ElevenLabs and the Voice AI Bet Nobody Wanted

A quick note on sourcing before we start: Sequoia's event is billed as "AI Ascent 2026" in the published video description, which is worth flagging if you're reading this in 2025 — the year designation either reflects a forward-dating convention or a genuine scheduling quirk. I haven't been able to independently confirm the event date, so take the "2026" as Sequoia's label, not a verified dateline. The revenue figures that follow come from Staniszewski's own conference remarks, not from audited financials or independent reporting.

Mati Staniszewski's founding story for ElevenLabs starts in Poland, where a foreign film means one guy — one monotone guy — reading every character's lines aloud while the original voices play underneath. Male, female, villain, love interest: one voice. Flat affect. The audience is expected to do the emotional labor themselves.

I ran a bookstore for twelve years. We had audiobooks, and I watched what happened when a narrator was wrong for a book — customers returned them, sometimes apologetically, like it was somehow their fault they couldn't connect to a performance that just sat there dead on the page. Audio is not neutral delivery. It is half the meaning. Poland's dubbing tradition isn't just annoying; it's a small daily proof that when you strip emotional information from a voice, the content loses something that can't be recovered in the text.

That's what sent Staniszewski and his co-founder — childhood friends from the Warsaw suburbs — down into audio AI in 2022, when, as he put it at the event, "that was year of crypto and metaverse. Nobody was still working on the AI side." Even the people who were working on AI were chasing text and images. Audio was considered, in his words, "a big niche."

This is the part I find genuinely interesting, not because it's a classic underdog story (it is, but so what), but because of the specific structural reason audio got overlooked. The models were smaller than their text counterparts, which meant lower compute requirements — but the data still had to be transcribed and annotated before it was useful. That's not a glamorous problem. It's a lot of painstaking work before you have anything to show. The researchers who might have tackled it were mostly chasing the bigger, louder frontier. ElevenLabs walked into a room that was nearly empty.

They built remotely from the start — London and Warsaw — scraping GitHub to find researchers by their actual work output rather than their geography or institutional pedigree. This was 2022, when "remote-first" still made some investors nervous. Staniszewski doesn't frame it as ideology; he frames it as the only way to find the right people for a niche domain with a thin talent pool. That tracks.

The other early decision that shaped everything: they monetized immediately. "We try to stay healthy on the margins so we can continue investing," Staniszewski said, "with the assumption that it's better for us to figure out that stream and be able to be independent in that development." I've heard a hundred founders describe early monetization as a philosophical choice. What he's actually describing is simpler and more pragmatic: don't be hostage to your next fundraise if you can avoid it. Revenue buys you the ability to make decisions based on what's right for the product, not what's right for your cap table. They eventually raised externally as ambitions grew, but the baseline was earned revenue, not runway arithmetic.

The model stack they built followed the work, not a predetermined roadmap: text-to-speech first, then speech-to-text, then the combination of both for dubbing, then real-time streaming models for conversational agents, then music. Today they claim coverage of the full audio research domain. Staniszewski says ElevenLabs now has "just over 400 people" and "over $400 million in revenue" — figures he stated at the conference, with a moderator having referenced more than $100 million in net new ARR in Q1 alone. These are his numbers, self-reported from a stage, not independently verified.

The voice agent conversation is where things get interesting and, depending on your disposition, a little uncomfortable.

Customer support is the obvious application — automated phone trees that actually work, basically. Staniszewski doesn't spend much time there. He's more interested in what he calls "revenue generating opportunities," meaning sales. He described, as conference examples, Deliveroo supposedly using voice agents to contact restaurants for operational data, and Deutsche Telekom reportedly deploying them for inbound sales inquiries. These are his anecdotes from a CEO presentation; neither company has confirmed these deployments publicly, and I'd treat them with the same skepticism you'd apply to any founder's stage patter about named customers.

But the underlying claim is real enough to examine on its own terms: voice agents are being used not just to answer questions but to close transactions. And Staniszewski made an observation about information yield that I hadn't heard framed this way before. When customers call in and speak rather than fill out a form, they don't just give you the information you asked for — they give you context. Why they're calling, what isn't working, what else they're considering. He said ElevenLabs uses this on its own inbound flow and collects substantially more signal than form submissions would produce.

That's the part that gave me pause, not the "AI is replacing the phone tree" version of this story, which is mostly fine, but the version where emotionally responsive voice agents are optimized for extraction — gathering information customers didn't intend to share because a warm, adaptive voice made them feel like they were just talking. I've seen what skilled salespeople do with a customer who's chatty and comfortable. The distance between a voice agent that adjusts its tone when you sound stressed and one that's been calibrated to keep you talking until you've said something useful to the company is not a technical distance. It's a design choice.

Staniszewski's answer to a Q&A question about agent-to-agent negotiation was honest about the current state: "We haven't seen any truly successful on the negotiation front. It was like more, you know, kind of order taking." But he thinks emotional intelligence is what changes that — not just matching tone, but knowing when to pause, when to slow down, how to read the other side. If that sounds like what good salespeople and negotiators already do, that's because it is.

The most striking operating detail Staniszewski shared had nothing to do with audio. Every team at ElevenLabs — including HR, legal, and go-to-market — has an embedded engineer. Not as a tech liaison, but as a functioning member who builds automation and helps the non-technical staff level up their own tool use. When everyone at the company started using AI coding assistance, the embedded engineers became the reviewers — checking outputs for security and infrastructure implications that non-technical teammates couldn't reliably catch.

I've covered enough small businesses to know that the question of who owns the technical literacy in an organization is never purely academic. When I was running the bookstore, I was the person who had to figure out the inventory system, the website, the email marketing — not because I was the most suited to it, but because that's what you do when there's no one else. The ElevenLabs model is essentially a deliberate answer to that problem at scale: stop assuming technical capability is siloed in one department and start treating it as a resource that lives inside every team. Whether that survives as the company grows past 400 people is the real question — these kinds of flat, no-title, small-team structures tend to fracture under their own success — but the instinct is right.

They also built a scoring system to take contract negotiation decisions off Staniszewski's plate. Sales team asks for an indemnity provision or a liability cap; the score for that customer size tells you how many concessions you can make. Fully automated now. That's not a flashy product story, but it's the kind of operational detail that means the difference between a founder spending their time on product and a founder spending it refereeing internal dealmaking. The mechanics of running a company are always the unsexy part, and the companies that solve them quietly tend to outlast the ones that don't.

Staniszewski ended by sketching a future where voice is the primary interface for AI agents, robots, and whatever computing looks like next. The education examples he cited — Masterclass reportedly building interactive Gordon Ramsay and Chris Voss experiences using ElevenLabs, per Staniszewski's own conference remarks, not confirmed independently — are the ones that stuck with me. Not because of Gordon Ramsay shouting cooking instructions (though, sure), but because the idea of a patient, available, emotionally responsive tutor that adjusts to how you're absorbing information is genuinely different from what we've had. Not "AI teacher" as a gimmick, but as a structural solution to the fact that good individualized instruction is expensive and scarce.

Whether the same technology that makes a better tutor also makes a better manipulator is not a question ElevenLabs is uniquely responsible for answering. But it's the question that follows voice AI everywhere it goes, and emotional intelligence — reading stress, adjusting pace, knowing when to push and when to back off — makes it sharper, not duller.

The voice that sounds like it understands you is more persuasive than the one that doesn't. That's been true since long before there were machines capable of it.

Dorothy "Dot" Williams is Buzzrag's small business and entrepreneurship correspondent. She covered Main Street economics for twelve years while also running an independent bookstore in Asheville, which she sold in 2020.