All podcasts / All-In Podcast / Summary

The Trillion-Dollar Industries AI Is Disrupting: Voice, Law & the End of the Billable Hour

2026-07-14 - 51 min - source - Read full transcript
Jason Calacanis (host)Mati StaniszewskiMax Junestrand

Key insights

ElevenLabs scaled from zero to $600M in revenue in under three years with zero attrition on its founding research team.
Staniszewski said the company launched its first human-sounding text-to-speech model in early 2023, then took about 20 months to reach $100M ARR, 10 more months to $200M, and 5 more months to $300M, closing 2025 near $600M with 600 employees; he credits retaining the entire original 10-person research and engineering core.
voice-ai-product-and-culture
ElevenLabs has never hired a product manager because AI now lets individuals cover the full code, design, and customer skillset a PM role used to require.
Staniszewski said the ideal PM was always someone who could code, understand the customer, and understand design at once, which is rare; AI tools now let someone go from amateur to 'advanced' across those functions, removing the bottleneck that made dedicated PMs necessary, so the company runs five-to-ten-person cross-functional pods instead.
voice-ai-product-and-culture
ElevenLabs embeds engineers inside every function, including legal and talent, both to automate work and to catch unreviewed AI-generated output before it ships.
Staniszewski described embedded engineers in non-technical teams with two jobs: building automations for that team, and acting as a security/quality check on what people in that team are shipping with AI tools, since employees new to coding can generate software without knowing whether it is secure.
voice-ai-product-and-culture
ElevenLabs pays voice talent directly through a licensing marketplace, turning one-time hourly voiceover work into recurring royalty income.
Staniszewski said the company has paid out over $22M to the community of voice talent through its marketplace, where people record a sample, get authenticated, and either take a default licensing rate or set their own price, replacing the old model where voiceover actors were paid hourly with only occasional backend deals.
voice-identity-and-ip
Voice cloning without consent is largely unregulated in the US, leaving companies to self-police through detection and blocking rather than law.
Calacanis described someone using his own podcast archive and ElevenLabs to clone his voice for an unauthorized video channel; Staniszewski said ElevenLabs' actual safeguards are internal - tracing all generated audio, moderating for commercial or scam intent on both voice and text, and building public tools that let anyone upload a sample to check if it is AI-generated, including for other companies' models.
voice-identity-and-ip
ElevenLabs has restored voice to people who lost it to illness, including a first-ever congressional speech and a wedding vow renewal.
Staniszewski cited working with US Congresswoman Jennifer Wexton, who lost her voice and used a recreated version to deliver the first synthetic-voice speech in Congress, and a woman who lost her voice before her wedding and was able to redo her vows in her own voice with her family present.
voice-identity-and-ip
ElevenLabs stays model-agnostic on purpose so customers aren't locked into any one frontier-model provider, even as those providers compete with it.
Staniszewski said the platform lets customers plug in Anthropic, OpenAI, open-source, or Google models so they aren't dependent on any single one; he argued ElevenLabs' defensible layer is the interaction/communication stack itself (turn-taking, voice architecture, labeled audio data from over a thousand contractors), not the underlying LLM, even though frontier labs are increasingly competing on voice too.
ai-vs-frontier-model-dependence
Both guests said they already believe AGI has effectively been achieved but not yet deployed.
Staniszewski argued classic Turing-style tests are already passed and a new bar is needed (something like 'more intelligent than every person on the planet times 10'), then said outright: 'We've kind of achieved it. We just haven't deployed it,' regarding text intelligence and predicted voice will hit the same bar for indistinguishable conversation within the year.
ai-vs-frontier-model-dependence
Legal services is a trillion-dollar market that is only 4% software today, which Legora reads as the core growth opportunity.
Junestrand said roughly a trillion dollars a year goes into legal services but only about $40B of that is software spend - a 96% service, 4% software split he called 'bananas' - and argued the software share must expand because legal is also supply-constrained: demand for legal services already outstrips the number of available lawyers.
legal-ai-disrupting-the-billable-hour
The billable-hour model systematically overcharges junior associates and undercharges senior partners, which AI is now exposing.
Junestrand said law firms price associate time far above its real value while underpricing partner time that actually matters most (e.g. avoiding a company-ending litigation mistake), because hourly billing has no other way to capture that value; he expects firms to keep shifting toward fixed fees per transaction or success fees tied to litigation outcomes.
legal-ai-disrupting-the-billable-hour
Legora's 'forward deployed lawyer' role exists because law firms and enterprises need help redesigning workflows for AI, not just a tool.
Modeled explicitly on Palantir's forward deployed engineers, Legora embeds lawyers who sit with Kirkland & Ellis-level partners to help transform the practice from pre-AI to post-AI; Junestrand called it necessary because AI in law isn't a mild productivity gain like document management or PCs were - it can do a large share of the actual legal work.
legal-ai-disrupting-the-billable-hour
Enterprises are already pulling legal diligence in-house because AI tools make it faster and cheaper than paying outside counsel.
Junestrand said Legora has made four acquisitions in the past year and did diligence in-house with its own tool for all of them, closing one deal in 12 days from LOI to close; he framed the incentive gap directly - the founder wants the deal done fast, the outside lawyer's incentive (get paid more, avoid liability) is to drag it out.
legal-ai-disrupting-the-billable-hour
Legora deliberately avoids building one general legal-intelligence model, betting on narrow fine-tuned models for specific high-volume tasks instead.
Junestrand called fine-tuning a general legal-intelligence model 'a total waste of time and money,' preferring narrow models for tasks like 'tabular review' (documents multiplied by prompts, e.g. 100 documents x 100 prompts = 10,000 API calls) where a fine-tuned extraction model cuts cost and latency versus a general model competitors are trying to build.
legal-ai-disrupting-the-billable-hour
LexisNexis and Westlaw's legacy data moat is eroding because legal research now requires 100% data completeness, not just breadth, and AI can rebuild that from scratch.
Junestrand said you cannot build a legal research product without literally all the case law, since a top law firm litigating a billion-dollar case needs every relevant precedent, not just the most common 80% - 'the opposite of the power law.' He argued LexisNexis and Westlaw's historic advantage (physically scanning and double-typing case books, including a de facto US reporting arrangement for Westlaw) is now replicable with AI tooling, and both legacy players' stock is being 'crushed' on that uncertainty, while Legora and Harvey's combined revenue already rivals LexisNexis's roughly $2B/year.
data-moats-and-legacy-incumbents
Legal AI adoption is bottlenecked by trust and compliance, not technical capability, which is why most legal-AI startups fail to convert pilots into customers.
Junestrand said 'compliance is our currency' - the technical challenge of proving where legal AI creates value is easy, but selling into a highly regulated, high-stakes industry where a single data leak into a model would be disastrous is what kills most legal-AI companies; Legora hosts sensitive material for weapons manufacturers and governments but has chosen not to offer on-prem or VPC deployment because it slows the roadmap too much.
data-moats-and-legacy-incumbents

Media referenced

Companies

Techniques and frameworks

Summary

This is a special guest-interview episode recorded live in Paris: Jason Calacanis, solo, runs back-to-back interviews with two AI company CEOs whose products are each attacking a specific trillion-dollar service industry. Mati Staniszewski (ElevenLabs) covers the first half on voice AI - the company's revenue ramp, culture, and how it's handling celebrity licensing and impersonation risk. Max Junestrand (Legora) covers the second half on legal AI - how his company is unbundling the billable hour and challenging incumbents like LexisNexis and Westlaw.

Staniszewski walks through ElevenLabs' growth from its first human-sounding text-to-speech model in early 2023 to roughly $600M in revenue and 600 employees, with zero attrition on the original 10-person research team. He describes an organizational model built on small, five-to-ten-person cross-functional pods with engineers embedded even in non-technical functions like legal and talent - partly to build automations, partly to security-check what those teams generate with AI. Notably, the company has never hired a product manager: Staniszewski argues AI now lets one person cover enough of the code/customer/design skillset that a dedicated PM role became unnecessary. The conversation moves into voice-as-identity: ElevenLabs has paid out over $22M to voice talent through a licensing marketplace, worked with Disney and Epic Games to bring the late James Earl Jones's Darth Vader voice into an interactive Fortnite experience, restored voice for a US congresswoman and a bride who'd lost theirs to illness, and built internal safeguards (tracing, moderation, public detection tools) against unauthorized cloning after Calacanis found his own voice cloned from his podcast archive without consent. On competitive dynamics, Staniszewski frames ElevenLabs' defensibility as the interaction layer itself - architecture and labeled audio data, not raw model scale - even as Anthropic and OpenAI encroach on voice, and closes by asserting that AGI-level intelligence has effectively already been achieved in text, just not fully deployed, with voice expected to hit the same bar within the year.

Junestrand opens the legal-AI half with a stark framing of the opportunity: legal services is roughly a trillion-dollar annual market but only about 4% of that spend is software, a gap he expects to close because demand for legal work already outstrips the supply of lawyers. He walks through why the billable-hour model is structurally broken - firms overcharge junior associates and undercharge the partner time that actually carries the value (avoiding a company-ending mistake), a mispricing AI is now exposing as clients like Cooley build direct-to-founder platforms and enterprises pull diligence in-house. Legora itself made four acquisitions in the past year and used its own tool to do all the diligence, closing one deal in 12 days - illustrating the incentive misalignment Junestrand highlights between founders (who want deals done fast) and outside counsel (whose incentive, even unstated, is to drag work out).

The back half of the legal conversation gets into product and data strategy. Legora's "forward deployed lawyer" role, modeled explicitly on Palantir's forward deployed engineers, sits inside law firms like Kirkland & Ellis to help redesign workflows rather than just hand over a tool - a difference Junestrand insists is categorical, not incremental, versus prior legal-tech waves like document management. On the model side, Junestrand rejects building one general legal-intelligence model in favor of narrow fine-tuned models for high-volume tasks (his example: "tabular review," where 100 documents times 100 prompts becomes 10,000 API calls), arguing narrow beats general on both cost and latency, and pushing back directly on the idea that Claude's bundled legal-skills offering is a real competitive threat - he calls it a pipeline generator that drives customers to Legora once they hit its ceiling.

The episode closes on data moats. Junestrand argues you cannot build a credible legal research product without literally complete case-law coverage - "the opposite of the power law," since a firm litigating a billion-dollar case needs every precedent, not just the common ones - which is why LexisNexis and Westlaw built their advantage the hard way (physically scanning and double-typing law books) decades ago. He argues that advantage is now replicable with AI tooling and cites LexisNexis's roughly $2B annual revenue against Legora and Harvey's combined revenue already approaching similar scale, with both incumbents' stock reportedly getting hit on AI uncertainty. He closes by naming trust, not technical capability, as the real bottleneck for legal AI: "compliance is our currency," since a single data leak into a model in a regulated industry would be disastrous, which is why Legora has deliberately chosen not to offer on-prem or VPC deployment despite hosting sensitive government and defense-contractor material.

Notable Quotes

"We've kind of achieved it. We just haven't deployed it." - Mati Staniszewski, on AGI

"The way that business model works is you overcharge for the associates, and you actually undercharge for the partners." - Max Junestrand, on the billable-hour model

"It's the opposite of the power law. You don't just need the top 80%, you actually need all of it." - Max Junestrand, on why legal research requires complete case-law data

"Compliance is our currency." - Max Junestrand, on why trust, not technology, is the bottleneck for selling AI into law

"The motivation of the lawyer is to not have you sue them if they [screw] up the deal... and to make as much money as possible. Which means to drag it out." - Max Junestrand, on outside counsel's misaligned incentives