After watching pitches from 199 companies at YC W26 Demo Day, one intense sense of discomfort remained. AI is no longer just a "new technology," but has become simply an infrastructure.

60% of participating companies are AI-native, and an additional 26% are AI-enabled. In other words, only 14% of companies are not using AI. But the important thing isn't this number. What has changed is that everyone is thinking not about "how to use AI," but "what to replace with AI."

Expressions like co-pilot, assistant, or copilot have become obsolete. What startups are aiming for now is complete substitution of high-paying jobs. Beacon Health replaces pre-authorization staff, Mendral handles engineering tasks, and LegalOS achieves a 100% approval rate for visa applications. These are not auxiliary tools but genuine replacements.

Here, a new term is needed: "AI agent." It has shifted from a technical term to refer to entire professions. Recruiters, legal work, medical administration, DevOps, QA testing—all are targets for replacement.

It's no surprise that 87% of the market is B2B. Only 14 companies target consumers, and of those, just 7 are officially classified as "B2C." Why? Because AI agents are best suited for structured workflows. Consumer-facing applications are more ambiguous and still challenging for AI.

What’s most interesting is what the fastest-growing companies are doing. Law firms, recruitment, accounting, insurance brokers—these are existing professional service industries. Arcline has over 50 startups as clients, functioning as an AI-native service company. Panta explicitly states it is a "service business based on the software economy." AI handles 80% of human work, operates on a performance-based model, and enjoys software profit margins while maintaining service reliability.

The lesson from this model is clear: start with services, gather data, eventually release automation, and evolve into a platform. Services are the wedge, and data becomes the moat.

The distribution channel story is also fascinating. Of the top 15 rapidly growing companies, 60% initially acquired customers through founder networks or YC networks. In other words, these companies already knew "who to reach" rather than "how to sell."

The most typical pattern is selling to former colleagues. The founders of Fed10 are ex-lobbyists, and their business card became their sales channel. The founders of Squid worked at State Grid for years, aware of inefficiencies in power grid planning. They didn’t need to find customers—they were their own customers.

Founder backgrounds are also distinctive. 46% are two-person teams, with the most common being two technical co-founders with different expertise (35%). Not hacker + salesperson, but two technologists. And they are often college classmates, former colleagues, or partners from past startups.

The most successful founders share a common trait: they deeply understand the problems they aim to solve through personal experience. A dentist developing surgical AI at Mango Medical, an aircraft maintenance supervisor automating document creation at Zymbly, rancher sons building robotic cowboys at GrazeMate. Companies that penetrate industries that aren’t the talk of cocktail parties but are deep and steady tend to be the strongest.

The resurgence of hardware is also notable. 18% of the batches include hardware components, a significant increase from recent years. Remy AI and Servo7 ship warehouse robots, Pocket has shipped over 30k wearable devices. Companies founded by alumni of SpaceX and Tesla stand out most in this group.

The importance of data advantage was also emphasized. LegalOS trained on 12,000 visa application cases to achieve a 100% approval rate. Shofo is building the world’s largest indexed video library. Since they use the same basic model, proprietary data becomes a key defense.

Failure patterns are also clear. 8–10 companies building agent monitoring or testing features are in risky territory—because the foundational model providers are building these natively. AI-native services without a data advantage are the same: quick monetization but very low defensibility. Core technology can be replicated in weeks.

An interesting rephrasing was that market entry strategies often fail because companies "build first and then hope." Successful companies first ask "who can we reach, and what do they urgently need?" Failing companies ask "we built a great product, how do we sell it?"—and that difference determines everything.

The gaps in consumer, education, and government tech are also notable. Few companies participate in these fields. Historically, the areas with the least funding have generated the greatest returns later. The next big wave of AI will likely emerge in these overlooked sectors.

Five common traits among the most rapidly growing companies: selling results, not tools; founders building customer relationships before product launch; charging from day one; customers being in urgent situations; MVPs being unusually simple.

The quality of pitches varied greatly. Memorable pitches have seven elements: shocking data or problem statement; specific, uncommon issues; a bombshell statement summarizing the team’s achievement; explaining market inevitability; compelling traction exceeding expectations; unique insights; and a crazy closing line—like "The first AI Oscar will be born at Martini" or "Book your lunar hotel for 2032." How these words influence investors is everything.

GRU Space plans to build the first hotel on the moon by 2032, securing a $500 million letter of intent and an invitation to the White House. Terranox has discovered uranium deposits worth $200–700 million. Ditto Bio is decoding immune control proteins evolved by parasites with AI to design proprietary immunotherapies. These companies are not just tech firms—they are tackling fundamental human problems.

The biggest lesson from Demo Day is that the era of AI-native startups has truly begun. But the real winners are not companies that just use AI well, but those that deeply embed expertise and use AI to transform existing industries. Deeply penetrating traditional sectors, building data flywheels, and avoiding generic AI wrappers—these are the most important lessons emerging from 199 pitches.