Mati Staniszewski, co-founder and CEO of ElevenLabs, describes building an AI-native audio company from 2022 — before ChatGPT launched — that has become one of the fastest-growing AI startups by combining frontier audio research with a unified communication platform for businesses.
Origins: From Polish Dubbing to Frontier Audio Models
The founding insight came from a distinctly Polish problem: in Poland, all movie voices — male, female, every character — are narrated by a single voice actor, stripping away emotional intonation; this “lektor” experience persisted into 2021 and sparked the vision of a Babel Fish future where content crosses language barriers with original voice and emotion preserved.
Mati (background: Palantir, BlackRock, mathematics) and Piotr (background: Google text models, university visual research) were best friends of 15 years and first-time founders; they started with dubbing but quickly discovered the research wasn’t good enough — transcription, translation, and speech generation were all robotic and failed to cross the uncanny valley.
Creator feedback redirected them: before dubbing, creators needed post-production fixes, script narration previews, and AI voiceovers for videos; this pulled ElevenLabs into building high-quality text-to-speech first, then expanding to the full audio stack (speech-to-text, localization, conversational agents).
Research + Product Deployment: The Core Organizational Model
ElevenLabs is explicitly structured as a combination of a frontier research lab (audio models only: TTS, STT, orchestration) and a product platform that applies those models to transform how companies communicate — marketing (Ramp’s Super Bowl ad), support (Deutsche Telekom call center agents), sales (inbound qualification), operations (Polish government healthcare appointment agents).
Research focus on audio alone is a deliberate competitive moat: audio requires solving both science and subjective art (voice preference, emotional nuance); the team believes architecture breakthroughs — not just data or compute — are the key differentiator, and they concentrate the world’s best audio researchers on that problem.
The “Honda engine” analogy: like Honda R&D as a separate engine-focused entity funding manufacturing, ElevenLabs keeps research pure on audio, then deploys it across modalities (ElevenReader, ElevenProductions, ElevenAgents, voice marketplace) where audio is the bottleneck and unique advantage.
They explicitly avoid intelligence/knowledge/coding battles; the goal is to be one of three major platforms mediating all human-AI and human-human digital communication within a few years.
Deutsche Telekom: Land-and-Expand via Forward Deployed Engineers
Deutsche Telekom started with marketing (daily AI-narrated podcasts in their Magenta app) two years ago when agents weren’t reliable; after proving value, they expanded to call center voice agents integrated with CRM and phone systems (SIP/Twilio), then to in-network agents that join live calls for booking and real-time translation.
Expansion is driven by Forward Deployed Engineers (FDEs) who sit inside the customer’s office, build integrations, simulate and trial calls, then scale — FDEs are part of the product team, not sales, with a dual mandate: solve the customer’s problem AND bring learnings back to the product so the next customer benefits (e.g., healthcare deployment patterns abstracted into reusable product features).
This model comes directly from Mati’s Palantir experience: first-month deployment to Aberdeen oil rigs to work side-by-side with customers, “best idea wins” in small teams (~5 people), and the critical second step of feeding field knowledge back into the product rather than becoming a services shop.
Flat, Transparent, AI-Native Organization
No titles, maximum 5 layers (targeting fewer over time), wide spans of control (~10 direct reports); information flows from the front lines (engineers, FDEs, customer conversations) directly to leadership without manager summarization.
AI amplifies this: summarizing all team activity, customer calls, and data streams so leaders can proactively engage where things aren’t working rather than waiting for someone to surface issues; near-total internal document transparency enables this.
Every function (talent, ops, legal, RevOps) has embedded engineers who automate and elevate AI adoption bottom-up; revenue engineering builds tools that transcribe calls, auto-populate CRM, and deploy AI SDRs — voice-based intake on the website yields richer customer information than forms because speaking is lower-friction and more natural.
Taste, Art, and the Human Element
Palantir’s “artist colony” ethos — blending art and science — resonates more now: voice is subjective (every voice sounds different to every listener), benchmarks are hard, and brand communication requires taste; ElevenLabs hires creatives, runs a voice marketplace (20,000+ voices, creators paid per use), and treats imperfection (ums, pauses) as a feature that makes agents feel human.
Accessibility work is a proud core: 10,000+ people with ALS, throat cancer, or voice loss have regained their voice — including a musician touring with his AI voice, a congresswoman speaking in Congress, and author/Tim Green (ALS) winning an Emmy for his AI-voiced podcast.
Growth Trajectory and Strategic Conviction
Revenue is skyrocketing on the Agents and Creatives lines; fintech (Revolut, Klarna, PagBank, Customers Bank) moves fastest, followed by healthcare, telcos, and retail/e-commerce.
Three concrete acquisition offers (last in June 2024), all declined: “AI is changing the world. We can build at the frontier of that change. We are going all in.” Mati (31) and Piotr share the conviction that this is likely their best lifetime idea and the timing is unrepeatable — selling would mean working on a lesser idea with money that doesn’t buy fulfillment.
Scott Wu (Cognition) shares this mindset: the intolerable outcome isn’t failure, it’s not trying; the biggest risk is not taking the risk.
Voice as the Primary AI Interface
Twelve months out: voice agents will combine IQ + EQ — emotionally understanding the user, adjusting tone, pausing, thinking, re-entering conversation naturally; the paradigm flips from humans learning technology’s language (keyboards, code) to technology adapting to humanity’s primal interface (voice).
Like Bell’s telephone vs. the telegraph: no specialized language to learn, just pick up and speak; screens may recede to pockets as voice becomes the default interaction layer across devices.
Conferences as a Disciplined Business Tool
Mati does 2–3 conferences/quarter, not for sessions or “the circus,” but as forcing functions to pre-arrange 1:1s with partners/prospects who are already there; preparation is key — explicit agendas, no wasted time, and a clear distinction between pitching and partnering.
Early on, he attended unprepared, drank, and got nothing; now it’s a systematic channel for land-and-expand, especially for conversational agents which apply broadly across customer journeys.
