ElevenLabs, a four-year-old company building the voice layer of AI, is now reportedly valued at $22 billion by investors, with annual recurring revenue pacing at $600 million. In an interview at the Nrth conference in Toronto, co-founder and CEO Mati Staniszewski discussed the company’s strategy, competition, and the future of conversational AI.

Staniszewski acknowledged that while audio models themselves will become commoditized over the next three to five years, there is still significant quality to be gained at the model level today. He emphasized the company’s goal to be the first to pass the Turing test for conversational AI, which requires combining intelligence with emotional intelligence—understanding the other party’s emotions and adjusting pace and tone.

On enterprise adoption, Staniszewski revealed that 55% of the $600 million ARR comes from classic enterprise clients, with the remainder from SMBs, developers, builders, and creators. He noted that the lines between model, platform, and application companies are blurring, citing Anthropic as an example, and that competing with customers is an accepted part of the evolving landscape.

Regarding model choices, he explained that open-weight models work well for informational customer service where a knowledge base defines good experiences, while frontier models are essential for financial services requiring authentication and error-free transactions. He also addressed government clients, noting that deployments are customized per country, such as a Polish healthcare project where AI agents call patients to remind them of appointments, reducing no-show rates from 18%.

On disclosure, Staniszewski believes businesses should currently inform customers when they are interacting with an AI, but predicts this will shift in five years as agents become the norm. He suggested offering agents as a choice when human wait times are long, as most customers are pleasantly surprised by the quality.

When pressed on gross margins, he declined to provide specifics but indicated the company is willing to accept lower margins to prove value and expand market share, focusing on long-term value creation. On training data, he clarified that the key wasn’t volume but annotation—thousands of contractors, including voice coaches, annotate speech for content, timing, and emotion.

Regarding an IPO, Staniszewski confirmed the company is building foundations for one, potentially by 2028, but emphasized it depends on timing and place. On AI safety, he supports pacing and precautions but believes the debate should be handled thoughtfully. He distinguished ElevenLabs from Hugging Face, stating they don’t deploy self-replicating agents and have KYC for every customer, mitigating certain exposure risks.