Hajra Naz
ElevenLabs builds the voice layer of AI. These are models that turn text into speech that sounds genuinely human. Most people encounter this technology when talking to customer service. Often, they don’t even realize it. Klarna runs first-line phone support on this technology.
That covers 35 million U.S. customers. Several other companies use it too. That includes Deutsche Telekom, Cisco, Adobe, and a growing list of governments. ElevenLabs also sells to creators. They use the platform for audiobooks, dubbing, and music production.
ElevenLabs isn’t alone in this space, though. In fact, the company increasingly bumps into its own customers. That includes Decagon, a conversational AI platform. Decagon trained its voice product on ElevenLabs. Now, it competes directly with ElevenLabs.
Still, investors don’t seem too concerned about this dynamic. The company says it’s pacing at $600 million in annual recurring revenue. Reportedly, backers now value ElevenLabs at $22 billion. That’s a striking figure for a company just four years old.
To learn more, this interview took place with ElevenLabs co-founder and CEO Mati Staniszewski. The conversation happened at Nrth in Toronto. That’s a local entrepreneurship conference, formerly known as Elevate. Several topics came up in a short time. That included whether businesses should disclose AI conversations to customers.
He believes they should. The conversation also touched on gross margins. Unsurprisingly, Staniszewski couldn’t discuss specific details. Still, he made his position clear. He doesn’t mind margins shrinking further if it means growing market share.
Read More: ElevenLabs Launches AI Music Tool for Commercial Use
On model quality and long-term goals:
There is still a lot of work to be done. The quality difference achievable at the model level remains significant right now. Looking further ahead, maybe three to five years out, those differences will likely shrink. One goal stands out for ElevenLabs specifically. The company wants to be the first to pass the Turing test for conversational AI.
This requires combining intelligence with emotional intelligence too. Systems need to understand emotions on the other side of a conversation. They need to know when to slow down or speak up. Nobody has achieved this yet.
On revenue breakdown:
ElevenLabs currently sits at $600 million in annual recurring revenue. More than 55% comes from classic enterprise customers. A significant portion of the remaining 45% comes from smaller sources. That includes small and medium businesses, developers, builders, and creators.
On competing with customers like Decagon:
The lines between companies are becoming increasingly blurry. When we think about model companies, platform companies, and application companies, the old divisions used to be clear. Today, those lines have blurred significantly. Take Anthropic as an example. What started as a model company is now clearly a platform too. It’s increasingly building a wide range of applications as well. This trend will likely continue.
Read More: Anthropic Reveals Its Biology Lab Has Already Found Something Important
On frontier models versus open-weight models:
This isn’t really a binary choice anymore. For customer experience, context matters a lot. If a call is purely informational, with no actions being executed, open-source models often work well. That’s because the knowledge base defines what a good experience looks like. However, financial services present different requirements.
Customers need authentication. They need accurate transaction information, such as refund details. There’s no room for error in these cases. Here, frontier models will continue leading.
On working with governments using Chinese open-weight models:
Each situation looks different. In every deployment, the models and voices used depend on the specific case. Working with the Polish government or Brazilian government comes with unique requirements. That could mean an open-weight model, a closed-source model, or their own fine-tuned model. In Poland specifically, this involves a healthcare use case.
Patients book appointments across the public health system there. Notably, 18% of patients never show up for these appointments. The deployment involves agents that call and remind patients directly. Poland had its own set of models, optimized on their specific knowledge. ElevenLabs integrated its technology while maintaining data residency requirements.
On AI disclosure to customers:
Disclosure should happen at this point in time. Right now, people simply aren’t used to talking with AI agents. The common concern is feeling deceived during a call. However, this may change within five years. Once everyone has their own agent working on their behalf, expectations will shift. People will start calling in and expecting to speak with an agent.
Society will likely adjust as a result. There are good ways to handle this transition already. For example, if there’s a 30-minute wait for a human representative, offer customers a choice instead. In almost all cases, customers choose the agent. They’re often surprised by how good that experience turns out to be.
On gross margins:
This answer will remain somewhat vague intentionally. Given the company’s research capabilities, ElevenLabs can fine-tune and constrain models in smart ways. When possible, the company passes savings on to customers directly. Proving value while staying close to the customer remains the biggest priority right now. Because of this, the company doesn’t mind seeing margins compress further. The goal is creating shared benefit as value develops over the next five years.
On training data:
For certain companies, ElevenLabs built models directly with them. These companies wanted something specific for their particular use case. Otherwise, the biggest training challenge hasn’t been about data volume. Instead, it’s been about properly annotating that data.
Thousands of contracted workers help annotate this information internally. They don’t just capture what was said. They also track when people were speaking, how they spoke, and what emotions came through. The company even brought in voice coaches. This helped ensure accurate accent detection across different training data.
On a potential 2028 IPO:
The company wants to build something that stands the test of time. Right now, ElevenLabs is preparing the foundation needed for this goal. Whether an IPO actually happens depends on timing and broader market conditions.
When pressed on the vague timeline, Staniszewski simply laughed in response.
Read More: How the historic SpaceX IPO is creating overnight millionaires among employees
On whether frontier AI labs should slow down:
Everyone in the industry seems aligned on working together to find proper pacing. Whether companies should discuss this publicly is a separate question. So is how much this should involve media conversation or formal regulation? Still, taking proper precautions while deploying this technology matters. Notably, ElevenLabs doesn’t train the text models or the core intelligence side of AI systems. That’s the central focus of most safety debates happening right now.
On security risks similar to Hugging Face’s breach:
ElevenLabs believes it sits a step further from this particular risk. The company doesn’t deploy self-replicating or recurrent intelligence components within its agents. Its technology doesn’t allow agents to create additional agents on their own. Every customer goes through a KYC verification process too. Cybersecurity risk remains a real concern for the broader tech industry overall. Still, ElevenLabs maintains a solid set of precautions to address these risks.
