Mati, founder of ElevenLabs: From a pain point in voice acting to building a $11 billion AI Communication Platform
Mati Staniszewski
co-founder, elevenlabsio
Original Statement
1. Core Views and Business Cognition
• Voice is the ultimate form of human-computer interaction: In the past computing era, humans were forced to adapt to the interaction methods of machines (keyboard, screen, programming languages). The evolution of AI will return the power of interaction to humans, with the most natural and primitive form of voice replacing complex peripherals and screens, becoming the core entry point for human-intelligence collaboration.
• Extreme focus builds a moat: ElevenLabs focuses solely on the audio field in basic research, not blindly expanding into pure text large models, images, or video generation. All efforts are dedicated to breakthroughs in audio architecture, subjective aesthetic quality, and emotional expressiveness.
• "Cutting-edge research + commercial implementation" as a dual driving force: It is neither a purely academic research laboratory nor a simple application company, but rather a combination of cutting-edge audio model research and product deployment that addresses real customer scenarios.
• Technology must empower humanity rather than replace humans: Opposing the arrogant narrative that "technology brings destruction/mass unemployment," it insists that AI should amplify human potential, creating positive value in barrier-free reconstruction, medical communication, and cultural and artistic creation.
2. Founder Background and Entrepreneurial Opportunities
• Founding team genes:
• Mati Staniszewski (CEO): With a mathematics background, he previously worked at BlackRock developing risk models and later served at Palantir in frontline deployment, skilled at deeply integrating optimization models with complex client business.
• Piotr Dabkowski (CTO): A close friend of Mati for 15 years, former core machine learning researcher at Google, specializing in knowledge graphs and visual multimodal algorithms.
• Reasons for starting the business and pain point discovery:
• Pain points in Polish film dubbing: In Poland, films generally use a single voice-over dubbing (Lektor), where the same voice actor reads without emotional fluctuations, resulting in a poor viewing experience.
• Initial idea: Break language barriers, allowing film content to retain the original voice, tone, and emotion of the original actors when crossing languages (similar to the "babel fish" concept in "The Hitchhiker's Guide to the Galaxy").
• Pragmatic adjustment of the entrepreneurial entry point: Initially, they wanted to do end-to-end dubbing, but during research on creator needs, they found that the most pressing issues for creators were "re-recording a misread line" or "directly generating high-quality narration." The team decisively adjusted priorities, first tackling the most urgent need for text-to-speech, breaking through the uncanny valley effect in voice.
3. Core Business Strategy and Product Matrix
ElevenLabs has built a three-tier progressive system around "voice and full-process communication": bottom-level cutting-edge model research $\rightarrow$ unified product communication platform $\rightarrow$ scenario-based terminal applications.
• Bottom-level cutting-edge model research:
• Covers text-to-speech (TTS), speech-to-text (STT), multilingual localization, and real-time voice dialogue orchestration (Voice Agent Orchestration).
• Core product lines and applications:
• Conversational AI / Agents: Enterprise-level dialogue agents, covering call center automation, real-time translation, and multi-channel interaction.
• Eleven Creative / Studio: Audio and video content creation, high-fidelity dubbing, sound design, and marketing advertisement production.
• Eleven Reader: A reading application for individuals that converts any document, long text, or report into high-quality, emotionally rich voice podcasts in real-time.
• Voice Marketplace: Over 20,000 certified voices are already available. Voice creators/voice actors share their voice tones and earn copyright revenue, building a strong bilateral network effect and creator ecosystem.
4. Key Data and Enterprise-Level Implementation Practices
• Core growth industries:
• Fintech: The fastest-growing application sector, with representative clients including Revolut and Klarna, fully implementing customer service inquiries and interaction automation.
• Telcos: Deutsche Telekom has fully integrated ElevenLabs into marketing podcasts and customer service call centers, even deploying real-time call translation and assisted agents at the network level.
• Healthcare and government public services: Collaborating with the Polish health department to deploy AI voice agents for automatic medical appointments and post-operative follow-ups.
• Retail and e-commerce: Expected to explode in 2024, fully penetrating pre-sales consultation and after-sales service.
• Public welfare and social value implementation:
• Cumulatively helped over 10,000 patients, musicians, and public figures who lost their voices due to diseases like ALS (amyotrophic lateral sclerosis) and throat cancer to regain their personal voice, including helping former NFL star Tim Green regain his voice and produce an Emmy Award-winning podcast.
5. Organizational Philosophy and Business Operation Model in the AI Era
• Inheriting and evolving Palantir's FDE (Frontline Deployment Engineer) mechanism:
• FDE belongs to the product team rather than the sales team: Engineers work closely with clients on-site (such as flying to client headquarters to work side by side) to solve complex system integration and business flow adaptation.
• Customer site as R&D laboratory: Extracting common pain points from individual client customizations and abstracting the general needs of specific industries into platform-level functions, feeding back to thousands of enterprises.
• Flat structure and small team autonomy:
• Advocating "Best Idea Wins," with no complex titles within the company and very few levels.
• Core business is broken down into agile small teams of 5-10 people, each with ample freedom and decision-making power.
• Engineering infused into all teams:
• Non-technical departments such as legal, HR, and operations are equipped with dedicated software engineers to enhance overall organizational efficiency through automation and AI tools.
• Deep self-use (Dogfooding):
• Fully utilizing its own AI entities to transform its business. For example, upgrading the traditional dropdown forms on the official website to a voice AI SDR, discovering that not only did the user experience significantly improve, but the business background and demand information collected far exceeded traditional text forms.
6. Business Competition, Independence, and Future Outlook
• Refusing to be acquired too early by giants:
• Facing substantial acquisition offers (having historically rejected 3-4 acquisition intentions), it insists on maintaining independent operations.
• The founders firmly believe that ElevenLabs is at the forefront of a generational opportunity and is their top entrepreneurial idea, and should not sacrifice the chance to shape the industry landscape for short-term capital returns.
• Evolution of voice interaction in the next 12 months:
• Voice interaction will leap from the current "text-to-speech unidirectional output" to natural communication that deeply integrates emotional intelligence (EQ) and intellectual intelligence (IQ).
• Future voice agents will be able to capture subtle emotional changes, understand when to pause, comprehend and express thought processes, and even incorporate just the right amount of colloquialism and imperfection (such as natural pauses and filler words) into conversations, eliminating the cold mechanical feel of human-computer dialogue.
Video source link: https://www.youtube.com/watch?v=RFccAuyPPOg
ABAB AI Insight
The direction of this draft is correct, and it captures several truly important points of this interview: Focus, FDE, Small Teams, Voice Interface, Independence. However, to achieve a truly high level, I suggest not limiting ElevenLabs to just a "voice generation company," but rather understanding it in the context of the migration of AI computing paradigms and next-generation enterprise software infrastructure.
Additionally, there are several data points and statements that I recommend correcting first.
First, the Voice Library is not 20,000+ certified voices. ElevenLabs' current official documentation states that the community Voice Library has over 10,000 voices, with some pages showing 11,000+.
Second, the data on public welfare projects can now be updated to be stronger. ElevenLabs recently disclosed that its "1 Million Voices" program has already covered over 11,000 people, with a long-term goal of helping 1 million people regain or preserve their voices.
Third, your statement that "ElevenLabs only focuses on audio and does not expand into images and videos" was basically true in the early strategy, but by 2026, it needs to be redefined. The official breakdown of the entire business now includes:
• ElevenAgents: Enterprise AI Agent;
• ElevenCreative: Voice, music, sound effects, which now also includes images and videos;
• ElevenAPI: AI Audio infrastructure for developers.
ElevenLabs itself has even explicitly stated that it has now "expanded well beyond voice."
So a more accurate statement would be:
ElevenLabs has not developed a general-purpose text large model, but has expanded into the entire Communication Stack with voice as its core capability.
The difference between these two statements is significant.
────────────────
1. Truly understanding ElevenLabs: It has undergone three upgrades in company identity.
Many people still understand ElevenLabs as:
Input text → Generate human voice.
This is the ElevenLabs of 2023.
Today, this is no longer the company.
It has actually gone through three evolutions.
First stage: AI Voice Model
The question it addresses is:
"Can machines speak like humans?"
TTS is the core.
The most important breakthrough in this stage is overcoming the "uncanny valley":
The problem with computer voices was not that they were unintelligible, but that they lacked a sense of life.
The rhythm was off.
The pauses were incorrect.
The emotions were wrong.
The breathing was unnatural.
The emphasis was misplaced.
The place where ElevenLabs first truly led was in making machine voices approach human-like:
Prosody—rhythm.
This is more important than whether the timbre sounds like a real person.
────────────────
Second stage: AI Audio Infrastructure
Next, the question became:
If all software needs sound, who will provide the sound infrastructure?
Thus, ElevenLabs began to cover:
TTS
STT
Voice Cloning
Dubbing
Sound Effects
Music
API
Now, ElevenAPI has unified capabilities such as Speech, STT, Voice Agents, Music, etc., for developers.
The significance of this step is immense.
It has transformed from:
An application
To:
Infrastructure.
Similar to:
Stripe is not just a payment page.
Stripe is payment infrastructure.
Twilio is not just a calling software.
Twilio is communication infrastructure.
What ElevenLabs truly wants to become is increasingly close to:
The Audio Layer of the AI world.
────────────────
2. The third stage is what is truly worth noting now: AI Communication Infrastructure
This is the most worthy aspect of this interview to understand in depth.
ElevenLabs is no longer satisfied with:
Making AI speak.
It is now solving:
How to enable AI to complete a full human conversation.
This is an entirely different market.
A phone customer service needs more than just TTS.
The entire chain is actually:
User speaks
↓
Speech-to-Text
↓
Understand intent
↓
LLM reasoning
↓
Call enterprise database/API
↓
Decide answer
↓
Text-to-Speech
↓
Turn-taking
↓
Determine when to speak
↓
Determine when to remain silent
↓
Handle interruptions
↓
Assess emotions
↓
Record CRM
↓
Security and compliance
↓
Monitoring, evaluation, auditing
This is what is called a Voice Agent.
So the real competitive unit in the future is no longer:
Whose voice sounds the best?
But rather:
Who can enable an AI conversation to truly complete a business task?
This is a very important cognitive upgrade.
────────────────
3. This is also why ElevenLabs' ARR has suddenly exploded
Here, capital data should be added, as it can explain why the company's strategy has changed.
ElevenLabs disclosed an ARR of approximately $350 million at the end of 2025; by the first four months of 2026, the company had already announced an ARR exceeding $500 million. In February 2026, during Series D financing of $500 million, the company's valuation reached $11 billion.
This is not ordinary SaaS growth.
It indicates a very important issue:
Voice AI has moved from being an "interesting generative AI feature" into the enterprise budget.
Why are enterprises willing to spend money?
Because it directly corresponds to costs.
For example, a traditional call center:
1,000 customer service agents.
Salaries.
Training.
Night shifts.
Office space.
Quality control.
Turnover rates.
Management.
Multilingual customer service.
These are all huge operational costs.
If an AI Voice Agent can handle part of the requests, the ROI is extremely easy to calculate.
When enterprises purchase it, they are not buying "AI."
They are buying:
A reduction in Cost per Resolution.
This is why the strongest enterprise-level AI products ultimately boil down to one question:
How much money did you save or earn for the customer?
────────────────
4. The Revolut case is very worth studying
Revolut is a typical case.
ElevenLabs disclosed:
Revolut has deployed ElevenAgents in the UK and Europe, covering over 4 million customers and supporting 30+ languages; during the initial deployment, the problem resolution time was reduced by over 8 times.
Note that what is truly important here is not the "30 languages."
But rather another question:
Revolut initially actually tried to do it themselves.
Because large companies might think:
STT + LLM + TTS, can't I just connect them myself?
Theoretically, it is indeed possible.
The real difficulty lies in production.
For example:
200ms latency.
Sudden loss of internet connection.
User interruptions.
Credit card information.
PCI.
Zero data retention.
Regulations in different countries.
Customer service CRM.
Recording.
Monitoring.
Evaluation.
Failover.
High concurrency.
This is a very classic entrepreneurial rule:
Demos are easy, production is hard.
The real commercial value of many AI companies lies in this gap.
────────────────
5. Therefore, the true moat of ElevenLabs is no longer the model
This point is very important.
If today we still say:
The moat of ElevenLabs is that its TTS model is the best.
I believe that is no longer sufficient.
Because model capabilities will ultimately diffuse.
OpenAI, Google, Amazon, Microsoft, Deepgram, Cartesia, etc., are all entering the real-time voice market, and there are a whole batch of infrastructure companies in the Voice Agent field, such as Vapi, Retell, LiveKit, etc.
The real moat is becoming six layers.
First layer: Model.
Voice quality.
Emotion.
Multilingual.
Low latency.
STT.
Speech-to-Speech.
────────────────
Second layer: Realtime Infrastructure.
The difference of tens of milliseconds or hundreds of milliseconds is not obvious in a chat window.
But it is very obvious on a phone call.
Human conversations are very sensitive.
Half a second delay:
Sounds like a robot.
One second delay:
Sounds like the phone is broken.
Two seconds delay:
User hangs up directly.
So the competition in voice AI is essentially also:
Latency Competition.
────────────────
Third layer: Turn Taking.
This is something that many non-technical people might easily underestimate.
Real human conversations are not:
A finishes speaking.
B finishes speaking.
A finishes speaking.
But rather:
Interruptions.
Hesitations.
"Um..."
Pauses.
Talking over each other.
Repetition.
Changing one's mind.
So a truly excellent Voice Agent not only "answers."
But must understand:
When not to answer.
This is a very deep product issue.
────────────────
Fourth layer: Enterprise Workflow.
Salesforce.
Zendesk.
Twilio.
CRM.
ERP.
Internal databases.
Payment systems.
Appointment systems.
Identity verification.
Without these, an AI Agent is just a chatbot.
────────────────
Fifth layer: Compliance + Trust.
Especially in finance, healthcare, telecommunications, and government.
SOC 2.
ISO 27001.
HIPAA.
GDPR.
Data Residency.
Zero Data Retention.
These things may not sound sexy.
But they are precisely the moat of multi-billion dollar enterprise software companies.
The FDE product page of ElevenLabs now clearly lists these capabilities as an important component of enterprise deployment.
────────────────
Sixth layer: Distribution.
This is the final and most important layer.
Being technically ahead by three months does not necessarily create a moat.
But:
1,000 companies have already deployed.
10,000 developers have accessed the API.
Customer service data has been accumulated.
FDE has already entered customer sites.
Developers have completed the integration.
Companies have completed compliance reviews.
At this point, the cost of switching suppliers is getting higher and higher.
This is the real compound interest.
────────────────
6. FDE is one of the organizational designs that entrepreneurs should learn from ElevenLabs.
Mati has worked at Palantir.
This experience may be more important than many people think.
One of Palantir's most influential organizational innovations is:
Forward Deployed Engineer.
Not traditional consultants.
Not sales.
Not after-sales service.
But:
Engineers directly enter the customer's battlefield.
ElevenLabs now publicly states that its FDE will participate in system design, integration, deployment, governance, monitoring, optimization, and subsequent expansion.
Why is this model particularly strong in the AI era?
Because what is truly lacking today is not the model.
But:
Last Mile.
Companies know AI is powerful.
But they don't know how to integrate AI into IT systems left over from decades ago.
Therefore, a very important structure for AI companies may emerge in the future:
Research Lab + Software Company + Consulting Capability.
These three capabilities used to belong to three types of companies.
AI has merged them again.
────────────────
7. But the FDE model has a danger: becoming a consulting company.
This is another side that needs to be recognized.
If:
Customer A wants a feature.
Write it once.
Customer B wants another one.
Write it again.
Customer C wants to redo it.
In the end, engineers are all doing custom development.
Then SaaS gross margins will slowly turn into consulting economics.
So the real problem that Palantir and ElevenLabs need to solve is always:
How to turn one-time customer customization into platform capabilities that all customers can use?
The ideal cycle is:
Customer Problem
↓
FDE Discovers
↓
Product Team Abstracts
↓
Platform Functionality
↓
All Customers Use
↓
More Customers
↓
Generate More Problems
↓
Abstract Again
This is called:
Customer → Product Learning Loop.
If this can be achieved:
FDE is not a cost.
FDE is part of the R&D system.
This is a very advanced organizational design.
────────────────
8. What is truly worth learning from ElevenLabs' "Focus" is not "doing nothing."
Many entrepreneurs misunderstand Focus.
Focus is not:
Always doing only one product.
ElevenLabs is a very good counterexample.
It started with:
TTS.
Later:
Cloning.
Dubbing.
STT.
Agents.
Music.
Creative.
It has even entered image and video workflows.
Why is this still called Focus?
Because all expansions revolve around a core capability:
Communication / Creation.
This is called:
Concentric Expansion.
Good companies do not avoid expansion.
But:
After expanding, the old moat can still provide advantages for new businesses.
Amazon expanded from books to all e-commerce.
AWS expanded from internal infrastructure to the cloud.
Apple expanded from Mac to iPod, iPhone, Watch.
More and more products.
Core capabilities become more concentrated.
ElevenLabs is following a similar logic.
────────────────
9. "Voice is the ultimate entry point of AI" needs to be understood at a higher level.
I suggest slightly modifying "ultimate entry point."
Because screens will not disappear.
Excel, Bloomberg Terminal, Photoshop, CAD, code editors all require high information density visual interfaces.
What may really happen is:
Voice becomes the default interaction layer of Ambient Computing.
What does Ambient Computing mean?
Computers no longer require you to:
Pick up a device.
Open an app.
Find a button.
Fill out a form.
Click submit.
But exist in the environment.
You directly say:
"Reschedule my meeting with the New York team for a time when everyone is available."
"Call this customer and negotiate the renewal for two years."
"Ask the hospital if there are any openings on Wednesday afternoon."
"Translate this article into Spanish and generate it in my voice."
At this point, humans are not:
Operating Software.
But:
Expressing Intent.
Software is responsible for execution.
This is where the true revolution of Agents occurs.
────────────────
10. From the perspective of computing history, this is actually a "shift in abstraction layers."
One of the best ways to understand AI is to look at computer history.
Every major technological revolution has reduced friction between humans and machines.
Machine language:
Humans adapt to machines.
Assembly:
Slightly closer to humans.
High-level languages:
C, Java.
GUI:
Mouse + Windows.
Touch:
Fingers.
Natural Language:
Text.
Voice:
Language itself.
Every technological advancement essentially means:
Machines increasingly understand humans, rather than humans increasingly understand machines.
This is also why the significance of Voice AI is far greater than "voiceover."
It represents:
A further naturalization of the Human → Computer Interface.
────────────────
11. What ElevenLabs is ultimately competing for is not the "voice market,"
But:
The Human-to-AI Communication Layer.
This is where its true imagination lies.
What is the most important entry point in the mobile phone era?
Touchscreen.
In the internet era:
Browser.
In the PC era:
GUI + Keyboard.
In the AI Agent era:
It is likely to be:
Voice + Vision + Context.
You speak.
AI hears.
AI sees the environment.
AI knows your context.
Then executes tasks.
So in the future, ElevenLabs' largest market may not necessarily be Podcast.
It may not even necessarily be Call Center.
It could be:
Cars.
Robots.
Smart glasses.
Headphones.
Home.
AI Companions.
Customer service.
Financial services.
Healthcare.
Education.
Game NPCs.
These all require one capability:
To give machines a trustworthy, natural, and emotional voice.
────────────────
12. Another direction that many people overlook: voice will ultimately become a "digital identity asset."
This could be very significant.
Today, identity mainly consists of:
Name.
Avatar.
Email.
Phone number.
In the future, it will add:
Voice Identity.
A person's voice:
Can be authorized.
Can be licensed.
Can be inherited.
Can be protected.
Can be revoked.
Can earn royalties.
This is also where the Voice Marketplace is truly worth paying attention to.
If this system matures:
Actors' voices.
Singers' voices.
Announcers' voices.
IP character voices.
Historical figures' voices.
Could all become digital copyright assets.
ElevenLabs has even further entered the music copyright system. In September 2026, Universal Music Group announced a collaboration with ElevenLabs to develop an AI music platform, emphasizing artist opt-in and licensing compensation mechanisms.
This indicates that AI audio is transitioning from:
Generation technology
To:
Copyright infrastructure.
This is a very important step.
────────────────
13. Why Mati does not sell the company is also easy to understand.
In the interview, he mentioned that he has already turned down multiple acquisition interests, which is consistent with the video description.
From an entrepreneur's perspective, it is easy to understand.
If you think of yourself as just:
An AI voiceover company.
10 billion dollars may already be very high.
But if your goal is:
AI Communication Infrastructure
Then the entire TAM is completely different.
Call Center.
Media.
Gaming.
Education.
Healthcare.
Financial Services.
Robotics.
Automotive.
Advertising.
Music.
Devices.
This is a foundational infrastructure market that could last for decades.
So the founder's calculation is not:
"How much can I sell for now?"
But:
"What is the value of the option I hold?"
This is also one of the differences in thinking between top entrepreneurs and ordinary entrepreneurs.
Ordinary entrepreneurs optimize for:
Exit Price.
Top entrepreneurs optimize for:
Optionality.
────────────────
14. I believe the 8 takeaways from this interview that are most valuable for entrepreneurs are:
1. Enter a huge market from a narrow problem.
ElevenLabs initially solved:
"Voiceovers are poor."
Not:
"Redefining global human-computer interaction."
Great companies often do this.
Amazon: selling books.
Facebook: college directories.
Uber: black cars.
Airbnb: renting air mattresses.
Stripe: seven lines of payment code.
────────────────
2. The first product must create a Wow Moment.
When users first hear ElevenLabs:
"Is this really AI?"
These moments are extremely important.
They will create natural word-of-mouth.
────────────────
3. The model is not the endpoint.
Model → API → Workflow → Platform → Ecosystem.
This is a very typical upgrade path for AI companies.
────────────────
4. The true value of enterprise AI lies in the Last Mile.
Not in benchmarks.
But in:
Can it be deployed?
Can it connect to the database?
Can it meet regulatory requirements?
Can it operate stably?
────────────────
5. The closer engineers are to customers, the faster the company learns.
This is FDE.
────────────────
6. The value of small teams is not in saving salaries.
But in:
Reducing coordination costs.
The larger the organization:
The more meetings.
The more approvals.
The greater the information loss.
In the AI era, the most precious resource is not people.
It is:
Decision speed.
────────────────
7. An AI Native Company does not just sell AI.
It first thoroughly uses AI itself.
HR.
Legal.
Sales.
Marketing.
Engineering.
Operations.
All redesigned.
This aspect of ElevenLabs' dogfooding is very worthy of study.
────────────────
8. The best AI companies ultimately do not sell AI.
They sell:
Lower costs.
Higher revenues.
Faster speeds.
Better customer experiences.
AI is just a means to achieve this.
This is a law in the business world that will never change.
────────────────
I will elevate the title one level.
Among your current titles, I would prefer to directly highlight the change in perception of "What kind of company is ElevenLabs."
My top recommendation:
ElevenLabs is not just a voice AI: How a 500M+ ARR company is competing for the next generation of human-computer interaction interfaces.
If more business-oriented:
From AI dubbing to 500M+ ARR: How ElevenLabs turns "voice" into the next generation of computing infrastructure.
If more entrepreneurial:
ElevenLabs founder Mati: From a dubbing pain point to building a $11 billion AI Communication Platform.
If more in-depth:
When machines truly begin to "speak": ElevenLabs and the next generation of human-computer interaction revolution.
Among these, I believe the second is most suitable for in-depth content.
Because it answers three questions at once:
Where it came from, how big it is now, and what it ultimately wants to become.
And this is precisely where the ElevenLabs case is most worthy of study.
M