Nikola
Mrkšić

Callers using PolyAI's voice agents occasionally do something strange. Mid-sentence, they pause - sometimes during a complaint, sometimes mid-rant about a late delivery or a billing error - and apologize. Not because someone asked them to. But because a small doubt has crept in: what if this is a person? That moment, that flicker of uncertainty, is what Nikola Mrkšić has spent the better part of a decade engineering. Not a trick, not a parlor gimmick. The actual, hard, unsolved problem of building a voice AI that handles real conversations the way a skilled human agent would.

Mrkšić grew up in Belgrade in the early 1990s, when Yugoslavia was coming apart at the seams. He attended the Belgrade Mathematical Gymnasium - a competitive school originally designed during the Tito era to produce mathematicians and scientists. Whatever the curriculum was supposed to do, it worked on him. He developed an early obsession with mathematics and computation, the kind that earns full scholarships to Cambridge University.

At Cambridge, he landed in the Machine Intelligence Lab under Professor Steve Young - himself a founder of voice technology work - and completed a BA, MSc, and PhD on spoken dialogue systems. His dissertation: "Data-driven language understanding for spoken dialogue systems." One mentor, the famous machine learning researcher Zoubin Ghahramani, offered an instruction that Mrkšić has clearly internalized: "Don't just go and become a PowerPoint monkey."

The Apple Detour That Wasn't

Before PolyAI, there was VocalIQ - another Cambridge spinoff working on real-world dialogue models. Mrkšić was its first technical hire, not an academic footnote but the person actually building the thing. When Apple acquired VocalIQ in 2015 to improve Siri, he went with it. For two years he worked on making Siri more conversational from inside one of the world's most secretive companies.

He left in 2017. Not to take a break, but to start over. He co-founded PolyAI in London with two colleagues from Cambridge's Machine Intelligence Lab: Tsung-Hsien Wen, who became CTO, and Pei-Hao Su, who became SVP of Engineering. Three researchers who had studied the same problem for years from a university lab decided to solve it properly, in the market, at enterprise scale.

The original bet was straightforward: voice was the natural interface for customer service. Not chat, not email - voice. People call when things go wrong, when stakes feel high, when they want a human response. The challenge was building a voice AI that could actually handle that complexity - identity verification, multi-step problem resolution, emotional tone, accents, interruptions - without falling apart the moment something unexpected happened.

The Architecture of Trust

PolyAI built a platform that handles the full surface area of a real customer service call. Not a basic IVR that routes calls to the right department. A system that can verify your identity, pull your account details, understand what went wrong, and resolve it in a single conversation - while adapting its voice, accent, and tone to match the brand it's representing.

The proprietary edge is data. PolyAI has accumulated a dataset of human-to-AI customer service conversations with outcomes that no competitor can replicate. Conversations with Marriott guests, Caesars Entertainment callers, FedEx customers. The system learns what resolution actually looks like in each industry, with each type of inquiry. The platform operates in 45+ languages, and is language-agnostic: an agent built in English can be ported to any other language without redevelopment.

UniCredit improved its Net Promoter Score by 14 points after deployment. Forrester calculated that PolyAI customers average $10.3M in savings and a 391% return on investment. At multiple enterprise clients, the system handles the equivalent of 1,000+ full-time employees' workload. And it never calls in sick, never has a bad day, never loses patience on call number 400.

The ChatGPT Vindication

When OpenAI launched ChatGPT in late 2022, it quadrupled PolyAI's inbound leads almost overnight. For most AI companies, that was a surprise. For Mrkšić, it was a vindication - seven years of patient building, finally meeting a market that suddenly understood what voice AI could do.

His position on the channel debate has been consistent and blunt: "Start automating the phone now. Don't distract yourself with chat." The data backs him up. PolyAI's own research shows 86% of Gen Z and younger millennials prefer voice for customer support when something goes wrong. The phone call isn't dying. It's just waiting to be handled better.

His vision for the contact center is a structural reorganization: what he calls a "command center" model where AI handles 90% of inbound queries, and human agents focus exclusively on the conversations that require real judgment. Not a cost-cutting exercise, but a reallocation - where the humans who remain become genuinely more valuable, and the AI stops pretending to be a shortcut.

From Belgrade to $750M

The Series D closed in December 2025 at an $86M raise, co-led by Georgian, Hedosophia, and Khosla Ventures - a remarkable show of consistency from the original backers. The round valued PolyAI at $750M, with total funding crossing $206M. New investors included Citi Ventures, British Business Bank, and NVentures (NVIDIA's venture arm), which had first backed PolyAI in the Series C.

By early 2026, the customer list had grown past 200 enterprise brands - Marriott, FedEx, Caesars Entertainment, PG&E, Volkswagen, UniCredit, Foot Locker, Starling Bank, and Whitbread among them. In March 2026, PolyAI won the Gold Award at the Stevie Awards for Best Customer Support Solution. The 270-person company runs out of a London headquarters with a growing San Francisco presence.

Mrkšić's academic work runs alongside all of this. His Google Scholar profile shows 7,477+ citations - researchers who have built on his work in machine learning and spoken dialogue systems. He was named advisory board member by Serbia's Institute for International Relations, and appeared on Forbes 30 Under 30 Europe in 2021. The research never really stopped; it just moved into a company.