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

The source is an authorized podcast/event transcript containing two separate interviews, with sponsor reads between/around them. The first interview is with a guest identified at the close as “Mati from 11 left”; the discussion and repeated references clearly concern ElevenLabs, but the transcript does not supply the guest’s full name. The host appears to be Jason Calacanis, based on references to “Jason,” “J-Cal,” This Week in Startups, his bulldogs, and his own investments; this identification should nevertheless be verified before publication.

The second interview is with a guest called Max, discussing an AI legal company rendered variously by the transcript as “Ligora”/“Lagora.” The exact company spelling and Max’s full name should be verified. This interview begins after an Oracle sponsor read.

The host opens by describing the guest’s company as being on a strong growth run while facing intense competition. After an Airwallex sponsor message, he asks about a reported revenue increase from roughly $350 million toward $500–600 million and asks how quickly the company reached those levels.

Mati/ElevenLabs guest says:

• The company started in 2022.
• Its first year was primarily devoted to research and product development.
• The company built what he describes as “the first text-to-speech model that finally could sound human.”
• That model was released at the beginning of 2023.

• It then took roughly:

• 20 months to reach the first $100 million in ARR;

• another 10 months to reach $200 million;
• another five months to reach $300 million.
• He says $300 million was approximately where the company closed the previous year.
• He then states that the company is now at $600 million in revenue. The change from ARR terminology to “revenue” should be checked before publication.

Asked how the company maintains culture during rapid growth, Mati says maintaining culture is critical, particularly through:

• the interview process;
• how people are brought into the company;
• onboarding.

He says the company has approximately 600 people.

He describes ElevenLabs as combining research and product around what he calls a communication platform for AI.

On the research side, the work includes:

• generating speech;
• transcribing speech;
• orchestrating speech during interactions;
• audio more broadly.

On the product side, ElevenLabs aims to cover the customer journey from:

• marketing;
• asset creation;
• international localization;
• customer support using voice agents;
• operational uses of voice agents;
• training;
• sales.

Mati argues that part of the company’s revenue growth reflects the number of business functions it has expanded into.

He says the original team was heavily oriented toward research and engineering and claims that among the company’s first 10 people there has been zero attrition: all remained with the company at the time of the interview.

He credits his co-founder, whom he describes as an exceptional researcher, and says ElevenLabs has assembled people who are genuinely motivated by solving audio and interaction problems.

His view is that, for researchers interested in those problems, ElevenLabs has become “one of the leading, if not the leading place” to work.

The host notes that the company began before generative AI had transformed software development and asks how the new ability of non-developers to create software has changed ElevenLabs internally.

Mati observes that in 2022, much of the technology conversation still centered on crypto and the metaverse.

He argues that this was actually a good period in which to start because the team could spend time concentrating on what it believed would become the future before the current AI frenzy developed.

Mati says ElevenLabs organizes much of the company into relatively small units, typically five to 10 people.

This applies to:

• product engineering;
• industry-specific go-to-market units;
• telecommunications;
• financial services;
• healthcare;
• other areas of the company.

He says one unusual organizational choice is to put engineers inside teams that would traditionally not be engineering organizations.

Examples include:

• the talent team;
• the legal team;
• revenue engineering;
• go-to-market teams.

These embedded engineers have two functions.

First, they create automations and bring software directly into the team.

Second, they help colleagues use AI while also providing a security check on what those colleagues build or deploy.

Mati describes a balance:

• Employees who are barely using current coding and AI collaboration software may be missing important productivity improvements.
• Employees using it excessively or carelessly may also create risk.
• People in functions that previously did not build software may now be capable of creating applications or automations without necessarily being capable of reviewing what those systems are doing behind the scenes.
• Embedded engineers therefore provide technical and security oversight.

The host reinforces the concern by pointing out that democratized software creation creates problems if poorly understood software reaches production, leaks information, is abandoned when its creator leaves, or quietly deteriorates.

The host asks how AI is changing product management and management generally.

Mati responds jokingly, “we're fired them all, right? We don't have any PMs.”

When asked whether ElevenLabs previously had product managers, he says it never did.

His reasoning is that even before the current AI wave, the ideal product person would have been someone capable of:

• coding;
• understanding customers;
• understanding design.

People who are genuine experts in all three are rare.

ElevenLabs therefore hires people who are:

• highly capable in at least one field;
• strong enough to understand at least one additional field.

Mati argues that AI has altered this equation because a person who previously had only limited ability in an adjacent discipline can sometimes move from an amateur level to an advanced level using AI, though not necessarily to true expert level.

That means employees are less dependent on separate functional specialists for every task.

He gives growth engineering as an example: one person may now be able to design an experiment, implement it, ship it, evaluate whether it works and iterate.

Mati says ElevenLabs has an advantage because it can heavily dogfood its own products.

If the company wants customers to build voice agents, its own teams need to be able to build and use voice agents.

This expectation extends beyond traditional engineering functions and into go-to-market work.

One example is an inbound AI SDR agent.

Alongside the normal form on the company website, a prospective customer can call an AI agent.

Mati says this creates two advantages:

1. People can provide information more quickly and naturally than by filling in a conventional form.
2. They often provide considerably more information.

That additional information can help ElevenLabs route the person more rapidly to the correct problem area and employee.

He sees this as an example of AI tooling making employees more capable at their underlying jobs while simultaneously improving the company’s understanding of its own product.

The host contrasts modern AI voice agents with traditional telephone automation, which he describes as an unpleasant “voice jail” experience that encouraged users to demand an operator or repeatedly press zero.

He argues that newer systems have crossed an important threshold because:

• their speech fidelity is substantially better;
• they can understand intent more precisely;
• users can interrupt them naturally;
• users do not feel guilty about interrupting;
• there is no need for social small talk;
• a user can get directly to the issue.

Mati agrees that behavior is changing and predicts that users may increasingly ask companies to “give me an AI agent” rather than asking for a human representative.

Mati says enterprise customers have been a major source of ElevenLabs’ recent growth and praises the company’s enterprise sales team.

He says the technology has improved because several elements are finally working together:

• reliability;
• orchestration across AI models;
• knowledge;
• integrations;
• the conversational experience.

He describes a major quality improvement over the previous 12 months, particularly the previous six months.

He predicts a coming period in which a customer can open a company website, speak with an agent and have that agent already possess relevant information from the customer’s previous interactions.

He believes the broader interface will evolve further:

• interfaces will adapt to the individual user;
• voice will operate in the background to locate information;
• support will move from reactive assistance after a user identifies a problem toward proactive assistance before the user explicitly asks.

He says ElevenLabs is already observing examples of this transition.

The host argues that historical speech-to-text systems suffered from both poor fidelity and social awkwardness.

He cites lawyers using Dragon Dictate roughly a decade earlier, with headsets, and says people often felt foolish speaking audibly to computers in an office.

He says a different behavior is emerging: people quietly whispering prompts to computers and agents.

He mentions a product transcribed as “Whisper Flow” and asks whether it uses ElevenLabs.

Mati says it uses ElevenLabs and several others and compliments the product.

The conversation then turns to physical controls.

The host says he uses a foot pedal with the voice interface.

Mati mentions a device or devices transcribed as “Plot” / “Plot, pocket” [transcript unclear], describing them as highly useful, particularly at events.

Mati argues that, provided people are informed that recording is taking place, such technology could capture conversational information that would otherwise disappear.

Instead of manually writing a few notes, a person could potentially:

• preserve the conversation;
• automatically populate specific notes;
• generate follow-up actions.

The host says he has three pedals under his desk.

Using his pedal with the voice product, he presses down to activate it, speaks, and releases it.

He argues that typing itself creates friction when working with an LLM: users become tired of typing and therefore shorten or stop their prompts.

Speaking removes much of that constraint.

He describes himself jokingly as a “professional bullsh-h-hirt artist” and says he can deliver a one- or two-minute stream-of-consciousness prompt.

His observation is that LLMs are particularly good at processing this large volume of loosely structured spoken thought.

He says the method has “changed everything” for him.

Mati adds that spoken thought can include a person beginning one idea and then changing direction.

Instead of losing the first thought, the AI receives both pieces of context, which he says can improve the final answer.

He sees this as another example of humans gradually changing the way they communicate with machines.

The host asks what differences ElevenLabs sees in human behavior when speaking with an LLM or voice agent.

Mati gives an emotionally sensitive example from financial services.

He names customers or companies transcribed as:

• Revolut;
• “Clarnow” [transcript unclear];
• Pag Bank.

One common use case involves reminding customers about payments or trying to collect payment from people who have not responded.

Mati says some people feel ashamed when explaining their real financial circumstances to another human.

He claims they can be more open with an AI system, disclosing what actually happened and providing more information.

He therefore argues that AI can remove or reduce a social and emotional barrier that exists in a human-to-human interaction.

He also says users tend to be:

• more abrupt or “snappy” with AI agents;
• comfortable giving quick answers;
• willing to interrupt the agent.

That enables conversations to move faster toward the desired outcome.

Mati says this behavior means designers also have to modify the interaction model.

The host raises the risks around celebrity and public-figure voices.

He tells a personal story involving his interest in bulldogs.

Someone told him they enjoyed his bulldog videos, but he did not know what they meant.

They showed him a channel featuring dogs telling jokes.

According to the host, the creator had taken audio from the This Week in Startups archive and used ElevenLabs to create a synthetic version of his voice for the channel.

The host says this occurred approximately one or two years earlier.

Rather than initially reacting angrily, he contacted the creator, described the work as flattering, and asked how it had been done.

The person told him ElevenLabs was used.

The host says he later contacted ElevenLabs about safeguards.

The host gives his own description of United States law surrounding voice and identity.

He says, in substance, that a person cannot simply take someone’s voice and use it commercially, although impersonation may be possible in parody and under fair use.

He contrasts U.S. regulation jokingly with France and performs an extended Donald Trump impersonation.

These legal statements are the host’s characterization, not verified legal guidance, and should be independently checked before publication.

The host later gives another example involving his own advertising.

His staff wanted to clone his voice to repair mistakes in recorded ads, such as an incorrect promotional code.

He says the ElevenLabs system prevented them from cloning his voice and required him personally to complete an authorization step.

He presents this as evidence that ElevenLabs had introduced stronger identity protections.

He also mentions Jamie Foxx as an example of celebrity involvement, though that specific relationship is not confirmed by Mati in the response and requires verification.

Mati says voice is a form of both identity and intellectual property.

People recognize individuals by the way they speak and associate emotion with that voice.

He notes that impersonation existed before AI and therefore is not an exclusively AI-created problem.

However, he says ElevenLabs believes that because it is helping lead the development of voice technology, it must also help lead development of safeguards.

He describes three areas of protection.

ElevenLabs traces generated content so it can investigate and act when necessary.

Mati says the company performs moderation at both the voice and text level.

He says content can be flagged or blocked when, for example, it is commercial in a prohibited context or appears intended to scam someone. The exact moderation rule described in the transcript is somewhat unclear and should be confirmed.

The company is developing systems that allow people to upload a sample and receive information about whether it appears to be AI-generated.

Mati says this detection work applies not only to ElevenLabs-generated material but also to material created using other open-source models.

Mati then turns from risk to the economic opportunity created by licensed voices.

He says ElevenLabs partnered with Matthew McConaughey on something rendered in the transcript as “creating World Cup” [transcript unclear].

The host performs a McConaughey impression during the exchange.

Mati’s broader point is that generative voice technology can take a recognizable voice beyond English into languages such as:

• Spanish;
• Italian;
• Portuguese.

He says the emotional qualities of the original voice can remain recognizable across those languages.

Mati cites a partnership/use case transcribed as “master-cats”, almost certainly a proper-name transcription requiring verification.

He describes a model in which content that used to be static becomes interactive.

His example is Gordon Ramsay teaching cooking in a kitchen and potentially reacting interactively to what the learner is doing.

The host interprets this as creating AI characters or interactive instances of recognizable people that subscribers can converse with.

Mati agrees.

The host jokes by imitating Ramsay and shouting that food is raw.

Mati says ElevenLabs created a marketplace in which people can:

• create a synthetic version of their voice;
• authenticate it;
• make it available for others;
• earn money when it is used.

He says ElevenLabs had paid more than $22 million back to its community of voice talent at the time of the interview.

The host observes that this creates a different economic model for voice actors who may previously have been paid largely by the hour.

He asks whether a voice actor can spend roughly an hour recording, create an ElevenLabs voice and then license it.

Mati responds: “100%.”

Pricing can operate in more than one way:

• a creator can accept ElevenLabs’ default pricing/distribution structure;
• in other cases, the creator can choose pricing.

Mati says the appropriate model depends on the use case.

He again emphasizes opportunities involving dynamic content and multiple languages.

Mati calls voice restoration some of ElevenLabs’ most important work.

The transcript says the company worked with people who lost their voices because of “ILS” and “fraud cancer” [both transcript unclear; medical conditions must be verified rather than silently corrected].

He cites U.S. Congresswoman Jennifer Wexton, saying she lost her voice but wanted to continue inspiring people and demonstrating that meaningful work remained possible.

He refers to what he describes as a first speech delivered in Congress using the recreated voice; the precise event and wording require verification.

Mati then tells what he calls a particularly heart-warming story about a woman who had planned to marry but lost her voice before the wedding.

According to him, she and her partner later repeated their vows using a recreated voice.

He describes the emotional effect on the family of hearing her vows in that voice.

His conclusion is that voice has unusually strong emotional and human connective power.

The host raises James Earl Jones and the Darth Vader voice.

He says his understanding is that Jones, before his death, made an arrangement involving Disney so his Darth Vader voice could continue to be licensed for his family.

The host presents this as his understanding and asks whether ElevenLabs was involved in subsequent uses.

Mati says he cannot disclose everything involving newer projects.

He does, however, discuss a public gaming example:

• Epic Games/Fortnite, in partnership with Disney, launched an interactive Darth Vader.
• Players reaching a certain stage could interact with Darth Vader live.
• The character could assist them with missions.

Mati sees this as evidence of a new model in which a person’s likeness, identity or publicity rights can be extended into interactive experiences rather than remaining limited to prerecorded media.

He expects this category to become increasingly common.

Mati identifies Headspace as another public relationship.

He says Headspace localized substantial amounts of content.

The host jokingly ranks Headspace behind “Com”, apparently another meditation company name rendered incorrectly by the transcript, and then remembers he had invested in it.

They discuss the possibility of moving beyond localized static meditation toward interactive, personally tailored meditation.

Mati expresses interest in a future in which a meditation experience could adapt directly to an individual.

The host jokes about using David Sacks’ voice in a meditation product.

The host says ElevenLabs is competing with companies led by people including:

• Dario at Anthropic;
• Sam at OpenAI.

He characterizes those companies as wanting ElevenLabs’ business and asks whether ElevenLabs risks helping future competitors by using their models.

Mati’s answer begins with model neutrality.

ElevenLabs wants its platform to make multiple model families available, including:

• Anthropic;
• OpenAI;
• open-source models;
• Google models.

He argues that model agnosticism benefits customers.

Customers can build:

• an agent harness;
• orchestration;
• a voice layer;
• communication and marketing experiences;

without making the entire application dependent on one underlying frontier model.

Mati acknowledges that distinctions between models, platforms and applications are becoming increasingly blurry.

Nevertheless, he defines ElevenLabs’ core layer as interaction and communication.

He claims ElevenLabs has successfully competed in:

• text-to-speech;
• speech-to-text;
• conversational turn-taking;
• music.

He describes the research team as “a set of magicians” that repeatedly produces competitive advances.

Mati offers several reasons ElevenLabs can remain competitive against larger model companies.

He argues that in audio research, architecture matters more than scale alone.

Simply making a model bigger is not necessarily sufficient; the way the model operates must change.

He says audio requires specialized data.

Although enormous quantities of audio exist, much of it is not properly labeled for the intended research.

ElevenLabs therefore invests heavily in labeling.

Mati says the company built an internal operation of more than 1,000 contractors who label audio assets.

The correct communication workflow differs substantially across:

• financial services;
• healthcare;
• telecommunications.

He says the product team spends significant effort learning the exact workflow each industry requires.

ElevenLabs also wants a broader ecosystem including:

• integrations;
• reusable voices;
• templates;
• agent authentication;
• other components that prevent customers from having to begin from scratch.

The host raises the possibility of reinforcement learning, model distillation and data leakage by upstream providers.

Mati says ElevenLabs is aware that some companies are continually attempting to understand how they can distill and use data.

He calls this an existing problem.

He says ElevenLabs has mechanisms that can slow such activity down, but he explicitly does not claim those mechanisms can stop it completely.

Asked whether the company has an internal open-source fallback or plans to replace frontier providers, Mati says ElevenLabs is looking more closely at where its own expertise can be extended.

He draws a boundary.

The company does not intend to make its focus:

• general knowledge work;
• coding.

Instead, it wants to own the technologies associated with interaction and the combination of those communication components.

He says ElevenLabs is spending more time on that area.

He also says the team enjoys competing directly against the largest AI laboratories and occasionally demonstrating that it can outperform them in its chosen domain.

The host speculates that ElevenLabs must be sending tens of millions of dollars annually to frontier-model companies.

Mati does not confirm a figure.

He says only that ElevenLabs “ship[s] a good amount” and describes the frontier labs as good partners.

He says the spend corresponds to value created for ElevenLabs and praises the work those companies have done.

Mati says it is remarkable that current agents can sometimes feel as though a user is communicating with another human.

He suggests that the traditional Turing test feels increasingly close to being satisfied in ordinary text interaction.

He states a goal for voice: within the year, ElevenLabs hopes to reach a point at which any conversation feels like speaking with another human.

The host says he believes the technology is already there for some applications, depending on what questions are asked.

He argues that tests invented 30 or 40 years ago to define artificial intelligence or AGI have effectively been passed and that society now needs more demanding tests.

The host proposes an intentionally extreme new standard: whether an AI can become “more intelligent than every single person on the planet times 10.”

He suggests society is “kind of there on AGI.”

The exchange ends with the idea that some capabilities have already been achieved, although not necessarily deployed universally.

Mati says: “There are definitely places where we did achieve it.”

These AGI statements are opinions and predictions rather than verified technical conclusions.

Following an Oracle Cloud sponsor read, the host asks Max about his company’s growth.

Max says growth is not technically exponential but has been a sustained 50% quarter-over-quarter for the previous seven quarters.

He then says that, as of the close of the previous Tuesday, the company had become one of the fastest enterprise companies with a direct-sales motion to grow “from one to 150, bidding Sierra with one quarter.”

This passage is materially unclear.

Possible missing units and the word rendered as “bidding” must be checked against the audio before publication.

The host frames legal AI as potentially able to reduce the cost of conventional legal services dramatically.

He gives the example of a startup that had:

• reached approximately $1 million in revenue;
• raised multiple funding rounds;
• hired a couple of dozen employees;
• still had no corporate lawyer.

When the host asked who reviewed contracts, the founders responded, in the transcript’s rendering, “chat cheap E, bruh.”

The same answer was given when he asked about:

• the capitalization table;
• HR matters.

The host jokes that the startup will make an entertaining future diligence target.

He says he warned the founders that Series A investors would eventually ask for issues such as IP assignments to be professionally reviewed.

The founders said they had learned how to prepare them by asking ChatGPT.

The host presents the story as evidence that some early-stage companies are already replacing a substantial amount of traditional legal work with general-purpose AI.

Max agrees that the behavior exists but emphasizes that a seed-stage startup operates very differently from a major U.S. bank.

He frames the broader market in economic terms.

According to Max:

• approximately $1 trillion per year goes into legal services;
• the market is highly fragmented;
• legal-technology software spending is approximately $40 billion;
• that implies roughly 4% software and 96% services.

He calls that balance “bananas.”

His expectation is that software will take a much larger share of work currently represented as service revenue.

Max argues that AI is not purely about eliminating existing legal spending.

He says the legal-services market is supply-constrained: demand for legal help exceeds the supply of lawyers and services.

As technology expands capacity, providers can therefore use it to:

• serve new kinds of client;
• enter new market segments;
• package new legal products;
• satisfy demand that previously went unmet.

Asked for an example, Max names a firm rendered as “Kooley” [verify exact name].

He says the firm created a platform directly for startup founders.

A founder can log in to a software interface containing:

• the firm’s material;
• its precedents;
• startup-oriented resources;
• embedded workflows that can review contracts.

Max sees this as important because it begins to break the traditional law-firm business model based on charging very high hourly rates for associates.

Max says large law firms effectively:

• overcharge for associates;
• undercharge for partners.

The host initially disputes the idea that partners are cheap and says he recently saw approximately:

• $1,800 per hour for a senior lawyer;
• $800 per hour for associates.

Max says Kirkland can charge as much as approximately $4,000 per hour.

His deeper argument is not that partner rates are numerically small.

He argues that 30 minutes of a top partner’s work on a critical issue can create value far in excess of the nominal hourly rate.

Examples include:

• bet-the-company litigation;
• avoiding a mistake that could cost tens of millions of dollars.

His view is that law firms historically capture that value by charging disproportionately for large amounts of associate labor because the billable-hour model provides no better way to price the partner’s highest-value judgment.

The host asks whether enterprises are responding to legal costs by taking work in-house.

Max says absolutely.

He gives his own company as an example.

He says the company had acquired four businesses so far that year.

For those acquisitions, it performed diligence internally using its own AI product.

The fastest transaction, he says, went from letter of intent to closing in 12 days.

The host uses the acquisition example to argue that founders and outside lawyers often have different incentives.

His characterization is:

• the founder wants the transaction completed quickly;
• the lawyer wants to minimize the risk of being sued for an error;
• the lawyer also has an economic incentive to generate fees;
• those incentives can encourage a slower process.

The host says this may create an incentive to drag out a deal even if the lawyer never explicitly describes it that way.

Max agrees with the broad point.

The discussion then notes that law firms are experimenting with alternative pricing structures, including:

• fixed fees for transactions;
• fixed fees for fundraising;
• success-based compensation in litigation.

Max sees this as evidence that AI is beginning to reshape the business model of one of the world’s largest professional-services industries.

Asked how incumbent firms are reacting, Max says there is “a lot of anxiety and a lot of fear.”

He notes that major firms are extremely large and profitable.

He says Kirkland Ellis produces approximately $10 billion a year and has roughly 4,000–5,000 lawyers.

He says profit per partner can be between approximately $5 million and $10 million annually.

Against that backdrop, Max says AI represents both:

• an existential threat;
• an existential opportunity.

He describes helping law-firm leadership understand how to respond as a major part of his own job.

His reasoning is that his AI company will succeed only if its customers successfully adapt as well.

Max describes a role at his company called a legal engineer.

He compares it to Palantir’s forward-deployed engineer model.

Instead of forward-deployed software engineers, his company has what he characterizes as forward-deployed lawyers.

Their job is to sit with law-firm partners and help transform the business from a pre-AI model to a post-AI model.

The host compares this transition to earlier law-firm adoption of PCs and document-management systems, when firms moved from printed drafts and physical libraries/storage toward digital workflows.

Max agrees with the analogy but emphasizes a critical difference:

• PCs and document management created relatively modest productivity gains.
• AI can perform significant parts of the underlying legal work itself.

The host asks whether junior legal jobs will survive.

Max says:

“The job will exist. The tasks will be different.”

His reasoning is institutional.

A partner-led profession still needs to develop future partners.

Just as software organizations need junior engineers who can eventually become experienced senior engineers, law firms need junior lawyers who can build enough experience to become senior lawyers and partners.

What changes is the training path.

Max says junior lawyers will no longer develop primarily by:

• locking themselves in a physical data room;
• manually reading every document;
• marking every issue;
• faxing results.

Nor will the modern equivalent simply be:

• opening a virtual data room;
• using Control-F to search files.

Instead, the lawyer’s job increasingly becomes orchestrating the AI agent performing that work.

The host notes that legal practice has historically been highly jurisdiction-specific.

Even within the United States, attorneys are licensed by state and legal rules vary.

He uses non-compete agreements as an example, asserting that they are generally not enforceable, or not very enforceable, in California but can be considerably more enforceable around Boston.

That legal characterization must be independently verified.

The host asks whether AI could enable lawyers and corporate legal departments to operate more effectively across jurisdictions.

Max says this is already beginning to happen.

Max describes two main categories of data beneath the product.

The product uses information belonging to the firm or enterprise, including:

• precedents;
• organizational data;
• internal legal material.

Max says the company also performs the difficult work of gathering:

• cases;
• legislation;
• regulatory updates;

for jurisdictions around the world.

He says this process is painful but becomes a meaningful data moat when performed at scale.

Max offers a cross-border example.

Suppose the general counsel of a California company lands the company’s first customer in South Africa.

Historically, the GC might call an outside lawyer, who knows another lawyer in South Africa, who eventually provides an answer.

Max says his company’s system can instead adapt to South African law directly.

He cites a real case in which the customer could obtain what he characterizes as an approximately 80% accurate response immediately, giving the legal team something useful from which to begin work.

He does not claim the response was a complete substitute for professional legal review.

He argues that as accuracy improves, more sophisticated work becomes possible.

Max says much of the world’s legal and regulatory information has never previously been structured in the way AI can now structure it.

He connects this not only to lawyers but to the many people involved in:

• public policy;
• regulation;
• related legal processes.

He considers this lack of structure a major source of inefficiency in society.

The host describes LexisNexis as a longstanding dominant legal-data company with a major moat in case law and regulatory data.

He says LexisNexis generates only a few billion dollars a year and suggests that the combined revenue of Max’s company and Harvey may already be in the hundreds of millions.

Those revenue comparisons are the host’s assertions and require verification.

He compares the incumbents’ potential reaction to the Tyrannosaurus rex appearing in the rear-view mirror in Jurassic Park.

He asks whether Max intends eventually to acquire LexisNexis or destroy its business.

Max does not answer in terms of buying or destroying LexisNexis.

Instead, he argues that established information providers often have difficulty becoming AI-native businesses.

He claims legacy organizations struggle because:

• they cannot attract the same talent;
• they do not operate at the same working tempo;
• internal politics slow decision-making.

He says that early in the AI cycle, many people assumed the incumbents owning the largest data collections would automatically win.

In his view, the market is beginning to show that data ownership alone is insufficient.

Max says his company sees opportunities to partner with content providers rather than necessarily reproduce every source independently.

He says it is already doing this in smaller jurisdictions, naming:

• Germany;
• France;
• Spain.

The United States is different, he says, because legal research is unusually concentrated.

The host asks whether Westlaw is the other major incumbent.

Max confirms that the U.S. market is effectively dominated by Westlaw and LexisNexis.

A later exchange suggests those incumbent businesses or their parent companies are being adversely priced by investors because of AI uncertainty, but the exact company/security being discussed is unclear and should be verified.

The host discusses how previous legal databases assembled historical case law.

His description includes:

• going to courthouses;
• obtaining old cases and books;
• scanning physical documents;
• sending material to India;
• having two people independently type or OCR documents;
• comparing the results for discrepancies.

His point is that the databases needed extremely high accuracy because legal researchers cannot tolerate missing or corrupted cases.

Max agrees with the importance of completeness.

The host notes that modern AI is far better at many of the tasks that legacy databases once performed manually.

However, he and Max also discuss continuing physical constraints.

They say source books may still need to be obtained and scanned.

They refer specifically to the importance of preserving page citations.

The host asserts that Westlaw effectively has a monopoly-like relationship concerning official U.S. case reporting; that characterization requires verification.

He mentions:

• a Harvard project;
• CourtListener or a similarly named source [verify exact reference].

He says those alternatives are trying to build broader public legal-data systems but considers their current coverage insufficient for the highest-stakes commercial legal research.

The host argues that legal research differs from many ordinary information problems.

For a high-stakes case, an attorney cannot accept a database that contains only the most commonly useful 80% of precedent.

He gives the hypothetical example of a litigator at Wachtel [verify spelling] handling a billion-dollar matter or litigation involving Elon [identity/context unclear].

The lawyer needs confidence that the system includes all potentially relevant cases.

The host calls this the opposite of the power law.

Max agrees with the requirement for comprehensive source material.

They discuss the practical burden of:

• obtaining records from courthouses;
• potentially paying per page;
• acquiring physical books;
• scanning those books;
• preserving page citations.

Max notes that there are methods other than the exact workflow the host describes but agrees that, in practice, assembling comprehensive legal data remains difficult.

The discussion distinguishes two generations of legal technology.

A lawyer searched a database, located relevant cases and then personally did the substantive analytical work.

Max says newer agents, particularly following releases identified as Opus 4.5 and 4.6, can increasingly:

• analyze witness statements;
• combine them with relevant cases;
• develop case strategy;
• perform more end-to-end workflows.

He argues that the industry is moving from AI merely augmenting lawyers toward AI doing meaningful portions of the work.

The human role increasingly becomes:

• orchestrating agents;
• supervising them;
• managing their output.

He compares this transition to what is already happening in software development.

The host asks whether Max’s company works with Anthropic and OpenAI.

Max says yes.

Asked whether the company spends millions or tens of millions of dollars on model tokens, Max says “Absolutely.”

The host then asks whether those same frontier companies are competing with him.

Max says:

“They are not competing in our product category at all.”

The host immediately adds: “For now.”

Max discusses Claude’s legal capabilities or offering.

He characterizes it from the outside as essentially:

• a bundle of Markdown skill files;
• several integrations.

He nevertheless sees value in its existence.

It demonstrates to prospective customers how useful generative AI can be in law.

Customers can begin experimenting with a general-purpose product.

In Max’s view, they eventually encounter a ceiling or discover that the offering is shallow relative to specialized legal workflows.

At that point, some contact his company.

He therefore describes general-purpose legal AI experiments as a pipeline generator rather than simply competitive pressure.

The host asks whether the company plans to use open-source models to build its own foundation model.

Max makes a clear distinction.

He says:

• he does not believe in fine-tuning or building general-intelligence models for this purpose;
• he considers trying to build a general legal-intelligence model a “total waste of time and money.”

He does believe in highly specialized models for narrow, high-volume use cases.

The reasons are:

• lower cost;
• lower latency;
• better economics when the same narrow operation runs at massive scale.

Max gives a product example called tabular review.

The computational workload is roughly the number of documents multiplied by the number of prompts.

His example:

• 100 documents;
• 100 prompts;
• therefore 10,000 API calls.

In that setting, a fine-tuned model specialized in extracting contract information could be highly valuable because it performs one narrow task repeatedly and efficiently.

Max contrasts this with competitors attempting to build broad, generalized “legal intelligence,” which he does not think is economically sensible.

The host raises the risk of confidential litigation or regulated information leaking into an underlying language model.

He says the consequences could be disastrous.

Max responds:

“compliance is our currency.”

He says this heightened responsibility is one reason selling AI into legal organizations is difficult.

According to him:

• many legal-AI companies exist;
• relatively few make it successfully through enterprise adoption;
• identifying potentially useful features is comparatively easy;
• securely getting those products into real customers is much harder.

He says his company solved the distribution/compliance problem relatively early.

Once the product is accepted inside a customer, he says expansion becomes significantly easier.

Max connects the difficulty of entering regulated legal customers to his company’s acquisition strategy.

He says the ease of expanding once the company is inside a trusted customer relationship is one of the factors behind its M&A strategy.

Earlier in the interview he had said the company had completed four acquisitions that year.

The exact acquisitions and date context should be verified independently.

To illustrate the trust and compliance demands, Max claims his company hosts or handles material involving:

• national secrets;
• contracts belonging to weapons manufacturers;
• government customers.

These are significant claims and should be verified before publication, particularly the phrase “national secrets.”

The host asks whether these sensitive customers force the company to deploy its product on-premise.

Max says:

“No, we don't do on-prem.”

When asked whether that is on the roadmap, he again indicates no.

He argues that deploying separately into customer VPCs is extremely time-consuming.

It creates dependencies that would:

• slow the product roadmap;
• slow execution.

The interview ends after this exchange.

Exact quote:
“We built the first text-to-speech model that finally could sound human.”

Context: Explaining the company’s initial 2022–2023 research and product period.

Topic: ElevenLabs’ original product thesis.

Exact quote:
“we are at $600 million in revenue.”

Context: After describing the path from $100 million to $300 million ARR.

Topic: Company growth.

Exact quote:
“from the first 10 people, we had zero attrition, everybody is still at a company”

Context: Explaining culture and continuity of the original research/engineering team.

Topic: Hiring and company culture.

Exact quote:
“we embedded engineers. in every place. And even in the places which aren't engineering.”

Context: Describing ElevenLabs’ unusual operating structure.

Topic: AI-native organizational design.

Exact quote:
“we're fired them all, right? We don't have any PMs.”

Context: Joking response when asked how ElevenLabs thinks about product management.

Topic: Product-management structure.

Exact quote:
“Never did.”

Context: Answer to whether ElevenLabs had ever employed product managers.

Topic: Organizational design.

Exact quote:
“it's slowly becoming that you will be asking for, give me an AI agent effectively on the call.”

Context: Discussing the transition from disliked automated phone systems to useful conversational agents.

Topic: Consumer adoption of voice AI.

Exact quote:
“It will shift from reactive to proactive to help you get that help before you potentially ask for it.”

Context: Prediction about customer-service interfaces.

Topic: Future of AI customer support.

Exact quote:
“So I'll give it a one to two-minute prompt. Then I let go. And it has changed everything. everything.”

Context: Describing pedal-assisted voice prompting.

Topic: Voice interfaces for LLMs.

Exact quote:
“With AI, people are much more open to share what actually happen”

Context: Financial-services example involving users discussing difficult payment situations.

Topic: Human psychology and AI agents.

Exact quote:
“the voice is identity and IP.”

Context: Introduction to safeguards and licensed voice use.

Topic: Voice rights and synthetic identity.

Exact quote:
“we are leading on other development. We also need to lead on a lot of the safeguards.”

Context: Why ElevenLabs believes it has a responsibility to build abuse protections.

Topic: AI safety and impersonation.

Exact quote:
“Today we paid back over $22 million back to the community of talent.”

Context: Describing the authenticated voice marketplace.

Topic: Voice licensing economics.

Exact quote:
“our research team is a set of magicians”

Context: Explaining why he believes ElevenLabs can compete with much larger AI companies.

Topic: Research advantage.

Exact quote:
“It's the architecture that matters, not the scale.”

Context: Explaining ElevenLabs’ research strategy in audio.

Topic: Specialized-model competition.

Exact quote:
“we have a few mechanisms to stop it, slow it down, not stop it.”

Context: Discussing other companies potentially distilling or using generated data.

Topic: Model/data protection.

Exact quote:
“But any interaction and how you can combine all those pieces together and make sure this is great. We want to own.”

Context: Explaining what ElevenLabs may build itself rather than relying on frontier providers.

Topic: Competitive strategy.

Exact quote:
“we hope this year we'll do that same thing for voice, where any conversation feels like you are speaking with another human.”

Context: Discussion of Turing-test-like experiences.

Topic: Voice AI roadmap.

Exact quote:
“it's sustained 50% quarter over quarter for the last seven quarters.”

Context: Opening description of the company’s growth rate.

Topic: Legal-AI company growth.

Exact quote:
“It's a trillion dollars every year into legal services”

Context: Explaining the size of the legal-services market.

Topic: Legal-AI market opportunity.

Exact quote:
“there's 4% software, 96% service, which is bananas.”

Context: Comparing roughly $40 billion in legal-tech spending with the claimed trillion-dollar legal-services market.

Topic: Legal-industry economics.

Exact quote:
“you overcharge for the associates, and you actually under charge for the partners.”

Context: His explanation of the traditional billable-hour model.

Topic: Legal pricing.

Exact quote:
“when something like AI comes along, that poses existential threat and existential opportunity.”

Context: Discussing highly profitable incumbent law firms.

Topic: AI disruption of professional services.

Exact quote:
“The job will exist. The tasks will be different.”

Context: Answer to whether junior-lawyer jobs will survive AI.

Topic: Employment and legal training.

Exact quote:
“It's orchestrating the agent that will be doing that work.”

Context: Contrasting future legal work with physical and virtual data-room review.

Topic: Agentic legal work.

Exact quote:
“once you start to do that at scale, it builds a real data most.”

Context: Transcript rendering while describing collection of global cases, laws and regulatory updates; “most” is likely a transcription problem and must be checked.

Topic: Legal-data competitive advantage.

Exact quote:
“some of the existing providers and the sort of legacy players have a really hard time pivoting into becoming AI-native businesses.”

Context: Response to a question about LexisNexis and other incumbents.

Topic: Incumbent disruption.

Exact quote:
“You don't just need the top 80% you actually need all of it.”

Context: Explaining why incomplete case-law databases are insufficient for high-stakes litigation.

Topic: Legal research requirements.

Exact quote:
“moving us from a world where AI is just augmenting to AI is actually really doing things”

Context: Discussing agents combining cases, witness statements and strategy.

Topic: AI automation versus augmentation.

Exact quote:
“I think that's a total waste of time and money.”

Context: Rejecting the strategy of building or fine-tuning a generalized legal-intelligence model.

Topic: AI-model strategy.

Exact quote:
“I do believe in very narrow models for narrow use cases”

Context: Explaining where specialized models make economic sense.

Topic: Legal-AI architecture.

Exact quote:
“compliance is our currency.”

Context: Asked how the company protects highly sensitive customer information.

Topic: Security and enterprise trust.

Exact quote:
“It's actually quite easy to understand where you can build value. But getting it to the customer is very hard.”

Context: Explaining why many legal-AI startups fail to gain enterprise adoption.

Topic: Enterprise distribution.

Exact quote:
“No, we don't do on-prem.”

Context: Asked whether sensitive government and defense-related customers require on-premise deployment.

Topic: Infrastructure strategy.