Untapped AI files patent for the handshake that lets AI agents prove who they are before they do business; will offer it free to SA companies

Lloyd Matthew, Founder and CEO of Untapped AI.

---

Untapped AI, a Cape Town company building multilingual voice AI for South Africa, has filed a patent application with the United States Patent and Trademark Office for a method that lets two AI agents on a call establish who they are dealing with before any sensitive information changes hands. The company sees it as the next stage in voice communication.

Within a few years, much of the world's phone traffic will run between machines as AI increasingly speaks rather than types. Your AI assistant will take calls from your bank's AI about a flagged payment, from a clinic's AI moving an appointment, from a retailer's AI arranging a replacement and from an insurer's AI discussing a claim, all while you are busy and without you picking up the phone. These assistants are already appearing overseas, and South African businesses have begun deploying early voice agents of their own. Untapped AI expects calls between machines to outnumber calls between people before long.

That future inherits every risk of today's phone fraud and compounds it. South Africa loses billions of rand a year to fraud, and not every case is reported. SABRIC, the South African Banking Risk Information Centre, names impersonation among the leading methods, and AI can now clone a voice from a short recording. When a fraudster's AI calls a consumer's AI, there is no person on either end to sense that something is wrong, and a spoofed number or a borrowed voice becomes far easier to pass off. Untapped AI set out to close that gap before this kind of fraud reaches scale.

The danger goes beyond criminals using voice synthesisers or AI speech to deceive. Rogue AI agents will multiply, and some will malfunction, some will be compromised, some will be built to cause harm and some will run in swarms of thousands, placing calls from rotating numbers under spoofed personal or business names.

Untapped AI designed its encryption and handshake with those machines in mind, drawing on every known fraud tactic, past and present. Every call through its network must prove itself afresh, and a failed attempt counts against the caller across the whole network, not only against the business it called. An agent that cannot prove who it acts for is refused and flagged for further testing. A swarm gains nothing from sheer numbers because the handshake is scrambled and recalibrated, and a hijacked agent cannot pass itself off as a business it does not represent because it cannot pass authentication.

Just as two people on a call might recognise a familiar voice, Untapped AI's method lets two AI agents recognise that the other side is a compatible machine, using a signal carried in ordinary-sounding speech that a human listener would not notice. The agents then open a separate encrypted data connection and use the call itself to confirm that the party on that connection is the same party on the line. Only when both sides pass that check does the business move from spoken conversation to the secure connection, where sensitive details can travel safely.

Unlike a code that could be recorded and replayed, the confirmation is worked out afresh for every call, so a copy taken from one conversation is useless on the next. If the other side turns out to be a person, or a machine that cannot prove itself but poses no threat, the agent carries on talking normally and discloses nothing sensitive. If the other side is posing as someone it is not, the agent can end the call and raise the alarm.

Developers demonstrated in 2025 that AI agents can switch to machine-to-machine sound, and published patents from large telecoms and technology companies describe AI callers that recognise each other and move a call onto a data link. Those approaches detect that a machine is on the line, and none of the published patents Untapped AI has reviewed proves who that machine is or whom it acts for. Untapped AI's method takes that next step, built from a defender's point of view. As far as Untapped AI is aware, no business anywhere in the world uses a method of this kind today.

Doorbells and intercoms, two-way radios and voice sessions inside apps are covered alongside ordinary phone lines. A delivery robot at the front door would have to prove itself to the household robot before the parcel is accepted and the door opens. Without that check, a fake courier could talk a home robot into opening up, and the flaw would only become obvious after the crime. Untapped AI will focus first on banking and healthcare clients, where the cost of a convincing impostor is highest.

Untapped AI's voice platform carries further safeguards that are not described in the application and will not be made public, so that fraudsters have nothing to study or work around.

"Today a call has to be secure before it can achieve anything," said Lloyd Matthew, founder and CEO of Untapped AI, who invented the method with Heinz Vollmer. "We establish trust first and only then deal with why the call was made. That reverses the old order of priorities, but the world has changed and we have had to change with it. We are now as much a security business as a voice business, and before we build anything we ask whether it can be made safe."

Because both sides of a call must run the same method, Untapped AI could have used it to corner the market. It chose not to, since a handshake nobody else could use would leave the local market exposed. South African businesses, including Untapped AI's direct competitors, will be able to build the handshake into their own services and use Untapped AI's code free of charge, though they may not repackage the method as a product of their own. For international companies, including the large global AI and voice platforms, Untapped AI will agree terms case by case if and when that becomes necessary.

"I have watched every generation of phone fraud, from call-back scams to SIM swaps, and each one worked because the network trusted whoever was calling," said Heinz Vollmer, who brings 30 years of experience in telecommunications and voice technology. "Within a few years, your own AI voice will speak to your bank's AI more often than you do. That conversation has to begin with proof of who is on the other end."

Untapped AI is now looking for its first partners to run the handshake on live calls. The company wants to hear from banks, insurers, medical schemes and hospital groups that already field impersonation calls, from the telecoms operators that carry those calls, and from South African voice AI builders, competitors included, that want to put the code to work. Early partners will help decide how the handshake is rolled out across South African networks before it reaches wider use, and regulators and fraud investigators can request a technical briefing through the same channel.

"Machines will soon be doing business with each other on our behalf, and the rules for how they prove themselves are being written now," said Matthew. "We would rather write those rules together with South African business than have them handed to us from overseas. If your organisation is about to let an AI take or make calls in its name, we want to talk to you before the first fraudster does."

To book a briefing or request access to the code, visit untappedai.co.za or contact Matthew directly.