How a Venture Studio Picks Which Verticals to Build

A venture studio that launches 3-4 ventures a year spends a quarter of its annual portfolio every time it picks a vertical. Get the vertical wrong and that is not a rounding error. It is three months of build capacity and $500K-1.5M aimed at a market the studio cannot win. So portfolio construction reduces to one decision made three or four times a year. Which vertical earns a build slot, and which large market gets passed over even when it looks tempting. This is selection at the studio level, a different question from defending any one venture's moat.

Avante Ventures is a venture studio building AI-native companies in Brazil and Latin America, and the question of how studios pick verticals is one we answer in front of every build. The short version is a four-part test. A vertical earns a slot only when it combines a large under-digitized services market, an available domain operator with deep scar tissue, a workflow where AI changes the unit economics, and a path to proprietary data that compounds. Miss one and the slot is better spent elsewhere.

The backdrop is the vertical AI thesis. Vertical AI that owns an industry workflow is proving more defensible and better valued than horizontal copilots, and Brazil happens to supply an unusual number of qualifying verticals. You can read the longer thesis at /why-avante. What follows is the test itself, and the discipline of using it to say no.

Why selection is the whole game

A venture studio is supply-constrained on purpose, and that constraint is the reason selection decides everything. With three or four slots a year, the studio cannot let portfolio math absorb its misses the way a fund writing 30 checks can. Every slot is a real bet that has to clear the bar before it is spent.

The model earns its return premium precisely because it concentrates. Per the Global Startup Studio Network, studio ventures posted an average internal rate of return that rounds to roughly 50% against roughly 19% for traditional venture-backed startups, which Avante cites as studio IRR of ~50% versus an industry-standard ~19% for traditional VC, roughly 2.5x over realistic time horizons. That is the studio-model benchmark, never any single studio's realized return. The honest caveat is that this figure runs on a self-reported sample of surviving studios, which is why measuring studio performance honestly matters, so read it as the ceiling the model can reach, not a number any one studio is owed. The traditional-VC figure traces to institutional benchmarks like the Cambridge Associates US Venture Capital index, built from quarterly fund financials across four decades.

That premium only holds when the operating partner can go deep. A studio venture reaches Series A in 25.2 months against 56 for a traditional startup, and 72% of studio ventures reach Series A versus 42% of traditional ones, per GSSN data via Bundl. Depth does not scale to dozens of bets at once. So the studio buys its edge by saying no far more often than it says yes. Selection is not the step before the work. It is the work.

The four-part test for a vertical

A vertical earns a build slot when four conditions hold at once. This is a gate, not a scorecard. A vertical that is strong on three dimensions and closed on the fourth still fails, because the missing condition is usually the one that decides the outcome.

• A large, under-digitized services market. The vertical has to be big enough to matter and far enough behind on software that the work still runs on phone calls, spreadsheets, and PDFs. A huge market already well served by software is a knife fight, not an opening.
• An available domain operator with deep scar tissue. The studio needs a co-founder with 10+ years inside the industry, carrying the regulatory, relationship, and process knowledge no deck can teach. If that operator is not findable and recruitable, the vertical fails no matter how good the market looks.
• A workflow where AI changes the unit economics. AI has to convert a labor-heavy task into software, not bolt on a feature. Per a16z, that shift can expand revenue per customer by 2-10x, on top of the 2-5x that fintech embedding already delivered.
• A path to proprietary data that compounds. The workflow has to throw off data the studio can accumulate and that makes the product better with use. Without a compounding loop, a well-funded incumbent or a horizontal model eventually catches up.

Why vertical AI clears the bar

The test points at vertical AI rather than horizontal copilots because vertical AI is proving both more defensible and better valued. The evidence is no longer theoretical. Per Bessemer Venture Partners in September 2024, LLM-native vertical companies were reaching 80% of traditional SaaS average contract value, posting roughly 400% year-over-year growth, and holding roughly 65% gross margins, with vertical AI market capitalization predicted to be at least 10x the size of legacy vertical SaaS.

What separates a durable company from a thin wrapper is ownership of the workflow plus a data loop. Bessemer ties vertical AI moats to proprietary data, depth of product integration, and economic value delivered, not to model access, which everyone shares. a16z frames the same point through the system of record. The company that owns the workflow captures the labor budget, not just the software budget, and U.S. software spending of $313B is only about 3% of the $10.5T spent on labor. Toast reached $1.5B ARR with over 80% of revenue from embedded financial services, which is what owning a vertical workflow looks like at scale.

The data loop is the part founders overstate most, so the test stays strict about it, and we draw the line precisely in how data network effects actually work in vertical AI. Per the NfX Network Effects Manual, network effects account for roughly 70% of the value created by tech companies since 1994, yet data advantages are weaker than commonly believed. Data is a real moat only when usage continuously updates a dataset central to the product, the way Waze improves with every trip, not marginal the way a recommendation feed is. For the studio, that distinction is the fourth condition made concrete. A vertical passes only when its workflow generates data that is both proprietary and load-bearing.

Why Brazil supplies so many candidates

Brazil produces an unusual number of verticals that clear the first condition, because services account for roughly 70% of Brazilian GDP with low software penetration across those sectors. That figure comes from IBGE data reported in July 2024, with the services share sitting 12.7% above its pre-pandemic level. A large, growing base of economic activity still run on manual workflows is a deep bench of exactly what the test screens for first.

The capital backdrop tells the studio these markets are open rather than crowded. LATAM startups raised $4.2B in 2024, up 27% from the prior year, with Brazil capturing close to half of all regional funding, per Crunchbase. Funding is recovering from a weak 2023 and remains well below the 2021 peak, which means most verticals do not yet have a well-capitalized incumbent owning the data loop. That is the window. It does not stay open forever.

The structural edge that lets a studio act is operator depth, and it is the core of why venture studios win in LATAM. Domain operators with 10+ years of Brazilian-market scar tissue, paired with a Silicon Valley playbook and first-ticket capital assembled on day one, is what a generalist fund cannot replicate from a board seat. AI infrastructure is now cheap enough to deploy without a Series A, which is why a studio can run 3-4 vertical bets a year in Brazil instead of one capital-heavy wager. The pattern repeats. Build an AI copilot to generate proprietary data, then use that data to raise and deploy capital.

When to pass on a big market

The discipline of the test shows up in the passes, not the builds. A studio can pattern-match itself into a crowded vertical where a well-funded incumbent already owns the data loop. The market is large, the workflow is clearly broken, an operator is even available, and three of the four conditions light up green. The fourth is closed. Someone got there first, their proprietary dataset is already compounding, and a new entrant would be feeding a loop it cannot win.

That is a pass, and it should be an easy one. Per the NfX Network Effects Manual, the company whose dataset is central and continuously updated by usage holds an advantage a later entrant cannot simply outspend. Entering a vertical where the data path is already closed spends a scarce slot on a second-place finish.

The size of the market does not rescue the decision. A studio with only 3-4 slots a year cannot afford one that produces a structurally capped outcome. The honest failure mode of studio portfolio construction is confusing a big market for an open one. Disciplined selection means walking away from a large, broken, tempting market precisely because the condition that compounds is the one that is gone.

How Avante runs the selection

Avante Ventures is a venture studio building AI-native companies in Brazil and Latin America, and the four-part test is how it decides where to build. Avante launches 3-4 ventures per year through a six-stage system. Research, Partner, Build, Traction, Revenue, Compound. The Research stage is where the test meets candidate verticals. The Partner stage answers the second condition, because no build starts without a domain operator carrying real scar tissue. Capital deployed is $500K-1.5M per venture across pre-seed, and Avante retains co-founder economics.

The operating discipline backs the selection. Operating partners stay engaged through the first revenue milestone, then move to board-level oversight. Solving company plumbing once routes roughly $300K-500K of effective capital per venture into product and traction rather than overhead, and a studio venture launches 6-9 months ahead of a comparably funded standalone team. That is the payoff for spending the slot well.

The portfolio reads as the test applied three times, by domain. Alphajuri builds in judicial assets, the precatorios and claims market, where the workflow is document-heavy and the data compounds with every case. WIR, with AXA, builds in insurance pricing and risk scoring, where an async API owns a workflow and the loss data is load-bearing. BR Auction Intel builds in real estate auctions, scraping, enriching, and scoring properties where the dataset improves with coverage. Each cleared the same four conditions before it earned a slot. The operating model lives at /principles. The work of a studio is not building. Most of the work is choosing what not to build.