ElevenLabs is now reportedly valued at $22 billion by its investors, a figure reached just four years after the company began operations.
The business builds the voice layer for artificial intelligence, supplying the models that convert text into speech sounding human. Klarna runs first-line phone support for 35 million U.S. customers on the platform. Deutsche Telekom, Cisco, Adobe, and a growing list of governments also rely on it. Creators use the system for audiobooks, dubbing, and music production.
Competition exists. Decagon, a conversational AI platform, trained its voice product on ElevenLabs and now competes directly with it. Despite this pressure, backers remain unconcerned. The company reports $600 million in annual recurring revenue.
Valuation and margins
To understand the situation, I interviewed Mati Staniszewski, co-founder and CEO of ElevenLabs, at Nrth in Toronto. The conference is a local entrepreneurship event formerly known as Elevate. We discussed whether businesses should inform customers when they are speaking to an AI, and whether Staniszewski could discuss gross margins.
He declined to provide detailed figures but stated he does not mind margins getting squeezed further if it means expanding the company’s market share.
Audio quality
Staniszewski joined TechCrunch Disrupt last year, where he predicted audio models would become commoditised within a couple of years. He views that prediction differently now.
There is still a lot of work to be done. The quality delta achievable just on the model level remains significant. In the longer term, perhaps three to five years from now, those differences will be smaller.
The company aims to pass the Turing test for conversational AI. This requires combining intelligence with emotional intelligence. The system must understand the emotions of the other party to know when to slow down or speak up. That capability has not yet been achieved.
Enterprise revenue
Current revenue breaks down as follows:
- Classic enterprise accounts: 55% plus
- Small and medium businesses, developers, builders, and creators: the remaining 45%
Competition with customers
The lines between model companies, platform companies, and application companies have become increasingly blurry. In the past, these categories had clear splits. Today, that distinction is much less defined.
Anthropic illustrates this trend. What began as a model company is now definitely a platform and increasingly a wide set of applications. Staniszewski believes this pattern will continue.
Model choices
Customers can select a reasoning layer from a menu of options at ElevenLabs. The choice between frontier lab models and open-weight models is less binary than it appears.
In customer experience scenarios involving simple information exchange, open source models suffice because the knowledge base defines a good experience. However, in financial services, authentication and transaction details are required. There is no room for error here, so frontier models will still lead.
Government deployments
Some open-weight models originate from China. The U.S. government and European governments are also customers. Conversations regarding these deployments differ based on the specific case.
Models and voices deployed depend on the situation. Working with the Polish or Brazilian governments involves their own set of requirements. Options include an open-weight model, a closed-source model, or a fine-tuned model created by the government.
In Poland, the deployment involved a healthcare case. Patients booking appointments across the public health system had an 18% no-show rate. Agents now call to remind them. The system used models optimised on local knowledge, integrated while keeping data residency.
Disclosure
Staniszewski believes businesses should disclose when a human is speaking to an agent rather than another person.
Currently, people are not used to this. The common pattern is that callers do not want to feel cheated. In five years, when everyone has their own agent working on their behalf, callers will expect an agent. Society will then shift.
Good methods exist for implementation. If there is a 30-minute wait for a human, offer the customer a choice. In almost all cases, they choose the agent and are surprised by how good the experience is.
Gross margins
Staniszewski offered a vague answer regarding margins given the cost of models and inference.
The company has a research element that allows it to fine-tune and constrain models in extremely smart ways. If savings can be passed to the customer, that happens. The biggest priority remains proving value and being present with the customer.
If the company can invest and prove that value, it does not mind margins going lower to benefit together as value is created in the next five years.
Training data
ElevenLabs has access to millions of hours of customer service calls. In certain companies, the company and the client created the models together to meet specific use cases.
Otherwise, the bulk of training involved annotating data rather than volume. Thousands of people work on a contracting basis to help annotate not only what was said, but when people were speaking, how they said things, and what emotions were used. Voice coaches were brought in to detect accents accurately.
IPO timeline
Reports suggest the company is looking at 2028 for an initial public offering. Staniszewski confirmed the company would love to create a business that stands the test of time.
The company is preparing the foundation to do so in the next years. Whether it proceeds depends on the time and place.
When pressed on the vagueness of “years”, Staniszewski laughed.
Safety and regulation
On whether frontier labs should slow down, everyone is aligned to work together on finding a way to pace. Whether they should be public about it, and how much media conversation or regulation it should involve, is another topic.
Staniszewski believes all parties should take the right precautions when deploying technology. The company does not train the text models and the intelligence side of models, which is the core key of the debate.
Regarding exposure similar to Hugging Face, ElevenLabs claims to be a step further. It does not deploy self-replicating or recurrent parts of the intelligence of agents. The technology does not allow agents to create more agents. Every customer goes through KYC.
Cybersecurity risk is definitely a risk for the wider world, but the company has a good set of precautions in place.
