How a voice agent remembers and follows a language switch

A good voice agent keeps track of what a caller has said by writing it down as the call goes: a running note, updated after every turn and kept in front of the agent for the whole call. The note holds the qualifying answers (name, budget, configuration, locality, timeline, a site visit), anything else the caller mentioned in their own words, and what has already been asked. On a follow-up call, the note starts from the CRM record. Following a language switch works the same way: the caller's language is decided on every turn, from what they actually said, and the agent follows from its next reply. Neither is left to chance.

That is how BlackWolf's agents work. This piece explains the approach, and the real calls that shaped it. For the basics, start with what an AI voice agent is.

Why agents re-ask

The obvious design is to let the agent remember the conversation as it goes. It fails in three ways.

The history does not stay whole. On a long call, the conversation an agent works from is trimmed or summarised, partly to keep replies quick (see voice latency). What goes first is the oldest part, which is exactly where the caller gave their name and budget.

A checklist pushes it to ask. An agent reminded after every answer which fields are still empty asks for them, whether or not the caller has just answered in passing. We have heard it happen: an agent asking for a contact number again and again through a call in which the caller wanted to talk about price and payment plans.

A mention in passing is easy to miss. "3 BHK dekh raha tha, budget around 90 lakh hai, but possession kab milega?" (I was looking at a 3 BHK, the budget is around 90 lakh, but when will I get possession?) contains two answers inside one question. An agent busy answering the question may not register that the budget is now settled.

The fix for all three is the same. Write the facts down where they cannot be lost, and put them in front of the agent on every turn.

The running note

What goes in

After every caller turn, the note is updated from what was said. Two kinds of thing go in.

• The qualifiers: name, budget, configuration, locality, timeline, purpose and site visit. These fill the CRM fields that make lead qualification worth something.
• Everything else they said, in their words. A spouse's name, where the children study, the rent they pay now, a preference for an east-facing flat. None of these has a field, and without a note all of them fall out of a long conversation a few turns later. Keeping the caller's own statements, not just the fields, stops the agent asking about something the caller explained ten minutes earlier.

What does not count

Most of the work is deciding what not to write down. Each of these rules comes from a real call.

• A loan amount is not a budget. Nor is an EMI, a rent, a salary or a down payment, unless the caller says it is their budget.
• A price the caller read in an ad is not their budget.
• The agent's own price, repeated back by the caller as a question ("80 lakh? itna?", meaning "80 lakh? That much?"), is not their budget.
• "Haan" (yes) is not a name, and neither is "sir" or "doctor".
• A placeholder such as "unknown" or "not given" is never stored. A blank is obviously missing; "unknown" looks like data, and once a field looks filled nobody asks for the real answer.
• A day agreed in answer to "kab aana chahenge, Saturday ya Sunday?" (when would you like to come, Saturday or Sunday?) is a booked site visit, even though the question never said "site visit".

When a caller corrects themselves ("Rohit... sorry, Rohan"), the note takes the new value and records that it replaced the old one. If both were simply left in the conversation, the agent could use whichever it found first.

How the note is used

The note is kept in a short, fixed form. Simplified, it reads like this:

Your notes: name Meera; budget 90 lakh; wants 3 BHK; site visit Saturday evening. Already answered, and asking again is a mistake: name, budget, configuration, site visit.

The settled questions come last on purpose: that is where nothing said later in the call can bury them. We checked it on replayed real calls.

Around the note sit the rules for asking:

• One question at a time, never a list.
• Never the same question twice in a row. If the caller ducked it, answer what they asked and come back later, as a person would.
• An unanswered question is asked at most twice, then dropped for the call. The name gets one last, polite try in the goodbye, where it reads as a courtesy: "before I go, may I note your name for the visit?"
• Nothing is asked of someone who is leaving. A caller who says they are not interested, or asks not to be called again, gets a goodbye, not the next question.

Memory across calls

On an outbound or follow-up call to a known lead, the note does not start empty. When the call opens, it is filled from the lead's CRM record: name, budget, configuration, locality, status and what earlier calls found out. The agent knows it rang this person, what it already knows and what it still needs, and it opens using their name. It never asks for what is on the record as though hearing it for the first time. If nothing is missing, the call is about moving the buyer forward, not qualifying them again.

The memory belongs to the lead and to one workspace. The agent does not learn across clients, and one client's leads are never read into another client's call.

Following a caller who switches language

Hearing a switch is not following it

Telling an agent to follow the caller's language works in typed tests. On real audio, it is not enough: an agent can hear a caller's Hindi perfectly and still reply in Gujarati. It hears correctly and chooses wrong, and telling it more firmly does not fix that reliably.

Two lessons follow. Test language behaviour on audio, not text. And make the decision where it can be made reliably: by fixed rules, on every turn.

Deciding the language on every turn

After each caller turn, fixed rules read what was said and decide which language it is in. Each one comes from a real call:

• Grammar, not script. Hindi can arrive written in Gujarati letters, so the decision rests on small grammar words (hai and chhe for "is", nahin and nathi for "not", kya and shu for "what"; the first of each pair is Hindi, the second Gujarati), recognised in all three scripts. Why this matters is in what it takes for voice AI to sound local in Hindi and Gujarati.
• A clear margin. One ambiguous word never switches a call. Between Hindi and Gujarati, one has to lead clearly.
• No evidence means no change. "Haan" (yes), "ok", "achha" (okay), a number or a name carry no grammar, so they move nothing. Guessing on them is how an agent ends up flip-flopping.
• English needs English grammar. "Price kya hai?" (What is the price?) is Hindi with an English noun in it, which is ordinary spoken Hindi, not a switch to English.
• Naming a language wins. "Can you please speak in Hindi?" is English by its grammar and a request for Hindi by its meaning. When a caller names exactly one language, that settles it.

Switching, and when

When the language changes, the agent gets a short, silent note: the caller is speaking Hindi now; reply in Hindi from your very next sentence; the opening line does not set the language, the caller does; keep the same accent; and if they change again, follow that.

Timing matters more than wording. The note goes in between turns, never while the caller is speaking: sent mid-reply, it cuts the agent off, which is worse than waiting. The cost of waiting is one turn: the reply to the turn in which the caller switched may still come in the old language, and every reply after it follows the caller. It is a trade worth making.

Drift and mixing

Two failures look like switching but are not.

Drift is the agent sliding back into its default language when the caller has not changed anything. Every reply the agent gives is checked too, and when one is in a different language from the caller's, a firmer note follows: you slipped; the caller has not changed; stay in their language for the whole of every reply, including the goodbye. It is deliberately different from the switching note, which says the language is not locked. Telling a drifting agent that the language is open invites it to drift again.

Mixing is Hindi words inside a Gujarati reply, or the other way round. A reply that is mostly right passes any check that asks which language it is in, so a separate check asks whether it is purely the caller's language. Ordinary English words such as project, booking and EMI are allowed.

The accent stays the same throughout, so a switch sounds like the same person changing language, not someone new taking the phone. The call closes in the language the caller was using. Where the client's WhatsApp Business account is connected, a caller who asks for details can get the brochure on WhatsApp after the call, once they confirm the number. That message goes out through the client's approved WhatsApp template, which is in English today, not in the language of the call.

What it sounds like

A sample call, with English translations in brackets. The dashboard stores each line as it was spoken; the translations are only for this page:

1. Agent opens in Gujarati, naming the company it is calling for.
2. Caller: "Haan, bolo." (Yes, go ahead.) No grammar to go on, so nothing changes.
3. Caller: "Mujhe 3 BHK chahiye, budget 90 lakh tak hai. Possession kab hai?" (I want a 3 BHK, budget up to 90 lakh. When is possession?) Hindi, with a clear margin. Budget and configuration go into the note.
4. Agent answers the possession question from the brochure. This reply may still be in Gujarati: it is the one-turn cost. The switching note goes in as soon as it ends.
5. Caller: "Aur parking ka kya hai?" (And what about parking?) The agent answers in Hindi, and every reply from here is in Hindi. Its next question is about locality, not budget.
6. Caller: "Site visit Saturday ko ho sakta hai?" (Can the site visit be on Saturday?) The agent agrees a time, in Hindi, and the note now holds a booked visit.

When the call ends, the lead shows 3 BHK, 90 lakh, the locality, a summary and the transcript, with the recording once the carrier supplies it, and the visit is on the Site Visits schedule with the Saturday the caller named, for the team to confirm. On the next call to this buyer, the agent starts from all of it.

What to test

On any vendor's agent: mention your budget in passing and see whether it asks again; call back as the same lead and see whether it remembers; switch from Gujarati to Hindi mid-call; answer only "haan" (yes) a few times; ask in English for Hindi. The rest of the checklist is in how to evaluate an AI voice agent vendor, and more on what an AI voice agent should do is in our pillar guide.

What to do next

Test memory and language switching on a real call before anything else, because they are where agents most often sound like machines. For how this works when selling property, see real estate.