Vobiz.ai is AI-native, API-first telephony infrastructure that sits underneath the voice AI stack at the layer that carries every AI conversation. In an exclusive interaction, Suman Gandham, Co-founder and CEO, Vobiz.ai predicts to Rajneesh De, Group Editor, APAC Media & CXO Media that over the next 18 to 24 months, voice AI grade latency would gradually become table stakes, observability would become a procurement requirement, and voice would fold into multimodal orchestration.
Could you briefly introduce Vobiz.ai and the problem it is solving in the voice AI ecosystem? What are the key pillars of the company’s strategy as it scales its AI-native telephony infrastructure?
Vobiz.ai is AI-native, API-first telephony infrastructure. We sit underneath the voice AI stack — not at the model layer, but at the layer that carries every AI conversation.
The problem we solve is a mismatch of eras. Telephony was engineered for humans calling humans, through contact-centre architectures of the 2000s. Legacy PSTN routing adds 400 to 600 milliseconds before the AI has done any work at all. Add speech-to-text, inference and text-to-speech, and you are at 1.2 to 1.5 seconds, a point where the conversation stops feeling normal and the customer drops.
Four pillars guide how we scale. Sub-80 ms telephony latency is an architectural commitment in how we’ve built the Vobiz stack—single-hop, and event-driven. Developer experience is a primitive: self-serve, frictionless KYC, numbers and deployment in minutes, not weeks of paperwork. Compliance built in rather than bolted on, with TRAI, DLT and consent handling automated. And model neutrality; we route across providers on latency and language, because we are not betting on which model wins.
What needs to change at the underlying telecom infrastructure layer for Voice AI to move from experimentation to enterprise deployment?
There are changes that will define how telecom infra can be built for AI. The media path has to stop being a pipe and become programmable – continuous streaming, barge-in handled at the media layer, and software able to intervene mid-call.
Audio conditioning has to move into the network, ahead of the model. Real-world calls are noisy, and inference struggles to isolate and understand a customer’s voice when it arrives mixed with traffic sounds and other human voices.
Two things need collective effort at an industry level: carrier engineering has to treat machine-originated traffic as a distinct class, and regulation drafted for human telemarketing needs answers on how an AI caller identifies itself. Both conversations are only starting.
Over the next 18 to 24 months, I expect voice AI grade latency to become table stakes, observability to become a procurement requirement, and voice to fold into multimodal orchestration. This is what the infrastructure needs to prepare for.
What changes when an enterprise moves from a few hundred AI-assisted calls to thousands or millions of concurrent voice interactions?
Almost everything that is negligible or invisible at pilot scale becomes critical.
Averages stop meaning anything. At a few hundred calls, average latency is a reasonable way to think. At a million, the tail is what customers actually experience, and even a small percentage means thousands of people saying your agent is broken.
It also drives the need for observability as a primitive for operators. At two percent degradation across a million daily calls, most teams discover that they have no visibility into where latency enters or why conversations fail.
Capacity behaves differently. A contact centre has fixed headcount, so volume has a ceiling and traffic is human-shaped. AI traffic is bursty and elastic — trunks and media servers have to scale on demand. We validated that going from 100,000 calls a week to over a million a day while holding latency flat.
And compliance exposure grows non-linearly. One mishandled consent flow in a pilot is an incident. The same flow at national scale is a regulatory event.
For enterprises deploying AI voice agents, what does “performance” actually mean — latency, call quality, uptime, completion rates, or something else?
Latency, quality and uptime are necessary conditions, not the measure of success. An enterprise does not buy eighty milliseconds. It buys resolved customer problems.
It helps to think in layers. At the infrastructure layer you measure P95 and P99 latency, MOS, jitter and connection success. At the conversation layer: completion rate, containment, interruption handling, whether context held across a long call. At the business layer, where the CFO lives: cost per successful interaction, resolution rate, compliance adherence.
What has genuinely shifted this past year is that buyers have moved down that list. Eighteen months ago, the question was whether it sounded human. Now it is cost per resolved conversation. But that is not to say that the first layer of metrics is becoming unimportant. You cannot get containment if packet loss dropped a syllable and the model mishears.
Could AI agents eventually become one of the biggest drivers of change in telecom infrastructure, in the same way cloud workloads reshaped data-centre infrastructure?
That is the right analogy, and I would push it further.
Cloud reshaped data centres because the workload broke the infrastructure’s assumptions. Elasticity, API provisioning, pay-per-use were designed to meet a workload that behaved very differently from the traditional ones. Similarly, voice AI traffic is machine-scale, bursty, latency-intolerant and needs software to intervene mid-call.
Legacy telephony is not designed for any of these.
The signal I would point to is the pace gap. Cloud workloads could tolerate a slow server; things just took longer. AI voice cannot tolerate a slow rail, because the conversation itself breaks. Speech and language models have improved every quarter for three years.
Telephony latency has not moved. That gap cannot hold while volume climbs.
What are the key pillars of Vobiz.ai’s growth strategy? How does the company’s ecosystem of telecom, technology, and enterprise partnerships come together to scale voice AI adoption?
Four pillars.
We own the infrastructure rather than orchestrate on top of it — carrier interconnects, SIP trunking, media servers, routing, compliance. Platforms built on third-party telephony cannot control latency, guarantee reliability at volume or own compliance in regulated markets.
We learned that limitation firsthand at the application layer.
We are moving from connectivity to intelligence. Revenue comes from numbers, usage billing, and value-added services like transcription, streaming and answering-machine detection. Voice infrastructure will differentiate on observability and optimisation, not minutes.
We are international by design and architected for global connectivity from day one. We’ve zeroed in on India first to solve local compliance and number provisioning because it is the hardest regulatory environment to get right. We already run volumes in several other markets and are focused on international expansion.
Finally, we serve builders before buyers. Carrier depth without developer experience is legacy telecom.
What are your unique differentiators over competitors like Exotel?
I would frame it as generational rather than competitive. Exotel built a genuinely strong business solving the problem of its era – connecting humans reliably at scale for contact centres. Much of India’s customer communication still runs on that generation of infrastructure.
Our starting assumption is different. We built for a caller that is software, not human. Everything else I have spoken about follows from there. And our adoption numbers are the proof: 100,000 calls a month in January to over five million a day now.
