How a 25-year-old built a $675M legal AI startup with no legal background

Executive overview

Legal software was fragmented and static — AI never performed well enough on unstructured text until LLMs changed the equation overnight. Max Junestrand co-founded Legora at 23, interviewed 100 lawyers before writing a line of product code, and grew from 10 to 100 people in under a year.

Legora is an AI-powered workspace for lawyers: a web app with an agent and tabular review grid, plus a Word add-in. It raised an $80M Series B and now serves tens of thousands of lawyers daily across Europe and the US.

Incumbents had data moats but couldn't ship; Legora won by out-shipping teams 100x its size and partnering with firms rather than selling to them.

Why legal AI works now

• Pre-LLM models could match clauses only when the language was identical across documents — meaning was invisible to them
• GPT-3.5 showed that unstructured legal text could be understood semantically, not just pattern-matched
• The same underlying model can replace dozens of separate point solutions (templating, redlining, research, translation)
• Cost of intelligence is falling fast enough that tasks once priced out of scope — deep cross-judgment research, full due diligence — are becoming standard

What Legora actually does

• Web app: chat agent that chains steps (research → conform to firm style → draft memo), using MCP to plug into firm-specific tools and workflows
• Tabular review (grid): load any number of documents as rows, run any queries as columns — built for scale, parallel execution, citation accuracy, and legal-specific chunking (definitions, cross-references)
• Word add-in: Cursor for lawyers — AI assistant in the right-hand panel that reads the document, proposes edits, and runs multi-step playbooks against open contracts
• Playbooks: collections of approve/reject rules with fallbacks; a lawyer's legal team creates the standard, then every sales rep or compliance officer can run it consistently

Selling to law firms

• Law firms rejected legal tech for years; what changed is that AI broke the service equilibrium — a firm doing faster due diligence steals clients
• Approach: "we win if you win" — align incentives, position as a long-term strategic partner rather than a vendor
• Large firms have innovation departments; target those people first, make a single partner and their team look like rock stars, then expand
• Bottom-up adoption is structurally blocked — software must clear IT security and data privacy checks before touching client data
• Buyers are now refusing 5-year lock-ins; 1-2 year contracts are the norm, and firms evaluate a vendor's rate of change as much as the product itself
• European compliance (data hosted in-region, no training use, no retention) was table stakes from day one and became a durable trust signal

Entering a market with no domain expertise

• Interviewed 100 lawyers before building; offered to pay their hourly rate for lunch — none charged, all showed up
• "We don't know exactly where the future is going, but neither do you — let's work together" was the honest pitch that built early partnerships
• Hiring lawyers directly into product teams came later; naivety in the early days was useful for questioning legacy workflows
• For any vertical AI founder: learn the domain deeply first, then build; your outsider perspective is a feature, not a bug

Scaling from 10 to 100 people in a year

• Spent the first months after raising Series A hardening reliability and scalability — deliberately paused sales until the system could comfortably onboard 1,000 lawyers a day
• Scaled from ~25 people in October to 100 six months later — roughly two hires per week
• Prioritised hiring former founders: they bring ownership, problem-solving agency, and the ability to run sub-companies within the company
• Always seeded new city hubs (New York, London, Madrid, Paris, Berlin) with the best Stockholm people first
• The culture is whatever the first hires embody — deliberate about velocity, high expectations, flat structure

Building defensible vertical AI

• Don't compete with AI labs on general capability; they ship constantly
• Find the narrow layer where domain knowledge, workflow integration, and trust matter — that's where moats form
• Build infrastructure so that when models improve, every feature improves automatically (hot-swappable model layer, classification routing by query complexity)
• The line between software and service is blurring: the category leader serves as a strategic transformation partner, not just a tool vendor
• Lawyers in 5-10 years will spend most of their time reviewing and directing AI agents, not doing the underlying work themselves