ElevenLabs is fine with its gross margins shrinking if that wins it more of the market. CEO Mati Staniszewski made that clear in an interview with TechCrunch AI at Nrth, an entrepreneurship conference in Toronto. He also argued that businesses should tell callers when they’re talking to a bot. The company is four years old, says it’s on pace for $600 million in annual recurring revenue (ARR), and its backers reportedly value it at $22 billion.
If you haven’t heard of ElevenLabs, you’ve probably still heard it. Its models turn text into speech that sounds human. Klarna uses it for first-line phone support for 35 million U.S. customers. Deutsche Telekom, Cisco, Adobe and a growing number of governments use it too. Below are the key questions from the interview and what the answers tell us.
💰 Why would a company welcome lower margins?
Staniszewski wouldn’t give margin numbers. He did explain the thinking behind them, though. ElevenLabs does its own research, which lets it “fine-tune and constrain models in extremely smart ways.” When that saves money, the savings go to customers.
“We don’t mind the margins going lower to actually benefit together as the value gets created in the next five years,” he said.
This is a land-grab strategy. ElevenLabs is betting that being built deep into companies’ customer service systems will matter more than profit per call. That bet makes sense when rivals are catching up fast. It gets riskier if the IPO window opens and public investors start asking about unit economics.
🔊 Is voice AI turning into a commodity?
At TechCrunch Disrupt last year, Staniszewski predicted audio models would become commoditized within a couple of years. He’s now pushed that out. He says the quality gap between models is “still significant” today and will probably narrow over the next three to five years.
His next goal is to pass the Turing test for conversational AI. That means making a voice agent people can’t tell apart from a human. He says getting there takes emotional intelligence as well as raw intelligence. The agent has to read the caller’s mood and know when to slow down or speak up.
Keep in mind who’s making this prediction. A CEO whose valuation depends on model quality has good reason to say that quality still sets products apart.
⚔️ What happens when your customers become rivals?
Decagon, a conversational AI platform, trained its voice product on ElevenLabs and now runs queries through its own models. Staniszewski doesn’t treat this as a crisis. He says the lines between model companies, platform companies and application companies are “much more blurry” now. He pointed to Anthropic, which started as a model company and now runs a platform and a growing set of applications.
The same thing is happening across the industry. Everyone is moving into everyone else’s layer, so relying on a single vendor is a real strategic risk for both sides.
🧠 Frontier models or open-weight?
ElevenLabs lets customers choose the “reasoning layer” behind their voice agents, and Staniszewski says it’s “less of a binary choice” than people think:
- Informational calls: Open-weight models work well here, because the knowledge base determines how good the answers are.
- Financial services, refunds, authentication: “There’s no room for error.” Frontier models still lead.
- Governments: Each one has its own rules. The Polish public health system uses ElevenLabs agents to remind patients about appointments, since 18% of patients never show up. The setup uses Poland’s own tuned models and keeps the data inside the country.
🤖 Should bots say they’re bots?
Yes, for now. “The common pattern is you don’t want to feel cheated on that call,” Staniszewski said. He expects attitudes to change within about five years, once people have their own AI agents and expect to reach an agent when they call a company.
His practical suggestion is to give customers a choice. If the wait for a human is 30 minutes, offer the agent instead. He says most people take the agent and are surprised by how good it is.
What this means for teams using voice AI
- Choose models based on risk. Save frontier-model spending for calls that involve money or identity.
- Tell callers they’re talking to AI. It builds trust now, and regulators are likely to require it anyway.
- Don’t depend on one vendor. If your supplier can become your competitor, keep your options open.
- Lock in prices while they’re low. A company that’s cutting margins to win market share is a good one to negotiate with.
The interview also covered training data. ElevenLabs uses thousands of contractors to label not just what speakers say but how and when they say it. Voice coaches help with the emotional labeling. The full conversation is available at TechCrunch AI.