ElevenLabs says it is pacing at $600 million in annual recurring revenue
ElevenLabs says it is pacing at $600 million in annual recurring revenue, while its backers have reportedly assigned the voice AI company a valuation of $22 billion. The four-year-old company builds models that convert text into human-sounding speech and sells them to enterprises, developers, small and medium-sized businesses, and creators.
Co-founder and CEO Mati Staniszewski said more than 55% of revenue comes from what he described as classic enterprise. ElevenLabs is used for first-line phone support by Klarna for 35 million U.S. customers, and its customer list also includes Deutsche Telekom, Cisco and Adobe. Its platform is also used for audiobooks, dubbing and music.
Voice quality remains a competitive variable
Staniszewski said the gap in audio-model quality remains meaningful, despite his earlier expectation that audio models would become commoditized. Over a longer horizon of three to five years, he expects the differences to become smaller. ElevenLabs aims to help conversational AI pass a Turing test that includes not only intelligence but emotional intelligence.
That requires systems to detect the emotion of the person on the other end of a conversation and adjust their delivery, such as slowing down or speaking up. Staniszewski said this has not yet been achieved. He also acknowledged that the boundaries between model providers, platforms and application companies are becoming less distinct as firms expand their offerings.
Model choice depends on the customer interaction
ElevenLabs lets customers select a reasoning layer from a menu of options. Staniszewski said the decision is not simply a choice between frontier and open-weight models. For informational customer-service calls without actions, open-source models can be appropriate when a company knowledge base defines the desired experience.
Financial-services interactions have different requirements. Authentication, transaction information and refunds leave little room for error, he said, and frontier models are expected to lead in those settings. Government deployments are also configured case by case, with open-weight, closed-source or fine-tuned models chosen according to the customer’s requirements.
In Poland, ElevenLabs is involved in a healthcare deployment for public-system appointment reminders. Staniszewski said 18% of patients do not attend appointments, and agents call patients to remind them. The implementation integrates with models optimized on the relevant knowledge while maintaining data residency.
Disclosure, margins and safeguards
Staniszewski said businesses should currently disclose when a caller is interacting with an AI agent rather than a human. He suggested customers could be offered a choice when a human wait is long, arguing that many opt for an agent and are then surprised by the experience. He expects social expectations to change as people increasingly use agents on their own behalf.
He declined to provide detailed gross-margin figures, but said ElevenLabs is prepared to pass savings to customers and accept lower margins when that helps demonstrate value. The company uses annotation work to improve its models, including analysis of what was said, when speakers talked, how they spoke and the emotions used; it has also used voice coaches to identify accents accurately.
On safety, Staniszewski said ElevenLabs does not train text models or deploy technology that allows agents to create more agents. Every customer goes through KYC, he said, while cybersecurity remains a broader risk. For businesses, the practical implication is to match voice-AI deployment to the cost of mistakes, local data requirements and transparent customer communication rather than treating every call flow as the same technical problem.