ElevenLabs CEO on $22B Valuation, Margins and AI Disclosure

Most people who have spoken to ElevenLabs have no idea they did. The company builds models that turn text into speech that sounds human, and much of that speech now reaches the public through customer service lines. Klarna uses it for first-line phone support covering 35 million U.S. customers. Deutsche Telekom, Cisco, Adobe and a growing number of governments are also customers. A separate part of the business serves creators, who use the platform for audiobooks, dubbing and music.

The company is four years old. It says it is pacing at $600 million in annual recurring revenue, and its backers reportedly value it at $22 billion. In a short on-stage interview with TechCrunch at Nrth, a Toronto entrepreneurship conference formerly called Elevate, co-founder and CEO Mati Staniszewski covered competition, model choice, disclosure and money.

The quality gap is narrowing, slowly

A year earlier, Staniszewski had predicted that audio models would be commoditised within a couple of years. He now sees a longer runway. Model-level quality differences still matter a great deal, he said, and are likely to shrink over three to five years. His stated goal is for ElevenLabs to be the first to pass the Turing test for conversational AI. That, in his view, needs emotional intelligence as well as raw intelligence: reading the other person's mood and knowing when to slow down or speak up.

More than 55% of revenue comes from classic enterprise customers. Much of the remaining 45% comes from small and medium businesses, developers, builders and creators.

Customers who become rivals

Like many AI suppliers, ElevenLabs increasingly meets its own customers in the market. Decagon, a conversational AI platform, trained its voice product on ElevenLabs and now competes with it. Staniszewski said the old boundaries between model, platform and application companies are blurring. He pointed to Anthropic, which started as a model company and now runs a platform and a widening set of applications. He expects the trend to continue.

Picking the reasoning layer

ElevenLabs lets customers select the "reasoning layer" behind their voice agents from a menu of options. According to Staniszewski, the choice between frontier and open-weight models depends on the task. An informational call, where no action is executed and the knowledge base defines a good answer, can often run on open-source models. A financial services call that involves authentication, transaction details or a refund leaves no room for error. There, he said, frontier models still lead.

Government work adds more requirements. Deployments may use open-weight, closed-source or a government's own fine-tuned models. In Poland, the use case is healthcare: 18% of patients who book appointments in the public health system never show up. Agents now call to remind them, built on the government's own optimised models, with data residency preserved.

Should callers know it is a machine?

Staniszewski supports disclosure for now. People are not yet used to AI agents on the phone and do not want to feel cheated. He expects that to change within five years, once people have agents acting for them and expect an agent on the other end. His suggested approach: when the wait for a human is 30 minutes, offer a choice. In almost all cases, he said, callers pick the agent and are surprised by the experience.

Margins, data and an IPO

On gross margins he declined to share figures. He said research lets the company fine-tune and constrain models efficiently, and savings are passed to customers where possible. He is willing to accept lower margins if that proves value and grows market share over the next five years.

On training, he said volume mattered less than annotation. Thousands of contractors label not only what was said but when, how and with what emotion. Voice coaches were brought in to detect accents accurately. Some models were built jointly with specific customers.

Asked about a reported 2028 IPO, he said the company is laying the foundation to go public "in the next years", with timing still open.

Backstage, he said labs broadly agree on pacing and precautions. ElevenLabs does not train text models, and its technology does not let agents create more agents. Every customer passes KYC checks. He acknowledged cybersecurity risk as a wider concern but said the company has solid precautions in place.