EP58 — Unicorn Files — PolyAI: the European voice AI company rewriting the call center

In this Unicorn Files episode, we tell the story of PolyAI, one of the most interesting European startups in the new wave of enterprise AI applied to customer service.

The company was founded in London in 2017 by Nikola Mrkšić, Pei-Hao “Eddy” Su and Tsung-Hsien “Shawn” Wen, three machine learning researchers from the University of Cambridge. The starting point was not building a chatbot, but tackling a much harder problem: automating real phone conversations, where customers interrupt, switch context, speak with different accents, request escalation, provide sensitive information and expect immediate resolution.

PolyAI positions itself as a dialog-first platform. At the centre is the Agentic Dialog Platform, with tools for CX, operations, product and developer teams, and the proprietary Raven model, trained on more than one billion enterprise conversations. The promise is not simply “AI that speaks”, but voice agents that can complete processes, integrate with enterprise systems and generate measurable value.

The episode explores enterprise customer cases: PG&E, where PolyAI reports 67% containment, 35,000 labour hours saved and a 22% CSAT increase on outage calls; UniCredit/Zagrebačka banka, where a Croatian-speaking voice assistant reduced abandonment and improved NPS; and customers such as Marriott, Caesars, FedEx and Foot Locker.

We also analyse funding: an $86 million Series D announced in December 2025, co-led by Georgian, Hedosophia and Khosla Ventures, with NVentures among the investors. Total funding exceeds $200 million; the $750 million valuation was reported by SiliconANGLE, but was not disclosed in PolyAI’s official announcement.

The final section looks at risks: commoditisation of voice models, pressure from Big Tech and contact-centre incumbents, European regulation, the AI Act, GDPR, DORA, PCI, HIPAA, auditability and vendor lock-in. The playbook for founders and operators is clear: in enterprise voice AI, the winner is not the company that makes the model speak better, but the one that makes it work inside a real enterprise.

• PolyAI should be read as an enterprise voice AI platform, not as a simple talking chatbot.
• Voice is harder than chat because it requires low latency, accent and interruption handling, tone, context and immediate human fallback.
• The proprietary Raven model and the Agentic Dialog Platform represent an attempt to build a dialog-first platform for complex enterprise conversations.
• Customer value is not measured in AI hype, but in containment, labour hours saved, CSAT, reduced abandonment, correct routing and operational ROI.
• PG&E and UniCredit/Zagrebačka banka are useful cases because they show deployment in regulated, high-volume contexts with real reputational risk.
• The $86 million Series D and the entry of NVentures strengthen PolyAI's credibility, but the company remains below unicorn status according to the reported valuation.
• Competitive risk is high: Big Tech, contact-center incumbents and agentic AI startups could compress the advantage if Raven and conversational data do not become a defensible moat.
• The playbook for founders is to start from call type and ROI, not from the model: dataset, operational baseline, human fallback, governance and metrics before AI magic.