Give Your Chat Agent a Voice — Luke Harries, Head of Growth, ElevenLabs

Luke Harries, Head of Growth at ElevenLabs, presents the case for adding a voice layer to existing chat agents, framing voice as the natural next upgrade for teams that have already invested in building and tuning their conversational AI products. The talk opens by noting that 2025 was the defining year for chat agents — citing viral examples like Linear and PostHog redesigning their home screens around chat interfaces, and even gov.uk adopting chat as the default interaction model. While acknowledging that chat has become the standard AI entry point, Harries argues it already feels dated compared to what voice enables: faster interaction, broader accessibility for users with dyslexia or motor impairments, and entirely new interaction paradigms such as phone lines, Zoom call participation, screen readers, and ambient interfaces.

Harries identifies the core engineering problem that ElevenLabs set out to solve: most teams have already built their chat agents, run evaluations, and integrated RAG and tool calling into their pipelines. The conventional approach to voice — standing up a full voice engine with speech-to-text, turn-taking, text-to-speech, LLM orchestration, and integrations from scratch — requires rebuilding infrastructure that already exists. This was the gap ElevenLabs found when working with enterprise customers like Revolut's customer support team: many had sophisticated agents in production and were unwilling to scrap them.

The solution previewed in the talk is the ElevenLabs Voice Engine, described as an early-access product launching within weeks of the conference. Voice Engine is architected as a standalone primitive that wraps any existing chat agent rather than replacing it. It bundles ElevenLabs' best-in-class models: Scribe for speech-to-text (described as the most accurate model available), V3 for text-to-speech, and an emotion-context-aware turn-taking system that can detect pauses and apply semantic barge-in logic for natural interruption handling. The engine exposes a server SDK in which developers create a client, instantiate the Voice Engine, and attach a thin wrapper to their existing agent — the wrapper proxies sessions through to the underlying chat agent without modifying its logic. A separate client SDK reduces front-end integration to roughly three lines of code and automatically unlocks telephony and CCaaS integrations once deployed. ElevenLabs also ships a library of Shadcn/Vercel-styled UI components that can be dropped directly into projects.

A live demo shows a working text-based chat support agent being converted to a voice agent in a single prompt using a coding agent skill. The prompt analyzes the codebase, identifies the chat agent, and generates the wrapping code automatically. Tool calling is handled natively: since tool calls are already wired into the underlying chat agent, they pass through the Voice Engine wrapper without additional configuration. ElevenLabs also supports client-side tools that can manipulate the DOM directly from the front end, with plans to add proxying of tool calls to the wrapped agent for more complex scenarios.

Harries concludes with a prediction: chat agents that do not add voice will be left behind, just as products that did not add chat in 2025 were. He frames the talk as a call to action, noting that ElevenLabs is actively seeking design partners among the engineering audience to be among the first to deploy Voice Engine integrations in production.