The Trillion-Dollar Industries AI Is Disrupting: Voice, Law & the End of the Billable Hour
Summary
ElevenLabs says its commercial curve ran from an early-2023 launch to $100 million ARR in roughly 20 months, $200 million 10 months later, $300 million after another five, and $600 million today. It supports that scale with 600 employees, five-to-10-person teams, no product managers, and engineers embedded even in talent, legal, and go-to-market. Mati’s cultural proof point: the original 10-person research-and-engineering team has had “zero attrition.”
The voice-agent inflection came from combining human-sounding speech with reliable orchestration, knowledge, integrations, and interruptible turn-taking. Mati called the last 12 months—and especially the last six—a “step change”: callers move faster, disclose sensitive financial circumstances more candidly to AI, and may soon ask for an “AI operator” rather than escape one. The endpoint is proactive service informed by prior interactions, not another layer of “voice jail.”
Voice is becoming a licensable asset, but the market depends on identity safeguards. ElevenLabs traces generated audio, moderates both prompts and voices, and offers detection across its own and open-source models; its authenticated marketplace has returned more than $22 million to talent. Licensed voices can also cross languages or become interactive characters, as with Darth Vader in Fortnite, while restored voices for people who lost speech show why Mati calls voice “identity and IP.”
ElevenLabs’ defense against frontier-model bundling is to own the communication layer while remaining model-agnostic underneath. Customers can use Anthropic, OpenAI, Google, or open-source models without rebuilding their agent harness; ElevenLabs concentrates on audio architecture, more than 1,000 contractors labeling specialized data, industry workflows, integrations, and turn-taking. Mati said attempted distillation can be slowed, “not stopped,” and that the company is exploring more of its own interaction-focused intelligence.
Legora frames legal AI as a conversion of a $1 trillion services market where software captures only about $40 billion—“4% software, 96% service.” The host reported 50% quarter-over-quarter growth for seven quarters and said the company had just become one of the fastest enterprise companies with a direct-sales motion to go from $1 million to $150 million, beating Sierra by one quarter. Max described a million-dollar-revenue startup using ChatGPT for contracts, cap tables, HR, and IP assignments; Mati warned it could become a “fun diligence target” when someone must verify the work.
AI attacks law-firm economics by automating associate work and exposing the billable hour’s hidden cross-subsidy. Max’s framing was that firms “overcharge for the associates” and undercharge for decisive partner judgment; automation therefore pushes work toward fixed fees, success fees, and in-house execution. Legora used its own product across four acquisitions, with its fastest transaction moving from LOI to close in 12 days.
Legal AI’s defensibility rests on complete jurisdictional data, embedded workflows, and compliance—not a general-purpose legal model. A litigator cannot accept 80% of relevant precedent: “You actually need all of it,” including page citations and difficult-to-obtain historical records. Max called general legal-intelligence model building a “total waste of time and money,” favoring narrow models such as contract extraction, while describing “compliance is our currency” as the prerequisite for hosting governments, weapons manufacturers, and national-secret-bearing work.
Deep dive
1. ElevenLabs converted an audio breakthrough into a $600 million revenue business
Mati traced the company to 2022, when crypto and the metaverse dominated attention and the team could quietly build research and product. Its early-2023 release was, in his telling, “the first text-to-speech model that finally could sound human”; $100 million ARR followed in roughly 20 months, then $200 million after 10 more, $300 million five months later, and $600 million today.
Headcount has reached 600, but Mati offered continuity as the cultural counterweight: everyone from the original 10-person research-and-engineering group remains. The company recruits around a focused mission—“solving audio, solving interaction”—while spanning speech generation, transcription, orchestration, localization, support, operations, training, and sales.
The operating unit is typically a five-to-10-person team organized around a product or vertical such as telecom, financial services, or healthcare. Engineers also sit inside talent, legal, revenue, and go-to-market teams, both automating work and reviewing security; too little AI use is a warning, Mati said, but “using too much of it” without review is another.
The host’s management question produced the sharpest organizational answer: ElevenLabs has never employed product managers. Mati instead seeks people expert in one of coding, customer understanding, or design and fluent in another; AI can now lift an amateur capability toward an advanced one, letting a growth engineer design, ship, test, and interpret an experiment without waiting on several functions.
2. Voice agents escaped “voice jail” by becoming interruptible and contextual
Mati attributed the recent growth surge to enterprise selling finally meeting a reliable product: orchestration, models, knowledge, integrations, and voice now work together. He called the past 12 months, particularly the last six, a “step change” toward a “golden era” in which an agent remembers prior interactions and eventually shifts support from reactive troubleshooting to proactive help.
Financial-services deployments exposed a behavioral advantage. Customers discussing missed payments may feel ashamed with a human but tell AI “what actually happened”; they are also “more snappy,” interrupt freely, skip small talk, and reach the point faster. That requires a different interaction model, but Mati expects callers eventually to request an “AI operator.”
The host’s foot-pedal example showed the same interface shift in knowledge work: holding a pedal and dictating a one-to-two-minute stream-of-consciousness prompt into Wispr Flow creates richer context than exhausted typing. Mati extended the point to wearable recording tools such as Plaud Pocket, provided recording is disclosed, because otherwise-vanishing conversations can become notes and follow-ups.
3. Voice becomes an asset only if identity, licensing, and model control coexist
Mati’s governing principle was categorical: “Voice is identity and IP.” ElevenLabs therefore traces generated content, moderates at both text and voice levels to flag commercial uses or scams, and lets users test samples for AI generation across ElevenLabs and open-source models—not merely after misuse, but as infrastructure for a wider synthetic-audio market.
The host also recounted finding a channel that used a This Week in Startups archive and ElevenLabs to create his voice for bulldog videos. When his team later tried to clone his voice to correct advertisements, the system blocked the attempt until he verified it himself.
The opportunity side is authenticated licensing. Talent can create a voice, share it through ElevenLabs’ marketplace, accept default distribution pricing or set a price, and earn money; Mati said more than $22 million has been returned to the community. Licensed voices can preserve emotion across Spanish, Italian, and Portuguese or turn static MasterClass material into interactive instruction.
Interactive likeness extends beyond living talent. Mati would not discuss unspecified new productions, but confirmed that Epic Games’ Fortnite offered a live, mission-helping Darth Vader through partnerships with the estate and Disney. The more personal specimen was a woman who lost her voice before marrying and later repeated her vows with a restored version, letting her family hear them for the first time.
Against Anthropic, OpenAI, Google, and open source, ElevenLabs remains model-agnostic while trying to own communication end to end. Mati argued that in audio “it’s the architecture that matters, not the scale,” supplemented by specialized labeling from more than 1,000 contractors, vertical workflows, integrations, voice libraries, agent templates, and authentication that general model companies do not assemble.
4. Legal AI targets a trillion-dollar services pool and the billable-hour subsidy
Max described legal services as a $1 trillion annual market against roughly $40 billion of legal-software spend: “4% software, 96% service, which is bananas.” Because legal supply is constrained, his thesis is not simply fewer dollars for existing work; technology also lets providers serve new use cases and market segments and package new products. He pointed to Cooley’s founder-facing software platform, which embeds its material, precedent, and contract-review workflows.
The host reported that the company had sustained 50% quarter-over-quarter growth for seven quarters and had just become one of the fastest enterprise companies with a direct-sales motion to go from $1 million to $150 million, beating Sierra by one quarter.
Max supplied the cautionary edge: a startup with $1 million in revenue, several funding rounds, and dozens of employees had no corporate lawyer, relying on “ChatGPT, bro” for contracts, its cap table, HR, and even IP assignments. Mati predicted it could become a “fun diligence target” when someone must verify the answers.
Max’s explanation of law-firm pricing was that firms “overcharge for the associates” and undercharge for pivotal partner judgment. Mati said he had recently received a $1,800-an-hour bill; Max put associates at $800, while Mati said Kirkland can reach $4,000 an hour. They agreed that partner time can be worth much more when it avoids tens of millions in damage.
The discussion pointed toward fixed transaction fees or litigation success fees. Legora itself completed diligence for four acquisitions during the year using its platform; its fastest moved from LOI to closing in 12 days. Mati’s pushback was that founders want speed, while outside counsel’s incentives include minimizing liability and billing more time, creating pressure to drag work out.
5. Junior lawyers remain, but their apprenticeship becomes agent orchestration
Max called AI both an “existential threat” and an “existential opportunity” for firms such as Kirkland, which he put at about $10 billion in annual revenue, 4,000–5,000 lawyers, and $5 million–$10 million in profit per partner. Legora’s legal engineers—forward-deployed lawyers modeled on Palantir’s approach—sit with partners to redesign firms from a pre-AI to a post-AI operating model.
His answer on employment was measured: “The job will exist. The tasks will be different.” Firms still need juniors to become experienced partners, just as software organizations need junior engineers, but training will no longer center on manually reading every data-room document or searching virtual files with Control-F; the core skill becomes orchestrating, checking, and managing the agents doing that work.
Cross-border work illustrates both capability and its limit. Legora combines a customer’s precedents and organizational data with cases, legislation, and regulatory updates across jurisdictions; Max described a California GC entering South Africa and receiving an immediate response estimated at 80% accuracy. That is a starting point that replaces slow referral chains, not a claim that local legal judgment has already become unnecessary.
6. Complete legal data and compliance matter more than a “legal brain”
Max disputed the original AI assumption that incumbent data owners would automatically win: legacy organizations struggle to match AI-native tempo, talent, and decision-making. Legora partners with content providers in jurisdictions including Germany, France, and Spain, while the United States remains unusually difficult because LexisNexis and Westlaw form a legal-research duopoly.
Mati’s power-law challenge exposed the real moat: for consequential litigation, 80% coverage is insufficient because a billion-dollar case may depend on a missing precedent. “You actually need all of it,” Max agreed—historical books, scans, verification, and exact page citations—turning unglamorous acquisition and structuring work into a prerequisite for trustworthy agents.
Max said Claude Opus 4.5 and 4.6 can move beyond database search into case strategy that combines precedents and witness statements. He rejected building or fine-tuning general legal intelligence as a “total waste of time and money,” but endorsed narrow models where scale pays: tabular review across 100 documents and 100 prompts creates 10,000 API calls, making specialized contract extraction valuable for both latency and cost.
Mati noted that Claude now offers legal capabilities; Max replied that a bundle of “Markdown, skills files, and a couple of integrations” demonstrates demand, then generates Legora leads when users hit its ceiling. His harder boundary was deployment: although “compliance is our currency” and Legora says it hosts national secrets and weapons manufacturers’ contracts and works with governments, it does not do on-premises deployment; Max said deploying in a VPC is time-consuming and creates dependencies that slow the roadmap.