Eleven Labs builds an AI speech layer, a model that converts text into human-like speech. Most people encounter this problem without realizing it when talking to customer service. For example, Klarna provides front-line telephone support for 35 million customers in the United States. So do Deutsche Telekom, Cisco, Adobe, and a growing number of governments. Eleven Lab also sells to creators who use audiobooks, dubbing, and music platforms.
It’s not done in isolation. In fact, the company trained its voice product on Celebrities and is increasingly reaching out to customers, including Decagon, a conversational AI platform that now competes with the company. But investors don’t seem too worried. The company claims to be generating annual recurring revenue at a rate of $600 million, but its backers reportedly value it at $22 billion, despite being just four years old.
To understand more, I interviewed Matty Staniszewski, co-founder and CEO of Eleven Labs, at Toronto’s Nrth (a local entrepreneurship conference formerly known as Elevate). We covered a variety of topics in a short amount of time, including whether companies should tell customers when they’re talking to AI (which he thinks they should) and whether they can discuss the company’s gross profit. Naturally, Staniszewski said he couldn’t discuss them in detail, but made it clear that he doesn’t mind them being squeezed further if it means increasing the company’s market share.
Our conversation has been condensed and lightly edited. You can read the full conversation here.
You attended TechCrunch Disrupt last year, where you said that the audio model will become commoditized within a few years. How do you evaluate that prediction now?
There is still much work to be done, and the quality improvements that can be achieved at the model level alone remain important. In the long term, perhaps three or five years from now, the difference will be smaller. What we want to do, and what we want to do first, is pass the Turing test for conversational AI. You need a combination of intelligence, but you also need emotional intelligence. You need to be able to understand the other person’s emotions so that you can slow down or speak up. That hasn’t been done yet.
What percentage of your business is currently enterprise?
Current ARR is $600 million. The remaining 55% are typical businesses, and the majority of the (remaining) 45% are small businesses, developers, builders, and creators.
Like everyone else in the AI industry, you are increasingly competing with your customers. Decagon now trains its voice products for users and queries them through its own models.
The lines become even more blurred. When you think about model companies, platform companies and application companies, there used to be a very clear separation of where they started and where they ended. Today, the line is even more blurred. In Anthropic’s case, what was once a model company has arguably become a platform that allows it to offer an increasingly wide range of applications. I think this will continue.
Customers can select “Inference Layer” from the menu of options in Eleven Labs. What are some considerations regarding the use of Frontier Lab models and open weights?
It’s not an either-or choice. In customer experience, if the call is just informational, no action is being taken. Many open source models can be used because the knowledge base defines what a good experience is. But if it’s a financial service, you need authentication, information about the transaction, and possibly a refund. There is no margin for error. The Frontier model will likely take the lead here as well.
Some of those open-class models are made in China. The US government is a customer. European governments are our customers. What are those conversations like?
different. For each deployment, the models and voices deployed vary by case. When we work with the Polish government or the Brazilian government, they have their own set of requirements. There are open-weight models, closed-source models, and proprietary fine-tuned models.[In Poland]this is a medical case. Patients across the public health system have appointments, but 18% do not show up. During deployment, an agent will call and remind you. They had a set of models optimized based on their knowledge, so we integrated them while maintaining data residency.
Should companies disclose when someone is talking to an agent rather than a human?
I think the information should be disclosed at this point. At the moment, people are not used to this phone and the common pattern is that they don’t want to feel cheated by it. But in five years, everyone will hire their own agent to represent them and expect an agent when they call. I think society will change if we do that. There is a better way. If a human wants to wait 30 minutes, give the customer a choice. Most of the time, they choose an agent and are surprised by how good their experience is.
What is your gross profit considering what you are paying for the model and inference?
I will give you a vague answer. Because we have that research element, we can fine-tune and constrain the model in very smart ways. However, if we can save you some money, we will do it. The biggest thing is still proving your value and being with your customers. So if you can invest and prove your value, you don’t care about the margins going down to actually make money as value is created over the next five years.
Millions of hours of customer service calls. Are you training with them? How much of the training data is synthetic?
Some companies have created models together. They needed a specific model that suited their use case. Otherwise, the majority of the training was about annotating the data, not the amount of data. We have thousands of contract staff in-house who help us annotate not only what is said, but also when it is said, how it is said, and what emotions are used. To be able to accurately detect accents, we had to implement a voice coach.
It is reported that the company is eyeing an IPO in 2028. Can you confirm that?
We want to build a company that stands the test of time. We are preparing the foundations to make that a reality in the coming years. But whether or not to do so depends on time and place.
“Year” is very ambiguous.
(lol)
Backstage: How do you know if it’s time to slow down at Frontier Labs?
We’re all going to work together to find a way to pace ourselves. Whether it should be publicized and how much media conversation and regulation should be involved is another topic. But certainly, we all need to take appropriate precautions when implementing technology. We do not train the text model or the intelligence side of the model, which is the core of the discussion.
Is it possible that Eleven Labs will be exposed like Hugging Face?
It goes a step further because it doesn’t introduce any self-replicating or repeating parts of the agent’s intelligence. Our technology does not allow agents to create more agents. All customers go through KYC. Cybersecurity risks are definitely a risk to the entire world, but we are taking appropriate precautions.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
