Notes from the latest YC batch's AI startups
The current YC batch is more AI-saturated than ever. The pattern of what's working — and what's getting funded but not working — has shifted again.
If you read the latest YC batch list looking for diversity of category, you will not find it. Roughly two-thirds of the companies are AI-shaped. Of those, maybe a fifth would have existed in their current form a year ago. The category is moving faster than the partnership can re-cluster it on the website.
The interesting question is no longer "are there a lot of AI startups in this batch." It is "what is the pattern of what is working." Watching demo day chatter, founder interviews, and the YC blog over the last batch cycle, a few things have shifted.
The vertical AI thesis is consolidating
The "AI for X" startups, where X is a specific vertical with regulatory drag, are getting more capital and more attention than the horizontal AI tools. AI for medical billing. AI for property management. AI for K-12 curriculum compliance. The pitches are duller. The metrics are better.
There is a defensible thesis behind this. Horizontal AI tools are competing on a feature surface that the model labs are slowly absorbing. Vertical AI tools are competing on workflow integration and trust, which the model labs are explicitly not interested in. That asymmetry favors the boring-sounding company.
The "AI-native CRM" wave finally broke
Two batches ago, there was an obvious cluster of "AI-native CRM" companies. By this batch, almost none of them are pitching themselves that way anymore. The ones that are still alive have repositioned. The ones that haven't are quiet.
The lesson is not that the category is dead. It is that "AI-native" was never a category. It was an adjective. The CRMs that found product-market fit found it because they served a specific persona — recruiters, GP-style fundraisers, small agencies — better than the incumbents. The ones that pitched "AI-native" as the value proposition struggled to find customers who cared about the adjective.
Voice is back, quietly
A small but real cluster in the current batch is voice-first. Not voice assistants in the 2018 sense. Voice as a primary interface for a specific job: customer support, outbound sales, scheduling, certain kinds of compliance reads.
These companies were unfashionable for a long time because voice AI was bad. It is now, in narrow domains, good. The companies winning here are not building general-purpose voice. They are building a narrow voice loop that does one job well and integrates into an existing CRM or PBX.
This is the kind of category investors miss until it is too late, because the demo is not exciting and the metrics are. It is also the kind of category where a small team can build a real business in twelve months.
Multi-agent is being walked back
Eighteen months ago, "multi-agent" was a pitch deck staple. In the current batch it is, conspicuously, less common. The teams who are doing meaningful multi-agent work have learned to talk about it as "orchestration" or "workflows," because investors have been burned by demos of agents talking to each other and accomplishing nothing.
The companies actually shipping multi-agent products in 2026 mostly have a humble explanation of what they mean by it. They have one to three role-specialized agents in a fixed graph, not a free-form swarm. That is much less interesting to read about and much more interesting to use.
The pattern under the pattern
If there is a single thread through the AI startups in this batch that are getting traction, it is that they have stopped trying to build "an AI company" and started trying to build a software company that happens to use AI heavily. The founders talk about their customers more than their models. The decks lead with the workflow, not the architecture.
That is, of course, exactly what good software companies always did. The AI hype cycle made it temporarily acceptable to pitch otherwise. The current batch is the one where the pendulum visibly swung back.