The Vertical AI Wedge Most Founders Are Missing

Most founders building vertical AI are racing in the wrong direction.

They’re trying to be faster. Faster chat. Faster workflows. Faster dashboards. The pitch decks all say the same thing — AI lets us move at superhuman speed. The thesis is that whoever ships the fastest AI-native version of an existing SaaS category wins.

That’s not the wedge.

The founder who’s most articulate about this isn’t a frontier AI lab researcher. He’s the CEO of a vertical SaaS company in the most overlooked corner of B2B software: municipal parks and recreation. Birju Kadakia spent 50,000 hours in city halls before scaling Rec Technologies into 100+ cities. When asked what survives AI disruption in vertical SaaS, his answer was direct:

“Some software is dead. You have to be a platform.”

The interesting question is what he means by “platform” — and which three pillars decide which vertical SaaS companies make it through the AI wave.

Why “vertical SaaS is dead” is the wrong question

The popular framing right now is binary: AI eats SaaS, or it doesn’t. ClickUp’s homepage says “software to replace all software.” The implied argument: when LLMs can vibe-code a working CRM in an afternoon, why pay $50 a seat per month for one?

The framing is wrong because it lumps two different categories together. Kadakia draws the line cleanly: feature-companies die; platform-companies survive. The dying category includes things like basic website builders, simple form generators, and any tool whose value is “one specific output that an LLM can now produce on demand.” That category was already commoditizing before AI; AI just accelerated the curve.

The surviving category is different. It owns three things horizontal AI can’t replicate: the data, the payment rails, and the infrastructure that makes the category fair and accessible. Those three pillars are the wedge.

Pillar 1: Own the valuable data

Every vertical has data that’s hard to get and high-trust to handle. In recreation, that data is schedule (when is the tennis court available?), customer preference (do you qualify for resident-rate pricing?), and compliance (allergies, emergency contacts, age requirements for kids’ summer camps).

A horizontal AI model can answer general questions about parks and recreation, but it can’t answer “is court 3 at Dolores Park available at 6pm on Tuesday for an Emeryville resident with a coaching subscription?” That question requires real-time access to scheduling APIs, identity verification, and pricing logic that lives in one specific city’s system. Owning the data is owning the questions the AI can actually answer.

For builders evaluating their own startup: what data does your category produce that’s hard to acquire, hard to standardize, and high-trust to handle? If the answer is “none,” you’re a feature, not a platform.

Pillar 2: Own the payment rails

This is the pillar most founders underestimate.

Payments aren’t just charge-and-collect. In a category like recreation, payments include resident-rate pricing (San Francisco residents pay one rate, non-residents pay another), income-bracket eligibility (lower-income families qualify for sliding-scale fees), store-credit logic (your refund from last summer becomes credit toward this summer’s camp), and complex tax/jurisdiction handling.

Kadakia frames it: “Transactions and money movements are really high-risk, high-important things. You saw OpenAI kind of back out of commerce. They’re just really high-risk, high-important things. You need a platform that’s designed around that.”

Horizontal AI players have repeatedly pulled back from commerce because the trust requirements are different. Your vertical AI moat lives partly in the willingness to do the unsexy compliance work to handle real money in your domain. If you don’t, the customer takes a bigger trust risk every time they integrate you, and the buyer-side procurement process collapses.

Pillar 3: Own the infrastructure for equity and access

The third pillar is the one that’s hardest to copy because it’s about workflow knowledge, not technology. Recreation runs on equity rules — limited supply of summer camp slots, financial aid programs that get early access, ADA accommodations, age-based eligibility. The system has to be fair, transparent, and auditable.

Building this requires sitting with the operators long enough to learn the unwritten rules. Kadakia’s team has been in the field 50,000 hours. That’s not a brag; it’s a moat. A horizontal AI model can read documentation about how summer camp registration works, but it can’t replicate the knowledge that comes from watching 100 different cities handle 100 different edge cases over three years.

The same logic applies in legal (Harvey), medical billing, insurance, and any vertical where the workflow is shaped by regulation, equity rules, or compliance. Field hours compound into knowledge that’s hard to compress into a model weight.

The pickleball test

If you’re building vertical AI right now and want a sanity check on whether you’re playing the platform game or the feature game, here’s the test from Kadakia’s playbook:

When Rec started in 2022, the founders’ vision was a full operating system for the entire $1.9T US activity economy. The obvious move was to build the platform on day one.

Instead, they picked pickleball-court booking software for 10 cities. Specifically. Pickleball. Booking. Just that.

The reason: pickleball was blowing up, cities needed better software for those specific courts, and shipping it gave Rec real schedule data and real customer data. With that data, they earned the right to expand into facility reservation, then memberships, then learning, then payments, then agents.

The discipline is rare. Most founders building a platform-sized vision try to ship the platform on day one, burn 18 months on infrastructure no one is using, and run out of runway. Kadakia held the platform vision in his head AND shipped a pickleball-booking tool the next month. That combination — vision plus wedge discipline — is the actual ship.

If you can’t articulate the pickleball-equivalent for your own vertical AI startup — the specific, embarrassingly narrow wedge you could ship next month that produces data the bigger platform needs — you’re probably not building a platform. You’re building a feature pretending to be a platform.

What this means if you’re building

Three takeaways for anyone in the vertical AI build phase:

Audit your moat against the three pillars. Be honest about what your startup actually owns. If you don’t own the data, the payments, and the access infrastructure for your vertical, ask why. If the answer is “we haven’t built it yet,” put it on the roadmap. If the answer is “the platforms will commoditize that,” you’re betting against yourself.

Sequence ruthlessly. Pick the smallest possible wedge that produces data your eventual platform needs. Don’t try to be useful to everyone in the vertical on day one. Be useful to one narrow user group, completely, and earn the right to expand.

Put in the field time. The 50,000-hours number is a thesis, not a brag. Vertical AI companies that win in 2027 will be the ones whose founders sat with operators in 2024. If you’re a vertical AI founder and you haven’t spent at least 1,000 hours in your customer’s actual workflow yet, get out of your office.

The wedge isn’t speed. It’s the patient, unglamorous work of building the data layer, the payments layer, and the infrastructure layer that horizontal AI will struggle to replicate because the value lives in domain understanding, not in inference quality.

The companies that figure this out won’t beat ChatGPT. They’ll be the operating system ChatGPT integrates with.

---

Frequently Asked Questions

What is the vertical AI wedge? The vertical AI wedge is the defensible position a vertical AI startup builds by owning three things horizontal AI models can’t replicate: the data specific to its domain, the payment rails for the industry, and the infrastructure for equity and access. Rec Technologies CEO Birju Kadakia argues these three pillars decide which vertical SaaS companies survive AI disruption.

Is vertical SaaS dead in the AI era? Some vertical SaaS is dead — particularly feature-companies whose value is one specific output an LLM can produce on demand. But platform-companies survive. Survivors own the data, payment rails, and infrastructure that horizontal AI can’t replicate. The dividing line is whether the company has built a platform-sized moat or just shipped a single feature.

What are the three pillars of vertical AI defensibility? The three pillars Birju Kadakia identifies are: (1) Own the valuable data — the schedule, customer preference, and compliance data unique to your vertical; (2) Own the payment rails — including complex pricing logic, eligibility rules, and credit handling; (3) Own the infrastructure for equity and access — the workflow knowledge that takes thousands of field hours to build.

Why won’t horizontal AI like ChatGPT eat vertical SaaS? Horizontal AI struggles to replicate three things: real-time access to vertical-specific data, the trust requirements of moving real money in regulated industries, and the workflow knowledge that comes from spending thousands of hours in a specific industry. Birju Kadakia notes OpenAI has repeatedly pulled back from commerce because the trust requirements are different.

How long does it take to build a vertical AI moat? Rec Technologies’ team has spent 50,000 hours in the field over three to four years. Birju Kadakia argues that field time is the actual moat, not the AI model quality. Horizontal AI models will keep improving, but the domain knowledge that lets vertical AI companies build trustworthy workflows takes years of operator embedding to acquire.

What is the pickleball wedge strategy? The pickleball wedge is Birju Kadakia’s term for picking the smallest possible specific problem to solve before tackling the platform. Rec started in 2022 with pickleball-court booking software for 10 cities — the most boring possible wedge. Shipping it gave Rec real schedule and customer data, which earned the right to expand into facility reservations, memberships, learning, and AI agents.

How can founders apply the platform thesis to their own startup? Three steps: (1) Audit your moat against the three pillars — be honest about what your startup actually owns; (2) Sequence ruthlessly — pick the smallest wedge that produces data your eventual platform needs; (3) Put in the field time — get out of your office and spend hundreds of hours in your customer’s actual workflow before scaling.

What kind of AI startups are most vulnerable to horizontal AI commoditization? Feature-companies — startups whose entire value is one specific output an LLM can now produce on demand. Examples include basic website builders, simple form generators, and AI tools that wrap a single ChatGPT prompt in a UI. These were already commoditizing before AI; horizontal model improvements just accelerated the curve.

What is the difference between a feature-company and a platform-company in AI? A feature-company solves one isolated problem with one specific output. A platform-company connects different entities, owns the valuable data, handles the payments, and provides infrastructure that makes the category function. Platform-companies have multi-layered moats; feature-companies have one-layer moats that AI typically erodes.

---

This insight post is based on a conversation between Angelina Yang and Birju Kadakia, CEO and Co-Founder of Rec Technologies, on Heroes Behind AI.