Vertical SaaS: The Legora $100M ARR Playbook

Vertical SaaS is software built for one industry's workflow instead of one business function across every industry. A horizontal product sells the same help desk to a hospital, a bank, and a law firm. A vertical product sells only to the law firms, and in exchange it gets to model how legal work actually runs: the documents, the review steps, the billing units, the regulator, and what a finished work product looks like.

That trade used to be the cautious one. You gave up market size for product fit, and you accepted a slower, smaller company. AI turned the trade around. When the cost of producing software collapses, as Y Combinator puts it in its call for SaaS challengers, generic capability stops being a moat, because everyone gets it from the same three or four labs. What does not collapse is knowing precisely how one industry works and being trusted enough to be handed its real data.

The clearest evidence available right now is Legora, a Swedish company selling an AI workspace into law firms. It went through Y Combinator's Winter 2024 batch, hit about $1M in ARR at general availability in October 2024, and crossed $100M in ARR roughly 18 months later. Co-founder and CEO Max Junestrand walked through how on YC's stage with Gustaf Alströmer, and the number is the least interesting part of it. Every decision he describes is a business model decision rather than a technology one.

What vertical SaaS actually means

Three things separate a vertical product from a horizontal one, and only the first is obvious.

It targets one industry's workflow, not one function. The unit of design is the process a professional in that field runs, end to end, not a feature like search or reporting.

It absorbs the industry's constraints as product surface. Retention rules, privilege, audit trails, and who is allowed to approve what are not compliance checkboxes bolted on later. In a vertical product they are the schema.

It sells against the labor line, not the software line. This is the economic difference founders underrate. A horizontal tool competes with other tools for a software budget. A vertical AI product competes with the hours a firm currently bills for the same work, and that budget is much larger. It is the same shift behind service as software, where the buyer is paying for output rather than seats, and it is why pricing an AI product is a strategic decision rather than a pricing page exercise.

The reason all three matter more now than in 2019 is that the first two used to be a cost and now they are the asset. Modeling the workflow was expensive engineering nobody thanked you for. Today it is the part a model provider will not do for your customer, because doing it means learning one industry deeply enough to be wrong about it in specific ways.

Set against horizontal SaaS, the differences line up like this:

• Unit of design. Vertical: one industry's end-to-end process. Horizontal: one business function, generalized until it fits everyone.
• Buyer budget. Vertical: increasingly the labor line, because the product does part of the work. Horizontal: the software line, competing with other tools per seat.
• Defensibility in the AI era. Vertical: domain depth, permitted data access, and evaluation sets nobody else has. Horizontal: mostly scale, brand, and distribution, since the capability itself is now cheap.
• Market ceiling. Vertical: bounded by the size of one profession, but multiplied by every country where that profession works the same way. Horizontal: larger in theory, and far more contested.
• Time to trust. Vertical: years, because the buyer is handing over regulated material. Horizontal: weeks, which is also why the switching costs are lower.

Legora is the case study in this piece, but the pattern generalizes. AI for field service is the same model aimed at home services contractors, and AI ERP is the version aimed at the finance system of record.

Why the vertical model got sharper in the AI era

Y Combinator's argument to founders is that the decades of accumulated code that protected legacy software is no longer the barrier it was. Their framing is blunt: "AI has collapsed the cost of producing software by 100x or more, and that changes everything." If that is true, the follow-on question is what is left to defend, and the answer is everything that is not code.

For a vertical company, four things are left, and each one takes time to build:

1. Access. Being permitted to read a firm's documents, email, and matter history is a multi-year trust process, not an integration ticket.
2. Workflow specificity. Knowing that a diligence review has these seven steps and this failure mode is knowledge that lives in customer conversations, not in a model.
3. Evaluation data. Knowing what a correct output looks like in your field, and being able to prove it, is a private asset. Generic benchmarks do not tell a law firm whether an answer is usable.
4. Behavior. The habits your users build inside your product are switching costs you did not have to invent.

Note what is absent from the list: the model, the interface, and the feature set. Those are the parts a competitor rebuilds fastest. That is the same conclusion as the broader case that AI is compressing every classic moat: what survives is the layer a model cannot supply on its own. Verticals are simply the shape of company where more of the business sits in that layer by default.

The bundle beats the point solution

The most useful decision in Junestrand's account happened when Legora was small enough that the decision looked wrong.

In October 2024, going into general availability with about 30 people, the company had three features: an assistant, a tabular review tool, and a Word add-in. Junestrand says they wrote a three page product manifesto that committed to being the best at all three and combining them. His summary of the bet: "if we can be the best at all these three and bundle them, we will win."

The uncomfortable part was the scoreboard. Legora was at $1M in ARR. A competitor doing nothing but the tabular review piece was near $50M. In Junestrand's words, "They were doing 50 times our revenue focused on one single thing." Every piece of standard startup advice says the company doing one thing 50 times better is the one to copy. He reports that Legora has since passed that competitor and won over a meaningful number of its clients.

This is not an argument against focus. It is an argument about what you focus on. The point-solution company focused on a feature. Legora focused on a workflow that happened to require three features, which is the harder thing to copy precisely because it looks unfocused from outside. Y Combinator names the same play in its challenger thesis: take ten point solutions and bundle them into a single suite.

Two conditions make the bundle work, and without them it is just an unfocused roadmap:

• The customer already experiences the pieces as one job. A lawyer moving between chat, a review grid, and Word is doing one task in three tools. The seam is the customer's pain, not yours.
• You can genuinely be best at each piece, not merely present. A bundle of mediocre components loses to a great point solution. The manifesto only worked because it committed to winning each one.

Junestrand's broader version of this is a warning about horizon. "you need to have a slightly longer horizon strategy of how you're going to win," he said, contrasting it with being opportunistic in a fast-moving market. A bundle strategy is exactly the kind of bet that looks worse than the alternative for about a year and then stops looking worse all at once.

The moat question every vertical founder gets asked

Alströmer put the modern version of the question to Junestrand: founders in the 2010s were asked what happens if Google does this, and founders now are asked what happens if OpenAI or Anthropic does this.

Junestrand's reframe is the part worth stealing. He reaches for the AWS precedent, pointing at how companies like MongoDB positioned against a platform that could obviously have built a database, and asks "what did they build which wasn't natural for a bigger platform like AWS to build". Then he restates the real question underneath: not whether a lab will enter your category, but "what's defensible in your business" if model intelligence keeps rising on a straight or exponential line.

His answer is a short list of inputs and outputs: the proprietary data, and then, in his words, "What are the workflow modes? What is the behavior that we're teaching our users?"

The honest version of this answer includes its own failure case, and Junestrand names it. If models get so capable that one of them writes all the code, fetches all the data, and figures out everything needed to solve a task, his conclusion is that we should all go have a pina colada instead. Read that as the boundary condition: at that point there is no vertical business, and no horizontal one either. He does not think that is the end state, and neither do I, but it is the correct way to hold the question: your strategy should be the one that survives a lot more model progress, not the one that requires progress to stall.

The practical test I would apply before writing a line of code:

• If a frontier lab shipped a very good general agent for your customer's job next quarter, what would still be true about your company? If the answer is nothing, you have a feature.
• What do you have that the lab cannot buy? Customer data you are permitted to hold, an integration into a system of record, a regulatory position, an evaluation set built from real work product.
• What is unnatural for a big platform to do? Selling to a few hundred law firms one partner at a time is unnatural for a company optimizing for a horizontal API business. That is the gap, and it is a durable one.

There is one more piece Junestrand is explicit about: "we understood that we had to get big fast". A vertical is defensible once you are established in it, and undefended while you are getting there. The window between those two states is the actual risk in this business model, which is why speed is a strategic requirement rather than a personality trait.

One workflow, many countries

The standard objection to vertical SaaS is the ceiling. One industry, therefore one small market, therefore a small company.

The counter is that a professional workflow repeats across geographies far more cleanly than consumer behavior does. Junestrand's version, describing an expansion from San Francisco to Chicago, Texas, New York, London, Stockholm, Germany, India, and Australia: "it turns out lawyers work the same way all over the world", give or take. Local law differs enormously. The work of reviewing a document, running diligence, and drafting to a template does not.

That is where a vertical actually compounds. You pay the domain cost once, then amortize it across every market where the profession exists. Legora said in its March 2026 funding announcement that it had grown from 40 to 400 people over the previous year, and by the time of this talk, a few months later, Junestrand puts headcount close to 500 across those markets. "This time last year we were 40 people," he said.

The second answer to the ceiling objection is the budget one from earlier. When your product does part of the work rather than helping someone do it, you are quoting against a labor line, and the ceiling moves. Legora raised a $550M Series D at a $5.55B valuation led by Accel in March 2026. The old vertical SaaS ceiling had no room for a number like that. If you want to see the same physics in other professions, the YC company hub for AI in legal is a useful map of who is attacking which workflow.

Distribution is part of the business model

Vertical markets are small enough that distribution is a strategy decision, not a growth channel. Legora's version has two halves, and founders usually copy the wrong one.

The half that mattered first was founder-led selling into a market with low expectations. Junestrand describes carrying a briefcase around Stockholm and finding that genuine enthusiasm was itself the differentiator, because "these chief innovation officers, these knowledge managers, these legal partners had never seen somebody be excited about selling them technology ever." He is candid that the product at that point was not good. What moved deals was energy plus social proof: the largest firm in the region was already a customer, so the next one did not want to be the holdout. In a vertical, that reference effect is unusually strong, because every buyer knows every other buyer.

The half that got attention later was the brand campaign. Junestrand's read on how his own category markets itself: "It is the most boring, the most bland." Legora's response was to hire the actor Jude Law for a campaign playing on his name, which The Drum covered as a deliberate move away from describing the product at all. Junestrand says they chased him for about six months, were told it was impossible, and got there by showing him customer testimonials rather than a pitch.

Nobody should go hire a film star off this. The transferable part is that category-level blandness is an exploitable asset. When every competitor markets identically, being legible and human is cheap differentiation, and in a market of a few thousand buyers it reaches most of them. That only works after you know exactly who you are selling to, which is the same discipline as writing a real ideal customer profile before you scale a sales motion, and choosing deliberately between top-down and bottom-up selling.

What agents change about vertical SaaS economics

The most recent shift Junestrand describes is the one that most affects how you should build now.

Legora's original design assumption was augmentation: help a lawyer do their individual task faster. What changed is that model capability plus accumulated trust plus tool access now allows the product to do multi-step work on its own. His example is diligence in an M&A transaction: point the agent at an unstructured data room, have it restructure the file tree against a template, give it the diligence questions, and let it work. Some of those runs take 20 to 30 minutes. The interaction moves from a conversation to a delegation, which he compares directly to how engineers work with coding agents.

Two consequences matter for anyone building a vertical product.

Evaluation becomes the bottleneck, not the model. Junestrand names it: "one of our main bottlenecks is evals for end-to-end work products." Scoring a single answer is easy. Scoring a finished work product produced over 30 minutes and many steps is a research problem, and it is the thing that gates every release. If you are building in a vertical, your evaluation set is a core asset and should be resourced like one from the start. The practical starting point is in our guide to LLM evals for founders.

Coding is your leading indicator. "I feel like we're also trailing like 6 months behind code," Junestrand said, on the grounds that code is easier for models: it is verifiable, the context problem is simpler, and the models are trained on it heavily. That gives every vertical founder a free forecast. Whatever is happening in AI coding tools now is a preview of what your domain gets in roughly two quarters, and you can design for it before it lands. That is the same logic as building for the next model rather than the current one, applied to a specific industry.

I would add the security note here, since it is the part that gets skipped. The reason Legora can run agents across a firm's documents and email is trust that was earned before the capability existed. If you are building a vertical product, the access boundary, the audit trail, and the question of what the agent is allowed to do unsupervised are architecture decisions you make early, not a security review you pass later. In a regulated profession that boundary is your product, not your policy.

Where the vertical model breaks

Four honest failure modes, because the Legora story is a survivor's account and survivors are a biased sample.

• The market really is too small. The workflow repeatability argument works for professions that exist globally at scale. It does not rescue a vertical with 300 total buyers and no international equivalent. Do the arithmetic on buyers times realistic contract value before you commit.
• You are early and never get big fast enough. The window Junestrand describes is real, and most vertical companies die inside it, out-executed by a better funded entrant while their moat is still theoretical.
• The workflow you modeled gets deleted rather than improved. Modeling a process deeply is a bet that the process persists in roughly that shape. If capability advances enough to collapse the whole workflow into one step, the depth you built becomes a liability.
• You bundle before you are excellent at anything. The inverse of Legora's play. A three-feature product where all three are mediocre loses to any focused competitor, and the manifesto does not save you.

None of these argue against verticals. They argue for choosing one where the buyer count and the contract value clear the arithmetic, and then moving faster than is comfortable. For a wider view of how these decisions connect to product, engineering, and go-to-market, the AI for startups pillar maps the rest of the operating picture, and Henrique Dubugras's playbook is a useful companion on building a vertical financial product for a single customer type.

What to do this week

• Write down the workflow you are targeting as a numbered list of the steps a professional actually performs today, with the time each takes. If you cannot produce that list from memory, you do not yet know the vertical well enough to build in it.
• Run the ceiling arithmetic honestly: number of buyers in your industry worldwide, times a realistic annual contract value, times a plausible share. If that number does not justify the company you want, pick a different vertical now rather than in year three.
• Answer the lab question in one paragraph: if a frontier model shipped an excellent general agent for your customer's job next quarter, what about your business would still be true. Keep only the answers that are data, access, integration, or habit.
• Audit your bundle honestly. For each component, decide whether you can be the best at it or are merely including it. Cut the ones you cannot win, and commit publicly to the ones you keep.
• Start the evaluation set this week, from real customer work product, not from synthetic examples. In a vertical, it is the asset that ages best and the one you cannot buy later.
• Look at what AI coding tools do today and write down what the equivalent would be in your domain in two quarters. Build toward that version, not the current one.

Choosing which layer you can actually own, deciding what the model is allowed to do alone, and building the trust and evaluation around it is the same operating job in every industry. That is what we teach in the AI Operating System for Startups.

Sources

• How Legora Went From YC to $100M ARR in 18 Months (Y Combinator), the conversation this article distills, with Legora co-founder and CEO Max Junestrand interviewed by YC group partner Gustaf Alströmer: the three-feature bundle manifesto, the $1M versus $50M comparison, the moat framing against foundation model labs, the global workflow argument, founder-led selling into law firms, and evaluation as the bottleneck for end-to-end agent work.
• SaaS Challengers (Y Combinator), on the collapse in the cost of producing software and the bundling play against legacy point solutions.
• Legora funding and scale, verified across TechCrunch and the company's own Series D announcement. The $100M ARR milestone and its 18 month timeline: The Next Web.
• Profiles: Max Junestrand, Gustaf Alströmer, and Legora on Y Combinator. The Jude Law campaign and Legora's marketing approach: The Drum.