Legal AI Is Repricing the $1 Trillion Legal Services Market

Both Max Junestrand of Legora and Mati Staniszewski of ElevenLabs describe businesses built on top of frontier models while resisting the idea that access to a model is the business.

Legora uses Anthropic and OpenAI and spends “millions or tens of millions” on tokens, Junestrand says. ElevenLabs makes Anthropic, OpenAI, Google, and open-source models available to customers. Neither executive claims the underlying models are irrelevant. Both describe them as powerful inputs. But their accounts of durable advantage sit elsewhere: in domain-specific material, workflow design, integrations, specialized models, and the ability to deploy into settings where the work or data is sensitive.

For Legora, that means a system built around firms’ own precedents and organizational information, alongside cases, legislation, and regulatory updates across jurisdictions. It also means Legal Engineers—lawyers who work directly with law-firm partners to redesign work around the technology—and narrow models for repetitive, high-volume extraction tasks.

For ElevenLabs, the differentiating layer is interaction: text-to-speech, speech-to-text, turn-taking, orchestration, voices, industry integrations, and the permissions required to use a person’s voice. Staniszewski says the company has more than 1,000 contractors labeling audio assets and is building workflows for sectors such as financial services, healthcare, and telecommunications.

That leaves the companies with different hard problems. Legal AI has to be complete enough for a high-stakes researcher to rely on it across jurisdictions. Voice AI has to make interaction useful while determining who can authorize, license, and challenge the use of a voice that may be inseparable from its owner’s identity.

Max Junestrand frames legal AI as an attack on an unusually large service economy. He puts annual spending on legal services at roughly $1 trillion, compared with about $40 billion spent on legal technology. In his formulation, legal is still 96% service and 4% software.

That imbalance does not mean all legal spending becomes software revenue. Junestrand’s argument is that technology can both absorb manual work and expand the range of services firms can offer. Providers can turn their precedents and materials into software-assisted products, serve segments that were previously uneconomic, and give customers more direct access to workflows that once required lawyer time.

He cites Cooley’s platform for startup founders as an example. The firm has put its materials and precedents into a system that gives founders access to startup-related material and contract-review workflows. That begins to break with the conventional arrangement in which associates conduct large volumes of repeatable work under a billable-hour model.

Junestrand describes the economics of that model bluntly: firms “overcharge for the associates” and undercharge for partner judgment. His point is not that partners are cheap. A senior lawyer’s half hour in bet-the-company litigation may be worth far more than even a very high hourly rate. But the historical way to price that judgment has included charging heavily for the associate layer.

When document review, comparison, extraction, and initial analysis become cheaper and faster, firms have less reason to treat associate hours as the mechanism that supports the rest of the partnership. Junestrand says firms are already experimenting with fixed fees for transactions or fundraises and success fees in litigation. The pricing question becomes more explicit: what is the value of a partner’s judgment, rather than how many junior hours can be attached to it?

Legora has used its own tooling for acquisitions, Junestrand says. The company acquired four businesses during the year of the interview and performed the diligence in-house. Its fastest transaction went from letter of intent to closing in 12 days.

In the exchange, Jason Calacanis argued that an acquirer’s interest is to close quickly while outside counsel’s incentives include avoiding liability and billing for work. Junestrand agreed that enterprises are increasingly asking whether more legal work can be brought inside.

The work does not disappear in Junestrand’s account, and neither does the need to train lawyers. Firms still need junior lawyers if they are to develop future partners, just as engineering organizations need junior engineers to develop senior ones. But the apprenticeship changes. The old route—reading every document in a physical data room, marking issues, and passing the material along—is no longer the central training mechanism. Nor is the future simply better search over a virtual data room. The junior lawyer’s work becomes orchestrating the agent doing the review and managing the result.

ElevenLabs makes a parallel argument about internal software work. Mati Staniszewski says the company has never had product managers. Rather than hire people who must be expert in engineering, customer understanding, and design simultaneously, it looks for people with deep skill in one area and meaningful understanding of another. AI, in his view, lets those people operate more effectively across boundaries.

A growth-oriented employee, for example, can design an experiment, ship it, observe what happens, and return with evidence rather than wait for a chain of specialists. ElevenLabs uses its own technology in that way: it built an inbound AI sales-development agent that prospects can call instead of only completing a web form. Staniszewski says callers provide information more quickly and often leave more useful context, allowing the company to route them to the right person or problem.

The operating model has limits. ElevenLabs embeds engineers in teams beyond conventional engineering, including talent, legal, and go-to-market functions. Their role is to build automations and help colleagues adopt AI tools, but also to ensure that people who can create software with those tools are not deploying something they cannot assess securely behind the scenes. Staniszewski’s view is not that every function should build unchecked software; it is that AI use needs to become widespread while technical judgment remains close to the work.

The distinction between access to legal material and the ability to act on it is central to Max Junestrand’s account of Legora.

The company sits on top of a customer’s own precedents and organizational data, he says, but also gathers cases, legislation, and regulatory updates across jurisdictions. That collection effort is difficult and expensive. It is also, in his telling, a source of defensibility because legal research cannot tolerate partial coverage in high-stakes work.

Junestrand’s example is a California general counsel whose company has landed its first customer in South Africa. Instead of waiting for one lawyer to find another lawyer in the relevant market, he says Legora can adapt its work to local South African legislation. In one example he cites, a user received what he characterized as an 80% accurate response immediately and could begin working from that starting point.

The immediate response changes the starting point for the legal team: rather than beginning with a referral chain, it has a jurisdiction-specific working answer to investigate and develop.

Completeness remains a practical constraint. Junestrand says historical decisions may require physically obtaining books, scanning them, and preserving page citations. A litigator at a firm handling a billion-dollar matter cannot rely on the most common authorities alone; the system must have the relevant corpus. He describes the United States as particularly hard because Westlaw and LexisNexis dominate legal research, while access to reported cases and archival material is difficult.

That is why he does not treat legacy ownership of data as a decisive advantage. Established providers, he says, move more slowly, struggle to match Legora’s operating tempo, and face political constraints inside large organizations. Legora is already partnering with content providers in smaller jurisdictions including Germany, France, and Spain, he says. The United States is more difficult precisely because a research product cannot be credible without comprehensive coverage.

But research completeness is only one side of the product. Earlier-generation legal databases allowed a lawyer to search, locate a case, and perform the substantive work. Junestrand says agents can now begin combining witness statements, cases, and other materials into case strategy and end-to-end work. The lawyer’s role shifts toward directing and evaluating that process.

Legora does not think it needs to train a general legal-intelligence model to do this. Junestrand calls building or fine-tuning a general model a waste of time and money. Instead, he favors narrow models for narrow tasks that occur at enough scale to make cost and latency meaningful.

Its Tabular Review feature is his example. The workload is the number of documents multiplied by the number of prompts: 100 documents reviewed against 100 prompts creates 10,000 API calls. A model specialized in extracting contract data can make sense there, Junestrand says. That does not imply that a company should build a universal legal model.

General models can still be useful to Legora’s commercial position. Junestrand says Anthropic’s legal offering helps establish how applicable AI is to law, but describes it as a bundle of Markdown skills files and integrations rather than a direct competitor to Legora’s category. Customers can begin experimenting with a general tool, encounter its limits, and then seek a system with legal workflows, organizational context, and research coverage.

Trust is part of the same sales problem. Junestrand says “trust and compliance” are Legora’s currency and that winning legal customers is hard not because it is difficult to identify valuable use cases, but because sensitive organizations must approve the system. He says Legora works with governments and entities handling national secrets and weapons-manufacturing contracts.

The company does not offer on-premises deployment and does not plan to, Junestrand says. Deploying into a virtual private cloud would create dependencies that slow the roadmap and execution. His position is that Legora can meet demanding customers without building a separate on-premises product for each environment.

Mati Staniszewski describes ElevenLabs’ growth as evidence that voice has crossed from a content-generation feature into a way companies operate customer interactions.

The company was founded in 2022 and released what Staniszewski calls its first human-sounding text-to-speech model at the beginning of 2023. It took roughly 20 months to reach $100 million in annual recurring revenue, another 10 months to reach $200 million, and five months to reach $300 million, he says. By the interview, he said ElevenLabs had reached $600 million.

Staniszewski attributes the most recent growth in particular to enterprise use. The company’s product spans speech generation, transcription, conversational orchestration, and voice interactions across marketing localization, customer support, operations, training, and sales. A voice agent, as he describes it, has to combine reliable voice technology with the relevant customer knowledge and integrations—not simply read an answer in a realistic voice.

He says quality changed sharply in the prior 12 months, and especially over the prior six. The old phone-tree experience is giving way, in his view, to an interface in which a person can call an agent, have it recognize context from earlier interactions, and get help without navigating a site or explaining the problem repeatedly. Over time, he expects the interface to become more proactive, using voice in the background to surface help before a customer explicitly asks for it.

The interaction itself differs from human service. In financial-services use cases involving missed payments or debt collection, Staniszewski says people are often more willing to explain their actual circumstances to an AI than to a person. Shame can be less of an obstacle. Users are also more direct: they interrupt the agent, answer briefly, and move quickly to the outcome they want. That changes the product requirement. A system designed around politeness and rigid call flows will not match an interaction in which interruption is normal.

ElevenLabs’ internal structure is designed around that wider use of AI. The company had 600 employees at the time of the interview and uses teams of roughly five to 10 people across product, engineering, go-to-market, and industry-specific efforts. Staniszewski says every function increasingly needs to use the tools it sells. If ElevenLabs offers customers voice agents, its own teams need to know how to create and use them.

Staniszewski says model architecture matters more than scale alone for voice performance, and that ElevenLabs has invested in a contractor operation of more than 1,000 people to label audio assets. The company also assembles industry workflows, integrations, voice inventory, templates, and authentication mechanisms so customers do not start from scratch.

It remains model-agnostic at the language-model layer. Customers can choose Anthropic, OpenAI, Google, or open-source models while retaining the orchestration, voice experience, and integrations around them. Staniszewski says that independence lets customers build an agent harness without being locked into one model provider.

At the same time, ElevenLabs is developing more of its own capability for interaction. Staniszewski says some companies are continually trying to distill or use data, and that ElevenLabs has mechanisms to stop or, more realistically, slow that behavior. Its response is not to build a broad model for coding or general knowledge work. It is to deepen its specialization in audio and communication.

The more voice resembles a person, the more it becomes a matter of rights and consent rather than only model quality.

Mati Staniszewski says ElevenLabs treats safeguards as a responsibility that comes with being a leading developer of voice technology. The company traces generated output so it can take action when needed; moderates both voice and text inputs to flag or block commercial and scam-related misuse; and has developed systems that let a person upload audio and receive an indication of whether it is AI-generated. He says those detection systems cover ElevenLabs output as well as some open-source models.

Those controls matter because a voice can be misused, but they also make licensed use more practical. ElevenLabs has a marketplace where a person can create and authenticate a voice, make it available, and earn from its use. Staniszewski says the company has paid more than $22 million to participating talent.

Voice owners can either accept a default pricing structure that allows ElevenLabs to distribute their voice more broadly or select their own price, he says. For voice actors, the model can turn a recorded performance into a reusable licensed asset rather than a one-time hourly engagement.

Staniszewski’s examples extend beyond commercial voice work. He cites Congresswoman Jennifer Wexton, who lost her voice and used the technology for a speech delivered in Congress. He also describes a woman who lost her voice before marrying and later repeated her vows so her family could hear them in her voice. In both cases, the value he emphasizes is the restoration of a personal connection.

The licensed-interaction model also applies to entertainment. Staniszewski says ElevenLabs worked with Matthew McConaughey on a voice that could carry across languages while preserving the recognizable emotional quality of the performance. He describes MasterClass moving from static content toward interactive lessons, and says Fortnite launched a live interactive Darth Vader experience with Disney and the estate.

For ElevenLabs, authorization is not a peripheral legal issue. It determines whether a recognizable voice can become an interactive product at all: a licensed performer can speak in multiple languages, a fictional character can respond to players, and a voice actor can set terms for reuse. The company’s product challenge is therefore broader than synthesizing realistic speech. It is making speech usable in a way that customers, rights holders, and the people behind the voice will accept.