The future of phone calls: toward 100% AI conversations?

A video recently went viral in the tech world: two AI voice agents exchange over the phone, realize they’re not talking to a human… and switch to a coded language, completely unintelligible to the human ear. This fascinating — and slightly unsettling — moment may mark the beginning of a new era: that of 100% AI-driven conversations.

Phone calls, once the exclusive domain of humans, have already been largely transformed by voice AI. And tomorrow, they could become fully automated, machine-to-machine. In this article, we try to imagine that future: what it changes technically, what it implies legally, and how a platform like Rounded could help shape and support that transition.

The evolution of phone calls

For decades, phone calls followed a simple model: one person calls, another picks up. Then companies began automating certain interactions at scale with the arrival of interactive voice response (IVR) systems. Those classic “Press 1 for...” menus marked the first break in the traditional customer relationship. Rigid and often frustrating, they paved the way for a new generation of voice interfaces.

Over time, callbots took over, offering better voice analysis capabilities. But it was the emergence of LLMs (Large Language Models) that truly propelled voice interactions into a new dimension. Thanks to these models, AI voice agents can now understand natural language, detect user intent, and carry out contextual, fluid conversations.

A symbolic turning point came in 2018 with Google Duplex, an AI capable of calling a hair salon or restaurant to book an appointment. The demo stunned observers: the voice sounded human, with realistic hesitations, and most importantly — the person on the other end never realized they were speaking to a machine.

Today, these technologies are accessible to businesses of all sizes. Voice agents handle inbound calls, follow up with customers, or schedule appointments — automatically. But a new chapter is beginning: calls with zero human involvement, where two AI agents talk directly to each other. A shift that changes the very nature of what we call a "phone conversation."

An AI ↔ AI call: efficiency, risks, and accountability

Having two voice agents talk to each other may seem like the logical next step. If an AI can talk to a human, why not let it talk to another AI to gain efficiency? But what sounds simple in theory hides much deeper complexity.

A conversation between two AIs is no longer an assistant tool — it’s an autonomous conversational system, with no human in the loop. And as soon as we let two machines converse unsupervised, a range of risks emerges: misunderstandings between models, infinite loops if one doesn’t grasp the other, incorrect decisions like validating an unwanted booking, or interpretation disputes about what was actually said.

Unlike humans, an AI won’t stay silent when uncertain — it keeps going, searches for an answer, acts no matter what. Where a person might pause or ask for clarification, the AI proceeds. That’s its strength — but also its limit. This makes structured efficiency a necessity: clear prompts, precise rules, and fallback mechanisms for when things go off track.

These kinds of conversations require a completely different technical approach. It’s no longer about crafting a human-friendly experience, but about building a machine-to-machine dialogue layer. That means short prompts, explicit decision logic, and — most importantly — optimized communication protocols. Which raises a core question: when two AIs talk, should they even use human speech?

Speech is a medium built for humans: slow, ambiguous, full of intonation, and limited in bandwidth. For machines, it’s inefficient. Researchers are already exploring alternatives where agents exchange not full sentences but compressed data packets — JSON structures, API calls, or even coded sounds that only the models understand. If the call goes over a regular phone network, nothing prevents the data from being translated into synthetic audio optimized for inter-AI exchange. It’s no longer a phone call, but a coded protocol carried by voice.

A striking example of this shift is the GibberLink mode. In February 2025, two engineers showcased this protocol at a hackathon. The idea: let two AI voice agents detect they’re talking to a machine and switch automatically to a language optimized for machines, incomprehensible to humans. No more English — just ultra-fast digital signals, generated using the open-source library GGWave. The result? Conversations were 80% faster than standard spoken dialogue. For AIs, it’s a massive efficiency gain. For us, it’s just noise — gibberish — unless post-processed into transcripts.

Which leads to a critical question: how do we monitor a conversation we can’t listen to? How do we make sure that two AIs communicating in code are doing only what they’re supposed to — and nothing more? Monitoring becomes non-negotiable. We need the ability to audit exchanges, track variables, spot edge cases, and if necessary, interrupt the call.

Some approaches exist — such as generating post-call summaries or reports readable by humans. But this requires robust infrastructure and a clear intention to keep humans in the loop, even if they’re no longer part of the conversation itself.

Then comes the legal angle: who’s responsible when something goes wrong? If an AI triggers an unwanted action — a payment, a cancellation, a binding decision — who’s liable? The company that made the call? The one that built the agent? The tech provider? And what if the client’s own AI agent responded incorrectly — can the user be blamed?

Just like self-driving cars required redefining liability, autonomous conversations will need new accountability frameworks. We’ll need clear rules, safeguards, and traceability to understand what happened in case of dispute.

In short, AI ↔ AI conversations promise major performance gains — but they require a complete cultural shift: new design methods, new monitoring tools, new legal standards. The challenge is technical, but also ethical and societal. Only by setting strong foundations now can we unlock the full potential of this new conversational era.

Short-term use cases

AI ↔ AI calls are no longer just science fiction. Several realistic use cases could emerge within the next 12 to 24 months, especially in sectors where human calls are still widespread, time-consuming… and largely standardizable.

1. Automated appointment scheduling

A personal voice assistant could call the front desk of a medical clinic or hair salon to schedule an appointment. If the reception is also handled by an AI, the entire conversation — checking availability, suggesting time slots, sending confirmation — could happen with zero human intervention in just a few seconds.

2. Logistics coordination between companies

In a supply chain scenario, an AI system in a warehouse could call the AI of a delivery company to confirm a drop-off time or acknowledge receipt. This type of exchange, often handled manually today, could become a fully structured voice-based transaction — with variables like parcel ID and timestamps passed and confirmed automatically.

3. Address confirmation before delivery

In e-commerce, a logistics AI could call a store’s customer service AI to confirm an address, phone number, or delivery window before dispatching a package. This type of pre-check between two agents can reduce failed deliveries, lower return rates, and streamline shipping — all without human involvement.

4. Cross-system technical assistance

A building management or cloud system detects a fault. It triggers an automatic call to the AI agent of a technical support provider. One AI reports the anomaly, the other suggests checks or confirms an open ticket. This AI ↔ AI front-line triage reduces the load on human technicians and ensures faster initial resolution.

Why Rounded is well positioned for this future

If AI-to-AI phone calls are coming, the next question is: which platform is built to handle them reliably, flexibly, and safely? That’s where Rounded stands out.

First, because every voice agent on Rounded is deeply customizable. Users can fine-tune the prompt, choose the LLM, define the voice and tone, select the most accurate transcriber, and — crucially — set guardrails like maximum call duration, fallback conditions, and auto-hangup rules. In a machine-to-machine context where everything must be tightly controlled, this level of precision is a major strength.

Second, Rounded is an API-first platform, designed for seamless integration into complex ecosystems. Using tools and variable management, an agent can send, receive, or transform data in real time. And unlike traditional agents designed for humans, these outputs can be immediately understood by another AI — paving the way for structured, efficient machine-to-machine dialogues.

Finally, all of this is backed by a robust monitoring system. Every call can be reviewed, sorted, or terminated in real time. Transcripts are accessible, variables are tracked, and users retain full control over the agent’s behavior at all times. This makes Rounded particularly suited to powering AI ↔ AI conversations at scale, with the reliability and transparency such use cases demand.

Everything about Rounded — from customization to integration to supervision — makes it a strong foundation for the future of fully automated voice communication.

And beyond? A voice agent for every individual?

If companies begin adopting AI-to-AI calls, what’s next for individuals? It’s easy to imagine a future where every person has their own personal voice agent, much like we all carry smartphones today.

This intelligent assistant could make calls on your behalf for everyday tasks: booking a dentist appointment, negotiating a bill, following up on a late delivery. It would call, explain your request, talk to either a human or another AI — and then report back with a clear summary. You stay hands-free the whole time.

In this future, every customer is represented by their own AI. No more waiting on hold, no more repeating information — your agent handles it all. This isn’t purely speculative: Google’s Duplex already showed a glimpse of this in action. But where Duplex was a closed, tech-demo experience, platforms like Rounded could soon make it accessible to anyone.

The implications are massive. It could become a tool for inclusion — helping those who struggle with speech or phone anxiety. It’s a clear time-saver, and potentially a new layer of client experience, where AI agents interact directly and instantly.

But this also raises new questions: Would you want to be called by someone else’s voice agent? How do we ensure they faithfully represent user intent? Will we need legal frameworks to define what these agents can and cannot do?

One thing is certain: voice remains one of the most universal and powerful interfaces we have — and AI is about to completely transform how we use it. The future of phone calls will likely be faster, smarter, and more automated — but also more impersonal.

Are we ready to hand over our voice to an AI agent? Maybe. What’s clear is that the revolution has already begun.