ElevenLabs: Flexible Margins and Uncertain IPO

The head of ElevenLabs claims $600 million in annual recurring revenue and a clientele ranging from Klarna to governments. He defends margins that are ready to erode if customer value is proven and advocates for the disclosure of AI agents during calls. On the security front, the company asserts that it does not enable self-replication of agents and verifies the identity of each client, while preparing the groundwork for a potential IPO at an unspecified date.

ElevenLabs Rules Out Self-Replication and Verifies Each Client

Mati Staniszewski states that ElevenLabs does not deploy self-replicating or recurring components of agent intelligence, and that its technology does not allow agents to create others. He specifies that each client undergoes an identity verification process. While describing cybersecurity as a global risk, he indicates that the company has put precautions in place. Regarding the pace of work in advanced labs, he believes there is alignment to find regulatory mechanisms, without deciding on the publicity to be given or the extent of media or regulatory involvement. He adds that ElevenLabs does not train the text part or the "intelligence" at the heart of the debate on a slowdown.

Revenue, Business Mix, and Margins Ready to Decline

According to ElevenLabs, the annual recurring revenue reaches $600 million, with over 55% coming from contracts with businesses, while the majority of the remaining 45% is generated by SMEs, developers, builders, and creators. Mati Staniszewski emphasizes that the refinement and constraints of the models rely on research, and specifies that the savings achieved are passed on to clients. He states he is willing to accept a decline in margins over the next five years if value is demonstrated to clients, while refusing to detail current gross margins. Regarding a potential IPO mentioned for 2028, he explains that the company is laying the groundwork to be able to do so in the coming years, emphasizing that the decision will depend on timing and location.

Deployments: Telecoms, Government Agencies, and Health Cases in Poland

The voice layer of ElevenLabs is already equipping customer service journeys, such as with Klarna for 35 million clients in the United States, and is used by Deutsche Telekom, Cisco, and Adobe. Government agencies are also part of the clientele, including in the United States and Europe. In Poland, where 18% of patients do not show up for their appointments in the public system, a system of agents that call and remind has been implemented, with models optimized for local knowledge and integration respecting data residency. Beyond customer service uses, the platform is also utilized by creators for audiobooks, dubbing, and music.

Blurred Boundaries with Clients and Model Choices

The boundaries between model providers, platforms, and application publishers are becoming increasingly blurred, observes Mati Staniszewski, citing Anthropic, initially perceived as a model player and now a platform with a range of applications. This porosity is also illustrated by cases of competition with former clients, such as Decagon. ElevenLabs offers a choice of reasoning layers: for informative interactions without action, open-source models may be preferred, while for financial operations including authentication, transaction information, or refunds, Mati Staniszewski believes advanced models are necessary. He also notes that some open-weight models used may be Chinese.

Technical Ambition and Training Method

For the CEO, quality gains remain accessible at the audio model level, even as he anticipates a narrowing of gaps in three to five years. ElevenLabs aims to have conversational AI pass a Turing test, which implies marrying intelligence with emotional intelligence, particularly the ability to perceive emotions to adjust pace or volume, a milestone he believes has not yet been reached. On the data side, some models have been designed with companies for specific use cases. The focus has been more on fine annotation than on volume, with thousands of people mobilized on a contractual basis to delineate content, timing, prosody, and emotions, and the use of voice coaches to refine accent detection.

Informing the Caller When the Agent is an AI, with Choice Offered

Mati Staniszewski advocates for the disclosure, at this stage, of the use of an AI agent during a call, noting that users are not yet accustomed to it and do not want to feel deceived. He proposes offering the choice when the wait for a human reaches, for example, 30 minutes; according to him, in almost all cases, clients opt for the agent and are surprised by the quality. He anticipates that within five years, everyone will have their own agent and will expect to interact with an AI agent.