ElevenLabs CEO Says Annual Recurring Revenue Has Reached $600 Million

Last year, ElevenLabs CEO Mati Staniszewski predicted audio models would become commodities within a couple of years. Now he says their quality still differs substantially. In a new TechCrunch interview, he also put his company at $600 million in annual recurring revenue, with large businesses supplying most of its business.

The model gap has not closed

Staniszewski has not abandoned his prediction that voice models will grow more alike. But he now places the narrowing of their differences three to five years out. For the moment, he said, the model alone can still produce a meaningful quality advantage. That matters to ElevenLabs: if speech generation becomes interchangeable, customers have less reason to choose its voice technology over a rival’s.

His longer-term goal is a conversational agent that can respond to how someone feels, including when to slow down or speak up. He said that has not yet been achieved. The work behind a convincing voice is not just a matter of gathering recordings: Staniszewski said thousands of contractors help label when and how people speak and what emotions they convey. ElevenLabs also brought in voice coaches to help identify accents.

$600 millionAnnual recurring revenue

Staniszewski described ElevenLabs as having $600 million in annual recurring revenue.

More than 55%Large-enterprise share

Staniszewski said what he calls classic enterprise customers account for more than 55% of the business.

Growth without a margin figure

The $600 million figure is the CEO’s recurring-revenue measure, not a disclosed full year of sales. TechCrunch also reports that ElevenLabs’ backers value the four-year-old company at $22 billion. Neither number answers how much the business keeps after delivering its services. Asked about gross margins, Staniszewski declined to give a detailed figure.

He said ElevenLabs would pass savings from more efficient models to customers and would accept lower margins to prove its value. That sets a clear priority—winning and keeping customers—but leaves the cost of that strategy hard to judge from the figures he shared.

When a customer becomes a rival

Decagon, an ElevenLabs customer, trained its voice product on the company’s technology and now runs queries through its own models, TechCrunch reports. Decagon now competes with ElevenLabs in voice AI. Asked about that overlap, Staniszewski said the old divisions between model makers, platforms and applications are becoming less clear.

The example makes his model-quality forecast more consequential. ElevenLabs can sell the technology behind another company’s voice product, yet that company can also develop a substitute. Staniszewski did not put a financial figure on the Decagon relationship in the interview, so the competitive threat is clearer than its effect on revenue.

The calls behind the business

Customer service shows where ElevenLabs is finding demand. Klarna uses its technology for first-line phone support serving 35 million U.S. customers, TechCrunch reports. In Poland, Staniszewski described an agent deployment that calls patients to remind them about public-health appointments. He said 18% of patients miss those appointments; he did not give a result showing how much the reminders reduce that rate.

ElevenLabs does not insist on one AI model for every conversation. Staniszewski said an informational call may work with an open-weight model—a model whose parameters are available to others—when the customer’s knowledge base supplies the answers. For financial-service calls involving authentication, transactions or refunds, he favors more advanced frontier models because errors carry higher stakes. He described the Polish deployment as using models adapted to local needs while keeping data in the required location.

There is also a question for the person answering the phone. Staniszewski thinks businesses should disclose when a caller is speaking with an AI agent, at least while people are still getting used to them. He suggested offering a choice between an agent and a longer wait for a human. It is a customer-facing judgment alongside the commercial one: making automated calls useful without making callers feel misled.