The phone rings. It is your son's voice, or something very close to it, and he is in trouble. He needs ₹60,000 on UPI in the next ten minutes and he cannot explain why. That is the exact shape of a case Scroll.in reported from Noida in April 2024, and the mother who received it later said one word had felt wrong: the caller said "mummy" when her son always says "mom". She sent the money anyway, because fear is faster than doubt.
Lacewing Technologies is an AI and software development company in Navi Mumbai. We build products of our own and we build systems for clients, and one of the things we build is AI voice agents for businesses. That puts us in an awkward spot when writing about voice scams, so let us say the position up front: every voice agent we ship announces that it is an AI at the start of the call. The rest of this post is about what happens when someone uses the same technology and hides it.
This is a protection guide, not a technical one. For how analysts tell synthetic audio from real audio, a separate post covers how to detect AI-generated voice. Here we care about the call you are on right now, and what to do before you reach for your UPI app.
Which voice scam patterns have Indian authorities documented?#
Four scripts account for most of what gets reported. They predate voice cloning. What cloning adds is a familiar voice at the top of the call, which removes the one moment of hesitation that used to save people.
The relative in trouble. A call from a child, a nephew, a parent. There has been an accident, an arrest, a kidnapping, a lost wallet abroad. Money is needed now and the person cannot talk for long. Tamil Nadu Police's Cyber Crime Wing warned in April 2024 that fraudsters were cloning the voices of people the victim trusts and inventing emergencies to push payment through UPI. The Scroll.in report describes the same pattern in Delhi-NCR, with the voice sample lifted from social media.
The fake bank or courier call. Someone from "your bank" says your KYC has lapsed and the account will be frozen tonight unless you confirm a few details, including the OTP that has just arrived. The Reserve Bank of India's February 2024 caution describes exactly this: unsolicited calls, a manufactured urgency, and threats of "account freezing/blocking/closure". The courier variant uses a parcel instead of a KYC form. A caller posing as FedEx says a package in your name has been held with drugs or fake passports inside. FedEx's own statement, reported by The Tribune, is that it "does not request personal information through unsolicited phone calls, mail, or email for goods being shipped or held".
The "digital arrest". This is where the courier call usually goes next. The call is transferred to a "police officer" or "CBI official", often on video, in uniform, in front of what looks like a police station. The Ministry of Home Affairs alert of May 2024 describes victims made to "remain visually available over Skype or other video conferencing platform to the fraudsters, till their demands are met", and notes that the fraudsters "use studios modeled on Police Stations and Government offices and wear uniforms to appear genuine". In the Bengaluru case The Tribune reported, a lawyer was held this way for two days and lost about ₹14.57 lakh. The Indian Cyber Crime Coordination Centre's advisory in October 2024 put the counter in one line: "CBI/Police/Custom/ED/Judges DO NOT arrest you on video call."
The CEO payment instruction. The business version. A finance officer receives a WhatsApp message, an email or a call from the MD asking for an urgent transfer to a new vendor account, often with instructions not to discuss it. SEBI's press release of July 2026, issued on information from I4C, calls this the "Boss Scam" and names the technique directly: "impersonation of MDs/CEOs through deep-fake methodologies like voice cloning, use of AI on video calls". The same release describes a second route, where a malicious zip file hijacks the finance officer's WhatsApp Web session or saves the fraudster's number under the CEO's name in the phone's contacts.
What is common to all four: a trusted identity, a deadline measured in minutes, a reason you cannot verify through the normal channel, and a payment to an account you have never used. The voice is the decoration. The script is the crime.
How does voice cloning actually work?#
In plain terms, without the maths.
A text-to-speech model is trained on a very large amount of recorded speech and learns the general relationship between written words and the sounds people make. On its own it speaks in a generic voice. To make it sound like a particular person, it is given a sample of that person's voice and told to match the sample's pitch, timbre, accent and pacing. Older systems needed hours of clean studio recording. Current systems can work from much less, and the Scroll.in report notes that scammers are using ordinary commercial voice services sold on monthly subscriptions, some with free trials.
Where does the sample come from? Anywhere a voice is public: Instagram reels, a wedding video on YouTube, a voice note forwarded through three WhatsApp groups, a "survey" call that keeps you talking for a minute.
Two things follow. The clone is a performance of a voice, not a copy of a mind: it can say anything the scammer types, but it does not know what the real person knows. And it is only as good as the model's guess about how that person sounds in a situation it has never heard them in. A son's voice cloned from cheerful reels has to be made to sound terrified, and that is where the seams show.
What are the tells that give a cloned voice away?#
Some of these are about the audio. Most are about behaviour, and those are more reliable because they do not require you to be an audio expert while frightened.
Urgency with no room. Every documented script compresses time. Money in minutes. The account frozen tonight. The arrest happening now. A real emergency can wait two minutes for a call-back. The script cannot, because the call-back is what kills it.
A push to another channel. "Don't call this number back, I'm on a borrowed phone." "Switch to WhatsApp video." Each move takes you off a channel where you could check and onto one the scammer controls. The SEBI release describes transfers instructed entirely through "messages/ fake calls on Social Media Platforms" for the same reason.
Refusal of a call-back. Say "I'll call you back on your number" and watch what happens. The real person says fine. The scam produces a reason: the phone is dying, they are in custody, the officer is only reachable on this line. No bank, courier or police officer is unreachable through the number on their official website.
Odd pauses and delayed replies. If the voice is being generated from typed text, there is a lag between your question and the answer, and the answer tends to be a complete, slightly formal sentence. Real people interrupt, trail off and say "haan, haan" while you are still talking.
Background that is wrong. Either there is none, which is strange for someone supposedly at an accident site or a police station, or it is too clean: a loop of traffic noise, a siren that never gets closer or further away. Real background sound is messy and changes.
The voice cannot answer a personal question. Ask something only the real person would know and that is not on any social media profile: what you ate at dinner last Sunday, what your grandmother called you. A clone is reading a script. The scammer behind it does not have the answer and will deflect with urgency: "Mummy, there's no time for this."
The payee is new. The money is going to an account or UPI ID you have never paid before, and the caller is adamant it must be that one. Real banks never ask you to move money to a "safe account". Real police do not collect bail over UPI.
Scam pattern, tell, action#
| Pattern | What the caller says | The tell | What you do |
|---|---|---|---|
| Relative in trouble | "I've had an accident, send money now, don't tell anyone" | Urgency, new payee, cannot answer a personal question | Hang up, call the relative on the saved number, ask the safe word |
| Fake bank / KYC | "Your account will be blocked tonight, confirm the OTP" | Asks for OTP or PIN, threatens closure | Hang up, call the bank's number from its website or your card, never share an OTP |
| Fake courier | "A parcel in your name has drugs in it, we are connecting you to police" | Transfer to a "police officer", request to stay on video | Hang up, check the shipment on the courier's official site, report to 1930 |
| Digital arrest | "You are under arrest, keep your camera on until this is resolved" | No agency arrests on video; demand for payment to clear the case | End the call. Do not pay. Report to 1930 or cybercrime.gov.in |
| CEO payment instruction | "Transfer to this vendor today, confidential, don't discuss" | New channel, new account, secrecy, pressure to skip process | Call the executive on their directory number, apply the two-person rule |
What is a family safe word protocol?#
Agree on a word or phrase in advance, in person, that is never written down or sent on any chat. Then any call that asks for money or help must include it. A version for an ordinary Indian household:
- Choose the word together, offline. Something ordinary and unguessable. Not a pet's name, not a birthday, not anything that appears on Instagram. Agree on it at dinner, not on the family group.
- Agree the rule that goes with it. Any call, from any number, in any voice, that asks for money, an OTP, a password or an urgent favour must start with the safe word. No safe word, no action, however convincing the voice.
- Decide what happens when the word is missing. The person receiving the call says "I'll call you back" and hangs up. They then call the relative on the number already saved in their phone. If there is no answer, they call a second family member. Nobody transfers anything until the real person has been reached.
- Protect the elders specifically. The relative-in-trouble script targets parents and grandparents because they are most likely to panic and least likely to question a child's voice. Walk them through the rule, let them practise saying "I'll call you back", and check the saved numbers in their phone.
- Rotate it if it leaks. If anyone says the word aloud on a call that turns out to be genuine, treat it as used and pick a new one at the next family meal.
The protocol moves the decision from "does this sound like my son" to "did this caller say the word". A frightened person will get the first question wrong against a decent clone. Anyone can answer the second.
What should a company's call-back protocol look like?#
The Boss Scam succeeds because finance teams are trained to act on instructions from senior people, and senior people are often unreachable. SEBI's advice is to "cross-check requests received through WhatsApp/ email/ social media platforms, by calling their seniors" and to "not transfer funds solely on the basis of instructions received on any social media platforms". Turned into a rule a small Indian business can run:
- No payment instruction is complete on the channel it arrived on. A WhatsApp message, an email or a phone call can start a request. It cannot finish one. The request is actionable only after a call-back.
- Call back on the directory number, never the number that called. The finance officer dials the executive on the number in the company directory, not the number in the message and not the one saved in their own phone under the boss's name. SEBI's release describes fraudsters saving their own number under the CEO's name on a compromised device. The directory is the one source a message cannot edit.
- Any new or changed beneficiary needs two people. A transfer to an account the company has never paid, or a change to a vendor's bank details, requires a second person to confirm with the vendor on a number taken from a previous invoice, not from the email requesting the change.
- Secrecy is a stop signal. "Confidential, don't discuss with anyone" is in SEBI's description of the scam almost word for word. Any instruction to bypass the approval chain triggers the call-back, with the founder copied.
- Never open an attachment to a payment request. The second route in SEBI's release is a zip archive carrying an executable that hijacks WhatsApp Web. Payment instructions come in text.
- Log out of idle WhatsApp Web sessions and report immediately. SEBI's list ends with both points: log out, and report any incident on 1930 or cybercrime.gov.in.
Write this as a one-page policy and have the founder or MD sign it too. The protocol only works if the person being impersonated has agreed in advance that their own urgent instruction will be ignored until confirmed. A boss who has said "call me back, every time" cannot be faked by a voice that says "there's no time".
Where do you report a voice scam in India, and how fast?#
Two channels, both named in every official advisory cited in this post.
1930 is the national cybercrime helpline for financial fraud. The MHA alert, the I4C advisory, the RBI caution and the SEBI release all direct the public to it. Call it the moment money has moved. Have ready: the number that called, the time, the account or UPI ID you paid, the amount, the transaction reference and screenshots.
cybercrime.gov.in is the National Cyber Crime Reporting Portal, run by I4C under the Ministry of Home Affairs. File there with the same details for a written record and a complaint number you can quote to your bank.
Then call your bank on the number from the back of your card or its website and dispute the transaction. The RBI's advice is to take a bank's number only from its official website or other verified sources, which applies doubly after a scam.
If no money has moved, report anyway. The MHA alert says I4C has blocked more than 1,000 Skype IDs used in digital arrest scams, along with SIM cards, devices and mule accounts. Each report adds to that list.
Why disclosure is the line between a product and a fraud#
We said at the top that Lacewing Technologies builds AI voice agents, including inbound call-answering agents for clinics, dealerships and service businesses that cannot staff a phone around the clock. Every one of them opens by saying it is an AI assistant. Clients sometimes ask us to drop that line because callers do not notice. We refuse, and this post is the long version of why.
Voice scams use the same underlying technology as a legitimate voice agent. The difference is not technical. It is three things the legitimate product does and the scam cannot survive:
It says what it is. An agent that announces itself as an AI cannot pretend to be your son, your bank manager or a CBI officer. The whole scam depends on concealing exactly this fact.
It calls from a number you can check. A business voice agent calls from, and answers on, a number published on that business's website. Hang up, dial that number, and you reach the same agent. Every scam script contains a reason you must not hang up.
It never asks you to move money inside the call. A legitimate agent books an appointment, confirms a delivery, answers a question about opening hours. It does not ask for an OTP and it does not direct you to a UPI ID.
SEBI's release, RBI's caution and the MHA alert all treat impersonation as the core offence. A voice agent that announces itself is not impersonating anyone. A voice that hides what it is, whether a cloned relative or an undisclosed bot in a sales call, is on the wrong side of the line before it asks for anything.
So the rule for a household and the rule for a business are the same rule. Any voice that wants your money, your OTP or your compliance, and will not let you hang up and call back, has told you what it is. Hang up. Call back on the number you already have. Ask the question only the real person can answer. Then, if anything moved, call 1930.