Why Voice AI Is Moving From Conversational AI to Operational Infrastructure

On May 22, 2026, SlatorCon London hosted a voice AI panel to discuss the current capabilities and limitations of voice AI, as well as enterprise deployments and frontier capabilities in this fast-moving space.

Moderated by Slator’s Head of Consulting, Alex Edwards, the panel included Neil Zeghidour, Co-Founder and CEO at Gradium, Arkadiusz Kwapiszewski, Head of Agent Design & Engineering at PolyAI, and Peadar Coyle, CTO & Co-Founder at AudioStack.

On the topic of voice AI’s current capabilities, PolyAI’s Kwapiszewski argued that enterprise value increasingly comes from voice agents that can complete actions rather than simply answer questions. 

“The real value comes not from deflecting [or] answering questions, but from actually doing something for users,” he said, pointing to deeper integrations with CRMs, ticketing systems, booking platforms, and operational workflows as the key to stronger ROI beyond answering basic FAQs.

AudioStack’s Coyle explained that today, global brands, publishers, and agencies are already deploying AI-generated audio in production across advertising and publishing use cases. According to Coyle, enterprise adoption is real and accelerating, particularly as organizations begin integrating voice AI into broader content production pipelines.

At the same time, Coyle stressed that “AI models alone are not enough,” arguing that successful deployments still require experts in the loop, QA, alignment work, and specialist language expertise to handle cultural nuance and multilingual complexity. He pointed to challenges around lower-resource languages and mixed-language speech patterns such as “Hinglish,” where cultural and linguistic nuances remain difficult for text-to-speech systems to handle reliably.

As for Gradium’s Zeghidour, natural multilingual voice experiences depend heavily on realistic data and culturally appropriate speech patterns. Zeghidour highlighted the importance of regional accents and localized speech styles, particularly in languages such as Spanish and Portuguese, where users expect voices to reflect regional identity. He noted that low-resource and unwritten languages remain a major technical challenge even for the largest AI models.

On Voice AI Infrastructure & Deployment

The panel also focused on the deployment of voice AI solutions, as Slator’s Edwards asked for real use cases of enterprise deployment from the panelists.

Zeghidour argued that while large-scale foundational models can achieve impressive conversational quality, they are often prohibitively expensive to operate at scale. The industry’s real challenge, he said, is building smaller, cheaper models that can deliver high-quality interactions while remaining economically viable for large-scale deployment.

That challenge becomes even more important as enterprises seek to deploy voice AI across millions of interactions. Zeghidour predicted that some future applications, such as real-time live phone translation offered by telecom providers, will require on-device inference rather than API-based architectures. He said Gradium is already developing models capable of running directly on edge devices and smartphones to support these emerging use cases.

PolyAI’s Kwapiszewski described how enterprises are using voice AI to address labor shortages, seasonal demand spikes, and multilingual customer support challenges. Retailers managing Black Friday call surges or restaurants dealing with peak reservation periods can scale AI agents instantly without onboarding additional staff. He also noted measurable customer experience improvements, citing a deployment at Zagrebacka Banka in Croatia that increased NPS scores by 14 points after replacing legacy Interactive Voice Response (IVR) systems.

The panelists also discussed how customer expectations are rapidly evolving alongside the technology itself. Coyle described how every new capability immediately creates demand for even more sophisticated personalization, integrations, and governance safeguards. He warned that increasing integration with enterprise systems also introduces new quality and risk-management challenges that vendors must actively address.

The Future of Voice AI

The conversation concluded with broader reflections on where voice AI may ultimately lead. Zeghidour pointed to robotics as a major future frontier, arguing that voice interaction will become even more critical as humanoid robots enter real-world environments. However, he noted that conversational AI for robotics introduces entirely new scientific problems, including identifying who is speaking, filtering background noise, and understanding conversational intent in noisy physical spaces.

Kwapiszewski emphasized that voice has become the most human-like interface yet for interacting with machines. He noted that users increasingly behave toward AI agents as though they were interacting with real people.

“The real value comes not from deflecting [or] answering questions, but from actually doing something for users.” — Arkadiusz Kwapiszewski, Head of Agent Design & Engineering at PolyAI

Kwapiszewski said, “They thank the agent for the assistance […], they apologize, they say something like, ‘Sorry, I need a minute to find that.’” According to him, that emotional engagement translates directly into stronger business outcomes, including lower abandonment rates and higher customer cooperation.

Taken together, the panel suggested that voice AI is entering a new phase defined by operational deployment, workflow integration, multilingual scalability, and increasingly human-like interaction. While technical hurdles around naturalness, economics, and low-resource languages remain unresolved, the consensus on stage was clear: voice is rapidly becoming one of the primary interfaces between humans and machines.

A recording of the panel will soon be available to Slator subscribers on the Video Library.