At its core, ElevenLabs does one thing exceptionally well: it converts written text into speech that is nearly indistinguishable from a human voice. Most people interact with its technology without ever knowing it. When customers call Klarna's phone support line - one that serves 35 million people across the United States - they are often speaking with a voice powered by ElevenLabs. The same is true for Deutsche Telekom, Cisco, Adobe, and a growing number of government agencies around the world. Beyond enterprise clients, the platform also serves individual creators who use it to produce audiobooks, dubbed content, and music.

The competitive landscape is getting more complicated. ElevenLabs is increasingly finding itself in the same market as companies it once served. Decagon, a conversational AI platform, originally trained its voice product using ElevenLabs technology and now competes directly with it. Even so, investor confidence in the startup remains strong. The company reports it is on pace to reach $600 million in annual recurring revenue, and according to recent reports, it has been valued at $22 billion by its backers - a striking figure for a company that is only four years old.

To get a closer look at where the company stands and where it is headed, ElevenLabs co-founder and CEO Mati Staniszewski sat down for an interview at Nrth, a Toronto-based entrepreneurship conference previously known as Elevate. The conversation touched on a wide range of subjects, from AI disclosure ethics to gross margins, IPO timelines, and the thorny question of what happens when a startup's customers become its competitors. Below is a condensed and lightly edited version of that exchange.

On the Commoditization of Audio Models

At a major industry conference last year, Staniszewski predicted that audio models would be commoditized within a couple of years. Asked how he would assess that prediction today, he acknowledged that significant technical differentiation still exists at the model level, but he expects that gap to narrow over the next three to five years.

What he sees as the real frontier, however, is something more ambitious: passing the Turing test for conversational AI. In his view, achieving that milestone requires more than raw intelligence. It demands emotional intelligence - the ability to detect what the person on the other end of the call is feeling and adjust accordingly, whether that means speaking more slowly or raising the energy of the conversation. That level of nuance, he said, has not yet been achieved by anyone in the field.

Breaking Down the Revenue Mix

With the company now reporting $600 million in annual recurring revenue, Staniszewski offered a breakdown of where that money is coming from. More than 55 percent comes from classic enterprise clients - large organizations with significant and ongoing voice AI needs. The remaining 45 percent is largely made up of small and medium-sized businesses, independent developers, software builders, and content creators.

When Customers Become Competitors

The situation with Decagon - a company that built its voice product on ElevenLabs and has since moved its queries to its own internal models - is part of a broader trend Staniszewski sees reshaping the AI industry. The boundaries between model companies, platform companies, and application companies, once clearly defined, are becoming increasingly blurred.

He pointed to Anthropic as a case in point. What started as a model-focused company has expanded into a platform and is now building out a wide range of applications. Staniszewski expects this pattern to continue across the industry, with companies at every layer of the stack gradually expanding into adjacent territory.

Frontier Models Versus Open-Weight Models

ElevenLabs gives its customers the ability to choose which reasoning layer - essentially which underlying AI model - powers their voice applications. Asked about what he is seeing in terms of how customers are choosing between frontier lab models and open-weight alternatives, Staniszewski pushed back on the idea that it is a binary decision.

For lower-stakes, purely informational interactions - a customer calling in to ask a simple question, for example - open-source models can perform well because the quality of the experience is largely determined by the knowledge base behind it. But in high-stakes environments like financial services, where customers need to be authenticated and may be requesting sensitive information about transactions or refunds, there is no margin for error. In those cases, frontier models remain the preferred choice.

Navigating Government Contracts and Open-Weight Chinese Models

Given that some of the most capable open-weight models in circulation originate from Chinese developers, and given that ElevenLabs counts both U.S. and European government agencies among its clients, the question of how those conversations unfold is a sensitive one. Staniszewski acknowledged that each deployment is handled differently, with model and voice selection tailored to the specific requirements of each client and jurisdiction.

He described one deployment with the Polish government as an illustrative example. The use case involves the country's public healthcare system, where patients booking appointments through a national platform fail to show up 18 percent of the time. ElevenLabs deployed agents that call patients ahead of their appointments to issue reminders. The system runs on models that have been optimized for that specific use case, with data residency requirements maintained throughout. The Brazilian government, he noted, has its own distinct set of requirements, and the approach there differs accordingly.

Should Businesses Disclose When a Customer Is Talking to an AI?

On the question of transparency - whether companies should be upfront with customers when they are speaking with an AI agent rather than a human - Staniszewski was direct. He believes disclosure is the right approach, at least for now. His reasoning is grounded in where society currently stands: most people are not yet accustomed to interacting with AI agents, and discovering mid-conversation that they have been speaking with one can feel like a form of deception.

That said, he expects the social norm to shift over time. In five years, he suggested, the expectation may well be reversed - with customers assuming they are speaking to an agent unless told otherwise, because they themselves will have their own AI agents acting on their behalf in other contexts.

He also offered a practical approach for businesses navigating the transition now: if a customer faces a long wait for a human representative, give them the option to speak with an agent instead. In his experience, customers almost always choose the agent - and are frequently surprised by how capable the interaction turns out to be.

Gross Margins and the Willingness to Sacrifice Them

Asked about the company's gross margins - a topic of considerable interest given the costs associated with running large-scale AI inference - Staniszewski declined to offer specifics, as expected. But he was candid about the company's philosophy. Because ElevenLabs conducts its own research, it has developed the ability to fine-tune and constrain models in ways that improve efficiency. When those efficiencies translate into cost savings, the company passes them on to customers rather than pocketing them.

More broadly, he indicated that ElevenLabs is not focused on protecting its margins in the near term. If investing in customer relationships and demonstrating the value of the technology means accepting lower margins for now, that is a trade-off the company is prepared to make. The logic is that the real value creation will happen over the next several years, and being deeply embedded with customers during that period is worth more than short-term profitability.

Training Data: Synthetic and Human-Annotated

With millions of hours of customer service call data flowing through its platform, the question of whether ElevenLabs trains on that data - and how much of its training data is synthetically generated - is a significant one. Staniszewski explained that in some cases, the company has worked directly with enterprise clients to build models tailored to their specific use cases, developed collaboratively.

Beyond that, he noted that the volume of data has been less important to the company's progress than the quality of its annotation. ElevenLabs employs thousands of contractors who are responsible for labeling not just what was said in a given audio sample, but how it was said - the timing of speech, the emotional tone, the specific accent. To get that annotation right, the company has brought in voice coaches who can help identify and accurately characterize subtle vocal characteristics.

IPO Timeline

Reports have suggested that ElevenLabs is targeting 2028 for a potential initial public offering. Asked to confirm that timeline, Staniszewski offered a carefully worded non-answer. The company is building what he described as a foundation capable of supporting a public listing in the coming years, but whether and when it actually moves forward with one will depend on the circumstances at the time.

Pressed on the vagueness of the word "years," he responded with a laugh and nothing more.

On Whether Frontier AI Labs Should Slow Down

In a more candid, off-the-record portion of the conversation, Staniszewski was asked where he stands on the debate over whether the leading AI labs should reduce the pace of their development. His response suggested a degree of alignment around the idea that some form of coordinated pacing is appropriate, though he drew a distinction between the underlying technical question and the public and regulatory conversation surrounding it.

He emphasized that ElevenLabs does not develop the text models or the core intelligence components that sit at the center of that debate. Its focus is on the voice and audio layer, which he positioned as a step removed from the most contentious aspects of the frontier AI discussion.

Security Risks and the Hugging Face Comparison

Asked whether ElevenLabs could face the kind of security exposure that Hugging Face encountered - a reference to vulnerabilities that have surfaced on the open AI model-sharing platform - Staniszewski argued that the company's architecture makes it a different kind of target.

The key distinction, in his telling, is that ElevenLabs does not deploy self-replicating or recurrent intelligence components. Its technology is not designed in a way that would allow AI agents to create additional agents autonomously. Every customer using the platform goes through a Know Your Customer verification process. He acknowledged that cybersecurity risks are a genuine concern for the broader AI industry, but expressed confidence that the company has put appropriate safeguards in place to manage its specific exposure.