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

Mati's account of ElevenLabs is unusually specific about shape. The first text-to-speech model that sounded human shipped in early 2023; 20 months to $100 million ARR, then 10 months to $200 million, 5 months to $300 million, and roughly $600 million now, against 600 employees. Two organisational choices stand out. There are no product managers at all — never were — because the ideal profile could code, understand the customer and understand design, and rather than hunt for that unicorn they hire people expert in one of those and literate in another. And engineers are embedded in teams that are not engineering teams: talent, legal, go-to-market. Their second job is a security check on whatever those teams now build for themselves, on the reasoning that people newly able to create software are frequently not able to review it.

The competitive answer is the technically interesting part. Asked why Anthropic and OpenAI have not simply taken voice, Mati's answer is that on the research side it is the architecture that matters rather than the scale — the model has to operate differently — and that the data problem is labelling rather than availability. ElevenLabs runs an internal team of over 1,000 contractors labelling audio assets. On distillation he is blunt that other companies are trying to extract the data and that the available mechanisms slow it down rather than stop it. The platform stays deliberately model-agnostic across Anthropic, OpenAI, Google and open-source models so customers build their harness and agent orchestration without depending on any one, and he draws the boundary explicitly: ElevenLabs will not chase knowledge work or coding, only interaction. On safeguards, the stack is three parts — trace everything generated, moderate at both the voice and text level, and offer detection so anyone can check a sample, including samples from other labs' open-source models. The marketplace has paid out over $22 million to voice talent, and the work he singles out is restoring voices to people who lost them to ALS and throat cancer.

Max's Legora half reframes legal as a supply problem. Roughly a trillion dollars a year goes into legal services against about $40 billion of legal software — 4% software, 96% service — in a market where demand already exceeds available lawyers. His structural claim about the billable hour is that firms overcharge for associates and undercharge for partners because there is no other way to price the thirty minutes of partner time that actually matters. Legora's answer to deployment is copied from Palantir: forward-deployed lawyers, whose job is sitting with partners at firms like Kirkland and moving the business from pre-AI to post-AI. He is dismissive of two popular strategies — he does not believe in fine-tuning or building general intelligence models, calling it a waste of time and money, while endorsing narrow models for narrow high-volume jobs like tabular review, where 100 documents times 100 prompts is 10,000 API calls and cost and latency actually respond. He also makes a point that inverts the usual power law: legal research is the one domain where the top 80% of the data is worthless, because a litigator betting a case needs all of it. On deployment he refuses on-prem outright, saying VPC deployments create dependencies that slow the roadmap.