ElevenLabs CEO discusses enterprise growth, model commoditization, and disclosure of AI voice use

ElevenLabs, a developer of AI voice models used in customer service, audiobooks, dubbing, and music, reported pacing at $600 million in annual recurring revenue, with more than 55% derived from enterprise customers. The company also noted that small and medium businesses, developers, and creators make up the remainder of its revenue base.

During an interview at a Toronto entrepreneurship conference, ElevenLabs co-founder and CEO Mati Staniszewski said commoditization of audio models remains years away despite rapid advances in model quality. He emphasized that achieving human-level conversational AI will require combining intelligence with emotional intelligence, a milestone he does not expect within the next three to five years.

The company’s customer base includes Klarna, Deutsche Telekom, Cisco, Adobe, and multiple governments, with deployments tailored to local requirements. For example, ElevenLabs integrated models and voices for a Polish healthcare system deployment that reminds patients of appointments, using models optimized for the use case while maintaining data residency.

Staniszewski acknowledged increasing competition from customers who train on ElevenLabs’ models and then compete directly, such as Decagon, a conversational AI platform. He described the blurring lines between model providers, platform companies, and application builders as a broader trend across the AI industry, citing Anthropic’s expansion from a model company into a platform and application provider.

On the question of whether businesses should disclose when customers interact with AI voices, Staniszewski said disclosure is appropriate now, given current user expectations, but predicted societal norms will shift as AI agents become commonplace. He suggested offering customers a choice between human wait times and AI agents, noting that most choose the AI agent once they experience it.

ElevenLabs serves U.S. and European government customers, including those with requirements for open-weight or closed-source models depending on deployment context. The company also uses extensive data annotation, including voice coaching to accurately detect accents and emotions, rather than relying solely on large volumes of synthetic training data.

The CEO declined to provide detailed gross margin figures but indicated the company prioritizes expanding market share and passing savings to customers over short-term margin preservation. He also stated the company is preparing infrastructure to support a potential initial public offering, though timing remains unspecified beyond “the next years.”