Context

Max Junestrand is co-founder and CEO of Legora, which he describes as "the agentic operating system for lawyers." Three non-lawyer founders started it in Sweden, were rejected by Y Combinator on their first try, and went from $1M to $100M in ARR in about 18 months (from three engineers to over 750 people), reaching more than 3% of the world's lawyers as active users. In this Y Combinator talk and Q&A, Junestrand walks through the specific product and company-building decisions behind that curve: how they bet on AI in a conservative, high-stakes industry, why they deliberately stopped selling for six months, how they think about models, and how they hire and build culture. It is dense with transferable lessons for anyone building an AI product in a real, unforgiving market.

The Big Idea

When a market is obviously about to be transformed by AI but nobody knows how, the winning move is to march into the space fast, learn it faster than anyone else, and bet that the models will keep improving, focusing your work on delivering the models' value to your market rather than on building the model itself.

Legora picked a space (legal plus AI) they were sure would matter, not a narrow problem, then out-learned everyone about how lawyers actually work. They refused to fine-tune a model, betting instead that frontier models would improve and their job was to channel that value to lawyers.

Key Insights

Bet on models improving, don't fine-tune

The central technical bet. In 2022 and 2023 the tribal wisdom was that you had to fine-tune your own model (Bloomberg reportedly spent millions building a legal model). Legora refused, partly for lack of money and partly from conviction.

• The bet: the models will keep getting better on their own, so your job is to figure out how to deliver the value the models generate to your market, not to build the model.
• The corollary: "build for the world today, and maybe one step ahead," not for the world of the distant future. Even the first ChatGPT was unusable in law, so they shipped what was good enough now (their first real pitch was literally "like ChatGPT but compliant in Europe") and rode the improvement curve.
• Why it matters: in a fast-improving-model world, effort spent hardening around a current model's weaknesses can be wasted when the next model erases them. Position to benefit from the exponential rather than fighting it.

Learn the market faster than anyone

The founders had no legal background, and Junestrand argues domain expertise was not the requirement; willingness to learn the market was.

• How they did it: they cold-emailed lawyers and messaged them on LinkedIn, offering to pay their hourly rate for a lunch to learn how each practice area actually worked. Many took the meeting, and many did not even charge them.
• The evidence: a Swedish VC passed on the pre-seed specifically because the team had no lawyers, and later called it the lesson that domain expertise is not what you need. Junestrand's caveat: for something like quantum or fusion, deep expertise matters more; legal was learnable as they went.
• The takeaway: immerse yourself in your users' real work aggressively and early. Learning velocity, not credentials, is the edge in a market you are new to.

Freeze sales to get the product right

One of the boldest calls: with about $35M in the bank and only 10 people (earning more from interest than from customers), they froze their entire sales motion for six months.

• Why: in law "you only really get one chance to get it right." If you show up and the product fails, lags, or goes down under load, "you are toast." The reliability bar had to be met before scaling sales.
• The result: the freeze is exactly where the $1M-to-$100M curve begins. They rebuilt the product on a platform adaptable to constantly changing models and frameworks, then re-entered.
• The principle: in a market where trust is fragile and failure is punished, deliberately slowing go-to-market to earn reliability can be what unlocks the fast growth, not a delay of it.

Evaluation is the real IP

Junestrand's strongest advice for AI builders: the core muscle to build is the ability to evaluate new models and new use cases, because that is what lets you route work effectively.

• How they built it: they hired lawyers whose job was partly customer-facing and partly building use cases to run evals on. They built "Legora Bench" internally over three years before releasing it.
• Why it matters more now: when only two frontier models existed, model choice was trivial. With many models (they were surprised to find Grok strong on cost-performance), the ability to eval and route across them by use case becomes a durable advantage.
• In law specifically: you often want maximum intelligence, because token cost is tiny next to the human expertise being replaced, so for complex litigation you throw the biggest model at it, while other use cases optimize for cost.

Hire for slope, not intercept

A hiring pattern they had to unlearn: chasing fancy logos on resumes, which Junestrand calls the "Y-intercept" mistake.

• The idea: someone's skill curve might start high (impressive pedigree) but have a flat trajectory. In a company scaling exponentially, you want a steep slope: high growth potential and willingness to work extremely hard, even from a lower starting point.
• The proof: their top seller is 23, jumped out of university with no sales background, and has sold over $10M of Legora.
• The lesson: optimize hiring for trajectory and drive over prestige, especially in a fast-scaling company where people have to grow faster than their starting credentials.

Small-company speed is the superpower

Junestrand is direct that a startup's real edge over incumbents is velocity of iteration on customer feedback. When Microsoft launched Copilot into Word and Outlook, Legora panicked ("we're so screwed, every lawyer already lives in Word"), but Copilot did not work well for a long time, leaving room. Even at 1,000+ people, when a customer problem surfaces on a call, he drops it in the product channel and asks how fast they can turn it around and delight the customer. The principle: identify your superpower (here, speed against slow incumbents) and lean into it relentlessly, while not zigzagging away from the core vision.

Mental Models & Frameworks

Pick a space, not a problem

Two distinct ways to start a company. Legora picked a space: it was obvious that legal plus AI would be big, but not how, so they marched in that general direction and figured out the specifics along the way. The alternative is to pick a specific problem, solve it clearly with measurable ROI, then expand from it. Neither is wrong, but they demand different strategies, and knowing which one you are running keeps you from expecting the certainty of one while playing the other.

Requalify for your job every quarter

Junestrand's model for a fast-scaling leader: running a sales team at $1M ARR is a completely different job than at $150M, so he treats himself as needing to requalify as CEO every quarter. It requires putting the company ahead of your ego, learning continuously, and accepting that your most important job keeps changing (his is now building the executive team). Applied to any role in a fast-growing company: assume the job you were good at last quarter is not the job this quarter, and re-earn it.

Ambition is learnable through peers

Junestrand argues ambition is not fixed; the fastest way to raise it is a peer group of very ambitious people. He was a lazy, video-game-playing student until a new, ambitious friend group made it "click," and he went from zero internships to eight jobs in a year (he cites this as a core value of YC and Paul Graham's essays). Inside Legora he pushes ambition across the whole company, not just the top (even the CFO challenges marketing to be more ambitious). The model: to become more ambitious, deliberately surround yourself with people whose ambition resets your sense of what is normal.

Decision Principles

Principle: One manifesto beats democratic voting

• When: deciding what to build with a growing team.
• Why: early on they made feature decisions by democratic team votes, and "too many chefs" led to working on too many features at once. They replaced it with a single written product manifesto that gathered everything they had learned into one simple document the whole 25-person team rallied around. A clear, owned point of view focuses a team far better than averaging everyone's preferences.

Principle: Run like there is no number two

• When: competing in a winner-take-most market.
• Why: Junestrand frames legal AI as winner-take-most, where "there is only winning, everything else is losing." He wants the company to behave like professional swimmers looking down their own lane rather than sideways at competitors, using rivals as a rallying carrot without being distracted by them. The intensity of acting like there is no safe second place is what let a latecomer overtake bigger incumbents.

Trade-offs & Nuance

Humility culture vs winner-take-all drive

Junestrand describes Jantelagen (the Swedish "Law of Jante," that you should not think you are anyone special). The upside: it supports a culture where the best ideas win and junior people feel empowered to speak up. The downside: it is very hard to build one of the fastest-growing companies in the world if you do not believe you are somebody or that your ideas are better. Legora had to blend US, European, and Asian cultural traits to keep the humility's benefits while adding the conviction and intensity the market demanded. The nuance: a cultural trait can be an asset and a liability at once, and scaling may mean deliberately keeping one side while grafting on its opposite.

Common Mistakes

Mistake: Chasing prestige hires

Optimizing for impressive resumes (the "Y-intercept") over growth trajectory. In an exponentially scaling company, a high-starting but flat-trajectory hire struggles, while a steep-slope hire compounds. The better approach is to select for drive, learning velocity, and growth potential, and to double down repeatedly on that kind of talent rather than on pedigree.

Practical Application

Build an eval and routing capability

• Do: treat the ability to evaluate models and use cases as core IP. Build your own benchmark grounded in your real use cases (Legora hired domain experts to construct and run these), and use it to route different tasks to different models by the cost-versus-intelligence trade-off each one needs.
• Why it works: as the number of viable models grows, systematic evaluation is what lets you exploit new and cheaper models quickly instead of being locked to one, which is a compounding advantage.

Immerse yourself in users' real work

Replicate Legora's learning sprint: go directly to the people whose work you are transforming and buy their time to understand it in detail (they paid lawyers for lunches to learn each practice area). Do this before and during building, and treat learning the market faster than competitors as a first-class goal, not a preliminary.

Slow go-to-market to earn reliability

If you sell into a market where one failure destroys trust (law, healthcare, finance), consider deliberately holding back sales until the product is reliable under real load, the way Legora froze sales for six months. Getting the reliability right first can be the precondition for fast growth rather than a cost to it.

Write a product manifesto

When feature focus is scattering, gather everything you have learned about your users and product into one short, opinionated document and rally the team around it. Use it to cut the feature sprawl that "too many chefs" produces, and to compete by doing a few things better rather than many things adequately.

Questions to Consider

• Are we spending effort fine-tuning or hardening around a current model's weaknesses that the next model may erase, instead of focusing on delivering the model's value to our specific market?
• In the market we are new to, are we learning how our users actually work faster than any competitor, and what would it take to immerse ourselves the way Legora did by paying lawyers for their time?
• Do we sell into a market where a single failure destroys trust, and if so, should we deliberately slow our sales motion until the product is reliable, rather than scaling on a shaky product?
• Have we built a real capability to evaluate and route across models by use case (a benchmark grounded in our own tasks), or are we locked to whatever model we started with?
• Are we hiring for trajectory and drive (steep slope) or for impressive resumes (high starting point but flat), given how fast we need people to grow?

Bottom Line

Legora went from $1M to $100M ARR in 18 months by out-learning a market they had no background in, betting that frontier models would keep improving so they never fine-tuned, freezing sales for six months to earn the reliability a high-stakes market demands, and treating model evaluation as core IP. The throughline for any AI builder: pick a space you are sure matters, learn it faster than anyone, and channel the models' improving value to your users rather than trying to build the intelligence yourself.

Notable Quotes

"You don't have to build for the world over here, you need to build for the world today and maybe one step ahead." (Max Junestrand)

"The amount of things you learn is a function of the amount of discomfort that you are willing to endure." (Max Junestrand)