Building One of AI’s Fastest-Growing Companies | Mati Staniszewski, ElevenLabs
In a Nutshell
ElevenLabs is building an AI-native audio platform by combining frontier research on voice models with forward-deployed engineers who embed directly with customers to solve real communication problems. The company’s edge comes from its singular focus on audio—text-to-speech, speech-to-text, and conversational agents—while treating taste and emotional nuance as critical product constraints, not afterthoughts. This approach has scaled ElevenLabs from a dubbing idea sparked in Poland to a platform used by fintechs, telcos, and governments, with a voice marketplace and forward-deployed model that turns customer deployments into reusable product improvements.
These notes were generated by AI and may contain inaccuracies.
Michael Dell has 40-plus years of experience and has been dominating for decades. The common narrative from biographies of history's greatest entrepreneurs is that "this time is different," but Michael believes this time is actually different. The conversation focuses on building an AI-native company from today.
Mati Staniszewski and co-founder Piotr are first-time founders who have been best friends for 15 years. ElevenLabs is their first venture, started in 2022. The company launched before the first version of ChatGPT, when crypto and the metaverse were the dominant topics. This timing allowed them to focus and build extensively on the AI side.
Mati was at Palantir building optimization models and bringing them to customers, including work with NHS during COVID response on vaccine distribution across the UK, and optimization work with the oil and gas industry. His co-founder Piotr was at Google working on text models for knowledge graph, and previously conducted university research on image visual models. Mati's background includes studying mathematics and working at BlackRock building risk models.
The founders' goal was to combine research and product deployment under one organization. On the research side, they built all the audio models, starting with a text-to-speech model in 2022 that was the first to cross human-like quality. Over time, this expanded to transcription models, localization models, and orchestration for voice conversations with agents. On the product side, they unified these capabilities into one platform that helps businesses transform how they communicate with their audience.
The initial trigger came from Poland, where movies are narrated with one single character voice regardless of whether it's a male or female character, causing emotional intonation to disappear. This still happened in 2021, and while cheaper and faster to produce, the quality is poor. The founders envisioned a future where content would have original voice, original emotions, and original intonation in local languages.
The founders saw how technology interaction would change, breaking language barriers and enabling different modalities and channels for digital conversations. In 2022, they knew they needed to unlock stories and content across modalities and channels, though they didn't know how quickly the shift from static to interactive would happen.
The first idea was dubbing static content to make movies feel native in Polish. Despite seeming like a tiny business, the founders thought it was huge because all content and stories could be delivered in local languages. They envisioned real-time translation scenarios, referencing the Babel fish concept from "The Hitchhiker's Guide to the Galaxy."
The dubbing process requires three steps: transcription, translation, and regeneration in another language. Research existing at the time was poor across all components, with robotic voices that couldn't cross the uncanny valley. NVIDIA had good research on speech-to-text, DeepL was doing well on translation, but all components for dubbing were insufficient.
Testing the dubbing idea with creators revealed immediate problems beyond dubbing: post-production line changes, pre-recording script narration, and AI narration over videos instead of human speaking. This feedback led them to solve speech generation research first and deliver a product for narrating content with high quality before addressing language shift.
Ramp helps companies cut expenses by 5% median and grow revenue by 16% median. The platform is designed to make finance teams faster and happier, used by top founders and CEOs. Deel enables global hiring, allowing companies to hire, pay, and manage workers anywhere. ElevenLabs uses Deel to hire exceptional talent globally.
The feedback from creators aligned with research needs. Since none of the three dubbing steps were good enough, Piotr assembled the best research team to solve the problems. The approach combined solving research components with delivering immediate product value.
The founder of Honda described their company as "a research lab for engines," separating Honda R&D as a distinct company funded by a percentage of sales. Research focused on experimentation while manufacturing applied the innovations. This philosophy aligns with how ElevenLabs operates today.
ElevenLabs maintains a focused research lab with research engineering, combined with small product teams (less than 10 people) that have flexibility and autonomy. The structure includes research, product deployment, and go-to-market strategies, with small teams running after specific problems using their best judgment.
ElevenLabs is introduced as a combination of research and product deployment. Research focuses on frontier audio models including text-to-speech, speech-to-text, and orchestration. Product is one platform helping companies transform communication through marketing, support, sales, and operations. Examples include Ramp's Super Bowl ad, Deutsche Telekom's voice agents, and Poland's government healthcare appointment system.
Research is solely focused on audio, while product combines audio with integrations, knowledge, and LLMs. The focus is important for building the best architecture for audio models. Audio presents a combination of science and art that is subjective, requiring getting voices right. The team believes many remaining problems exist on the architecture side rather than data or compute.
The company believes it has audio technology no one else in the world has, with the best audio researchers working on frontier audio models. When evaluating new products, the key question is whether they have a unique advantage applying audio models. If a product doesn't have audio/voice communication as a bottleneck, it's not their forte.
Two years ago, the company considered lip-dubbing and avatar creation for marketing campaigns but paused because video models weren't good enough. Even with the best audio, the experience wouldn't be satisfactory. Today they're revisiting this as open-source models have improved, but they're not creating text-to-video models from scratch. They focus on intersections where audio provides unique advantage, modifying existing assets.
The intense focus on audio serves as defense against larger labs. While research advantage may diminish over time, product and ecosystem components become increasingly important. The ecosystem includes a marketplace model where people create authenticated assets and earn compensation when their voice is used.
The marketplace has 20,000 voices where people create and authenticate their voice, then earn compensation when it's used. The company created the system and fills gaps by finding people to create missing voice types. Every time a voice is used, the creator gets paid. ElevenReader converts documents into podcasts, with voice actors receiving payment for usage.
ElevenLabs is building the best models to help deliver content, knowledge, and conversation in completely new ways. The company has three core product lines plus three developing bets, with approximately 20-30 different applications within those lines. ElevenLabs explicitly avoids intelligence work, knowledge work, and coding as those are not their strengths.
The company wants to become one of three major platforms where people set all their interactions over the next three years. The vision is to be the leading platform for engagement and communication, helping companies have one place to set up integrations, bring knowledge, bring brand assets, and deploy across the entire customer journey including marketing, sales, and support.
Deutsche Telekom is one of ElevenLabs' most deeply integrated customers. Their usage includes:
- Creating podcasts in the Magenta app using audio work
- Creating and distributing ads to their audience using ElevenCreative
- Using ElevenAgents for call center operations where customers get help through voice agents integrated with Deutsche Telekom's knowledge
- Deploying an agent inside the network for T-Mobile subscribers to schedule bookings or provide real-time translation during calls
All information across marketing, support, wider operations, and sales is captured back to understand customer engagement across the full spectrum of the journey.
Deutsche Telekom started with marketing two years ago as their first product. This was the most obvious use case because agents weren't very reliable or quick at that time. They created daily podcasts for customers to learn about what's happening in the world through ElevenLabs voices.
Expansion to support required building integrations with CRM systems, connecting to phone systems like SIP trunking or Twilio, and ensuring agents behave according to desired logic and have appropriate knowledge. Forward deployed engineers partnered with the team in Germany to work side by side on integrations.
ElevenLabs learned the forward deployed engineers approach from Palantir. FDEs are part of the product team, not the go-to-market team, and are deeply embedded in understanding product roadmap. They spend time working with partners side by side on problems in their offices.
The philosophy emphasizes being on the same side as the customer, obsessed with their problem, and understanding them deeply from the beginning. This differs significantly from previous experiences at BlackRock where there was minimal direct customer interaction.
Small teams at Palantir were about five people, considered big, and the same principle applies at ElevenLabs. FDEs solve customer problems while stretching the product in that direction, then bring knowledge back to improve the product for the next generation of companies.
The second step of learning expertise and bringing it back to the product is frequently missed. Without this step, FDE work becomes effectively just services. The approach uses customer base as a form of R&D, accelerating development through understanding domains. Problems are rarely unique to one company, so domain expertise remains with customers while product experience gets abstracted and made reusable.
ElevenLabs maintains a no-title organization carried over from Palantir, with a relatively flat structure and very few layers. Between Mati and the C-suite, there are maximum five layers today. The goal is to have fewer layers over time, not more, if AI helps run the organization more effectively.
Most people have close to ten direct reports, creating a wide-set structure that helps build the small team approach. This requires amplifying information across teams to summarize what's happening, how teams are performing, and identify gaps.
Currently, information flow relies on specific individuals surfacing information. With AI, there will be access to everything happening at better granularity than before, with AI summarizing back-and-forth conversations. This shifts from reactive to proactive engagement.
The company chose to be extremely transparent with data and documentation, giving everyone access to nearly all company docs, which helps create a system to tap into that knowledge.
The small-team approach will be a pattern across companies created in the future. It helps with adopting AI technology through bottoms-up adoption rather than top-down mandates. Even non-technical teams like talent, ops, and legal have engineers who help automate work and elevate AI usage.
This represents a bigger pattern where companies will infuse all teams with engineering resources. Centralized functions like legal exist but will have significant engineering embedded to scale operations without hundreds of people.
Revenue engineering is one of the biggest centralized functions. Tools help capture knowledge during conversations through transcription, making it easier to fill fields without manual time. The company dogfoods its own work, using AI agents to replace parts of operations.
An AI SDR example shows people can speak with a voice agent instead of filling dropdown forms. This results in people leaving more information than through manual forms, telling more about problems and use cases. The AI SDR function is developed centrally but deployed to each local team for refinement of local nuances.
The quote from "Zero to One" states: "The single most powerful pattern I have noticed is that successful people find value in unexpected places. And they do this by thinking about business from first principles instead of formulas."
The biggest revenue lines are on the Agents' side and Creatives' side. The number of use cases deploying conversational agents is skyrocketing. Fintech is adopting fastest, including Revolut, Klarna, PagBank, and Customers Bank. Healthcare, retail e-commerce, and telcos are the other biggest adopters, with fintech moving quickest due to regulatory aspects.
From Palantir, there was the concept of needing understanding of the art to do the job well, with the phrase "artist colony" used. This appreciation has grown as ElevenLabs intersects AI and creative space. Building audio research models requires nuance in delivery, as every voice has individual preferences that must be captured. Working with brands defining communication with the world also requires significant art.
ElevenLabs aims to combine technical excellence with creative sensibility. The company draws inspiration from Palantir's attempt to merge these domains, though success in this area remains open to judgment. The voice marketplace serves as one concrete example of this blend, alongside efforts to employ both for-the-point new engineers and for-the-point creatives who work directly with customers. Taste is increasingly recognized as a defining factor in AI product development, moving beyond buzzword status to represent a genuine limiting factor in creating superior product and deployment experiences.
Edwin Land, founder of Polaroid and Steve Jobs' hero, articulated the vision of building companies at the intersection of technology and liberal arts. Land's quote states: "Taste is as rare as a unicorn." This perspective contrasts with some current tech founders who appear to lack humanity, focusing on technology that replaces rather than enhances human capability. Historical figures including Steve Jobs, Edwin Land, and the founder of Honda consistently framed their work as inventing technology to enhance humanity.
AI development should amplify human potential rather than replace it. This principle guides ElevenLabs' approach, particularly in voice technology applications. The company has worked with over 10,000 people to restore their voices, including cases involving ALS patients, throat cancer survivors, and musicians who lost their voices. One notable case involved a musician who performed a concert with his original band using an AI-recreated voice and is now touring across the UK.
A congresswoman who lost her voice used AI voice technology to continue inspiring others while serving in Congress. Voice carries unique emotional impact and recognition value that text cannot replicate, including emotion, intonation, imperfections, and pauses. The company partners with 800 organizations to deploy voice technology across education and broader cultural contexts.
Tim Green, a former bestselling author and NFL player who developed ALS, started a podcast using his AI-recreated voice. He recently interviewed Howard Lutnick and won an Emmy for his work. His visibility inspired hundreds to thousands of people to reach out seeking similar voice restoration services.
Voice creates emotional attachment that exceeds text-based relationships. Podcast listeners develop deeper connections with hosts than readers typically form with favorite authors, even acclaimed writers like Cormac McCarthy. The imperfections in speech—including filler words, pauses, and verbal quirks—contribute to authenticity that audiences value more than technical perfection.
Initial attempts to create perfect voice agents without imperfections failed to sound human. Adding "ums," "uhs," and natural pauses dramatically improved user satisfaction, with people reporting the technology felt genuinely human and conversational. This represents a fundamental shift from trying to eliminate human-like qualities to embracing them.
Voice AI presents unique research challenges compared to text models. Every voice differs from every other voice, making direct model comparisons difficult since different systems produce different voices inherently. Text-to-speech benchmarks face this fundamental comparability problem that text language models do not encounter.
Growing up in Poland, entrepreneurship was not a common conversation topic. Mati Staniszewski gained exposure to this path through studying abroad and working at exceptional companies like Palantir, which demonstrated that working directly with customers to solve problems was achievable.
Over years at Palantir and his co-founder's time at Google, they conducted hack weekend projects together. The decision to start ElevenLabs crystallized in 2021, representing two to three years of exploration before identifying a genuine problem worth obsessing over.
ElevenLabs has received three or four concrete acquisition offers, with the most recent arriving in June of the previous year. The company declined all offers, choosing to remain independent and pursue frontier AI development. The founders view the current timing as uniquely favorable for building something special.
At age 31, Staniszewski recognizes ElevenLabs as potentially the best idea he will have in his lifetime. The philosophy prioritizes building over financial exit, recognizing that the money will come regardless. The focus remains on continuing to build rather than optimizing for acquisition.
Staniszewski's co-founder maintains a lower public profile but shares the all-in commitment. He brings both research excellence and business insight, deeply considering AGI implications and how the research team can participate in fundamental technological change. The timing is viewed as historically unique across the entire team.
Voice represents one of the largest future interaction methods with AI systems. The next bottleneck in AI development will be communication and collaboration methods. The goal is solving voice interaction for everyone, enabling technology to adapt to human communication preferences rather than requiring humans to learn technological languages.
Within twelve months, voice AI should understand users through both IQ and EQ dimensions, recognizing emotional states and adjusting responses accordingly. The technology will incorporate natural pauses, thinking time, and conversational re-entry. The fundamental shift moves from humans learning technology's language to technology adapting to primal human voice communication.
Alexander Graham Bell distinguished between telegraph and telephone by noting that telegraph required learning specialized language, while telephone allowed immediate use of existing communication methods. This same principle applies to voice AI eliminating the need to learn technological interfaces.
The original company vision referenced the Babel Fish from The Hitchhiker's Guide to the Galaxy, not as creating the universal translator itself, but enabling everyone to have translation capabilities within their existing devices, presence, and work contexts.
Staniszewski typically participates in two to three conference events per quarter. This differs from advice given by Daniel Ek, who advocates avoiding the "circus" of conferences where people gather to distract rather than work. Ek's perspective emphasizes that elegant businesses can attract opportunities without active conference circuit participation.
Conferences serve as an effective forcing function for bringing partners, prospects, and clients together in one location to explore collaboration opportunities. This approach works particularly well for conversational agents work, as it applies to how most businesses think about customer communication and journey transformation. Events provide dedicated time to connect with existing contacts and convert prospects into customers, or expand existing relationships from one to two or five clients.
Initial conference experiences proved unproductive when approached without clear purpose. The lack of preparedness meant flowing through events without targeted outcomes, attending sessions that often lacked relevance to the business, and engaging in activities that created the illusion of importance but delivered minimal value. This led to a period of avoiding events entirely.
Successful conference participation requires thorough preparation, including pre-arranging one-on-one meetings with desired contacts. Since the same people often attend multiple events, strategic selection of meetings with different individuals becomes essential. Explicit communication about meeting intentions prevents misunderstandings where one party expects business discussions while the other anticipates casual interaction.
Many conference attendees share the explicit goal of conducting business and exploring opportunities. When both parties approach meetings with clear business intentions, the interactions become productive rather than creating discomfort. Proper preparation and pre-arranged time slots have transformed conference effectiveness.
The most valuable conference time focuses on securing meetings with key individuals rather than attending sessions or content alone. While speaking at conferences occurs frequently, the primary objective remains maximizing direct interaction time with attendees. This approach has yielded consistently positive results.
Alcohol consumption has been largely eliminated except for specific social occasions like weddings. European social culture tends to emphasize drinking in ways that some find enjoyable, though this has become less frequent over time.
Keep David Senra in your library
Save the videos and channels worth coming back to, and find them again in one place.
