ElevenLabs develops the models that convert text into speech that sounds human — the voice layer many people encounter in customer service calls, often without realizing it. The company counts corporate and government clients among its customers: Klarna runs first-line phone support for 35 million U.S. customers on ElevenLabs' technology, and Deutsche Telekom, Cisco and Adobe are also users. Creators use the platform for audiobooks, dubbing and music.
Although ElevenLabs faces competition and sometimes competes with its own customers — for example Decagon trained its voice product on ElevenLabs and now runs its own models — investors appear sanguine. The company says it is pacing at $600 million in annual recurring revenue (ARR) and is reportedly valued by backers at $22 billion, despite being only four years old.
Interview highlights with CEO Mati Staniszewski at Nrth
I spoke with Mati Staniszewski, co-founder and CEO of ElevenLabs, at Nrth in Toronto (formerly Elevate). In a condensed interview he discussed model quality and commoditization, enterprise adoption, the role of open-weight vs. frontier models, annotation and training data practices, and questions of disclosure and regulation.
-
Business mix and scale: ElevenLabs reports roughly $600 million in ARR, with more than 55% coming from classic enterprise customers. Much of the remaining revenue comes from small and medium businesses, developers and creators.
-
Model choices by use case: Staniszewski argued the choice between frontier lab models and open-weight models is not binary. For informational calls where no transactions occur, open-source models can be sufficient because the knowledge base drives the experience. For financial services or other sensitive cases requiring authentication and transaction handling, frontier models still tend to lead.
-
Government deployments and data residency: ElevenLabs works with multiple governments and tailors deployments to each case. Deployments can use open-weight models, closed-source models, or locally fine-tuned models depending on requirements. He cited a Polish healthcare deployment where reminder calls help reduce no-shows — in that system 18% of booked appointments result in no-shows — and noted the company integrates while respecting data residency rules.
-
Competing with customers and blurred lines: Staniszewski acknowledged that the boundaries among model companies, platforms and application vendors have blurred. Companies that began as model providers often become platforms and then offer applications. That dynamic means vendors sometimes compete with customers who have built their own alternatives.
-
Disclosure when callers speak to AI: He believes businesses should disclose when a caller is speaking to an AI agent today, since people aren't yet accustomed to it and may feel deceived. He expects social norms to shift in a few years as agents acting on users' behalf become common; until then, disclosure and offering customers choices (for example, human agent vs. AI when wait time is long) are good practices.
Data, annotation and margins
Staniszewski said that for certain clients they build custom models together and fine-tune for specific use cases. A major part of their training effort has been annotation rather than sheer volume: ElevenLabs used thousands of contract annotators internally to label not just what was said but timing, delivery and emotional cues, and they brought in voice coaches to accurately detect accents.
On gross margins and inference costs he gave a guarded answer, saying he could not provide detailed numbers. He emphasized that the company uses research-driven methods to fine-tune and constrain models efficiently, and that if savings can be passed to customers they will do so. For now, proving customer value and expanding market share matter more than protecting margins in the short term.
Safety, governance and IPO timing
Staniszewski supports coordination among labs and companies to pace deployment and take precautions, though the optimal public posture and regulatory involvement are open questions. He noted ElevenLabs does not train text-generation models that would enable self-replicating agents; their technology does not permit agents to spawn additional agents. Every customer undergoes KYC, and the company says it has robust cybersecurity measures in place to mitigate wider risks.
Regarding reports of a 2028 IPO timeline, Staniszewski said ElevenLabs is preparing the foundation to be able to go public in coming years but would base any decision on timing and market conditions.
Why it matters
ElevenLabs illustrates how voice AI is being embedded rapidly across commercial and public services — from customer support to healthcare reminders — while the distinctions among models, platforms and applications blur. With the company reporting $600 million ARR and a reported $22 billion investor valuation, investors see significant potential in the speech layer of AI, even as questions about data practices, disclosure and governance remain active.
