of hotline volume is not complex. It is where do I send this form, what are your hours, has my application been received, what do I need to bring, how do I schedule. Those calls do not need a certified interpreter. They need to be answered, correctly, in the caller's language, immediately.
An AI agent takes that routine layer. The result is not that interpreters are replaced. It is that the interpreter queue is no longer clogged with opening-hours questions, so the people who genuinely need a human interpreter wait less.
How does a caller choose a language without a phone menu?
Two ways, and neither one is a menu.
In adaptive mode, the caller simply speaks. The agent detects the language in the first turn and continues in it. There is nothing to press and nothing to listen through, which matters most for exactly the callers who find a long English menu hardest.
For a service that wants an explicit choice, or that wants a different workflow behind each language, the cleaner design is a routing agent. A front agent greets the caller, works out what they need and which language they are speaking, and hands off to the right specialist agent. Each of those agents can have its own language, voice, knowledge and workflow. The context travels with the call, so the caller does not start again after the handoff. That is covered on the Routing and Transfer page.
The difference from a traditional phone tree is that the caller never navigates anything. They say what they want, in their own language, and the system sorts it out.
What does multilingual AI not replace?
Worth being direct about this, because the honest limits are the reason to trust the rest.
Certified interpreting for high-stakes matters. Legal proceedings, clinical consultations, child welfare, anything where a misunderstanding has serious consequences: that is interpreter work. An AI agent should be triaging and scheduling around those conversations, not conducting them.
Equal quality in every language. Coverage varies across the set. English is the strongest, the major European and Asian languages are close behind, and the long tail is thinner. If a specific language is central to your service, place a test call in it before you commit.
Every language in the world. More than 50 is a lot and it is not all of them. Check that yours is on the list before you plan around it.
A human, when the caller wants one. Every workflow can hand the call to a person, with a spoken briefing so your team picks up already knowing who is calling and why.
How do you set this up?
Three settings, not a project. Pick the voice, choose the language mode, define the persona. A standard agent is live in about an hour, and you can change the wording yourself afterwards without going back to anyone.
The quickest way to judge whether the language handling is good enough for your callers is to hear it. Place a free test call and try it in the language your customers actually use.