heading · body

YouTube

I analyzed 373 AI startups selected by Y Combinator in 2026

Harshit Tyagi published 2026-06-10 added 2026-06-16 score 6/10
ai startups y-combinator agents b2b saas
watch on youtube → view transcript

ELI5/TLDR

Someone read the descriptions of 373 companies Y Combinator funded in 2026 and looked for the pattern. The answer: 90% are AI companies, so “AI” no longer makes you special. What makes a company special now is owning a boring, repetitive task that someone does every single day — and not just talking about it, but actually doing it. The winning products don’t suggest; they take action, log what they did, and let a human stay in charge at the points that matter.

The Full Story

”AI” stopped being the interesting part

The headline number does the heavy lifting. Of 373 funded companies, 336 are AI companies — about 90%. When nearly everyone in the room is building with the same tool, the tool stops being a selling point.

AI is no longer the differentiator. It’s the default. The real differentiator is the workflow that the company owns.

Two other numbers fill in the picture. 92.5% are B2B — companies selling to other companies, not consumer toys. And the median team is two people, which is what you’d expect from very early startups.

The biggest single category, 122 companies, is developer tools and AI infrastructure — the picks and shovels. If everyone is building agents, everyone needs the surrounding plumbing: testing tools, data pipelines, ways to watch what the agents are doing. A surprising second place goes to industrial, manufacturing and robotics (54 companies), meaning AI is leaking out of software and into factories and logistics.

From answering to doing

The core idea of the video is a shift the presenter calls moving “from generation to action.” A chatbot hands you a block of text and you do the work. The newer products do the work themselves — they read the messy incoming data, decide what to do, go into your other software, complete the task, and update everything.

His example is Senta, which finds jobs for people, tailors the application, and submits it. The person keeps control; the dull repetition gets handled. The economic point underneath: buyers pay more for finished work than for suggestions.

Once software acts, it can break things

Here’s the catch the video is honest about. A wrong answer is annoying. A wrong action — an email sent, money moved — costs real money and trust. So a whole new category is appearing whose entire job is keeping agents on a leash: permissions, approvals, audit logs, sandboxes (a walled-off space where an agent can act without touching the real system), and rollbacks.

You need these guardrails to be provided to the agent before they can start operating in real tools.

The named examples are a small zoo. Clawweiser lets an agent use your Gmail or Slack without ever seeing the actual passwords, sitting in the middle to check each action is allowed. Stranger still: Mount sells insurance for deployed AI agents, and Heaven offers banking accounts for “teams and agents.” Whether the world needs an insurance carrier for chatbots is left as an exercise.

Trust as the product, not a feature

In law, compliance, finance and healthcare, the presenter argues, people don’t buy AI because it sounds clever. They buy it because it can show its work — what changed, why, what evidence was used, who signed off. RegBase, for instance, tracks regulatory changes across countries and government websites and has to be reliable enough that a compliance team will actually act on it.

Coding agents and the “looks like SaaS, is actually an OS” trick

Coding agents (Codex, Claude Code, and others) get their own section: they’ve gone from writing snippets to taking a ticket, branching the code, writing it, testing it, fixing the automated checks, and producing a change a human can review. Replicas-there lets teams hand tasks to these agents from Slack or GitHub, each running in its own isolated machine so many can work at once.

The cleverest observation is the last: some startups look like ordinary vertical software — “AI for logistics,” “AI for manufacturing” — but are quietly becoming the operating system for that industry. The recipe is to pick one painful workflow, then creep outward to where the data lives, where decisions get made, and where approvals happen. Day Job, building scheduling agents for transport that plug into existing systems, is the example. The related dream is the one-to-three-person company that runs large operations on specialized agents instead of a twenty-person team. The path there, via a company like General Legal (an AI-native law firm at a flat fee): start as a service, then turn the repeated parts into a product, keeping a human at the “trust boundary.”

The closing advice is the negative space: don’t build generic copilots, thin chatbots, shallow wrappers, or novelty content tools. Own a real, painful workflow or don’t bother.

Key Takeaways

  • 90% of the 373 funded companies are AI; 92.5% are B2B; median team size is two. AI is now the baseline assumption, not the pitch.
  • The dividing line is “AI feature” vs “AI operating system” — collect context, decide, act inside other tools, update the record, and keep a log.
  • Biggest funded segment is developer tooling and infrastructure (122 companies) — the plumbing everyone building agents will need.
  • Second-biggest is industrial / manufacturing / robotics (54) — a sign AI is moving into physical operations, not just software.
  • A new “guardrail” layer is emerging because acting agents can cause real damage: permissions, approvals, audit logs, sandboxes, rollbacks, even agent insurance and agent banking.
  • In regulated fields (legal, finance, healthcare), traceability is the product — what changed, why, what evidence, who approved.
  • The repeatable playbook for founders: pick one painful workflow you understand deeply, start as a service, then productize the repeated steps while keeping a human at the trust boundary.
  • What to avoid: generic copilots, thin chatbots, shallow wrappers, novelty content tools, generic outbound automation.

Claude’s Take

The central insight is genuinely useful and well-stated: the era where “we use AI” was a differentiator is over, and what’s left is owning a workflow end-to-end. The “feature vs operating system” framing and the “looks like vertical SaaS, is actually an OS” observation are the kind of thing worth keeping.

But treat the rigor with caution. The presenter says so himself — the categories and “opportunity signals” are inferred from company descriptions, not audited. That means the whole analysis is a careful read of marketing copy, which describes what 373 companies want to sound like, not necessarily what they do or whether they’ll work. The company name-drops (Senta, Clawweiser, Mount, Day Job, General Legal) are illustrations chosen to fit the thesis, not evidence for it — and a few, like agent insurance carriers, read more like froth than signal.

It’s also, underneath the analysis, a soft advertisement: the founder runs an AI consultancy, and the video ends pointing to a paid report and a chat-with-the-data tool. None of that makes the observations wrong, but it shapes which ones get emphasized. As a map of what YC funded and the vocabulary of the moment, it’s a solid 6. As a guide to what will actually survive, the survivorship bias is baked in — these are companies that got funded, which is a different question from companies that will last.