The AI Startup Shift: Forget LLMs, Vertical Agents Are the Real Opportunity

Here’s something interesting.

I’ve been talking to several friends doing AI startups lately, and they all have something in common: none of them are building “LLMs” anymore. They’ve all pivoted to “vertical agents.”

It’s not that they’ve stopped doing AI—they’ve just stopped dreaming the “I’m going to build China’s OpenAI” dream. Now they’re all building: legal agents, medical agents, financial agents, education agents… each going deep into a specific vertical.

Honestly, I think this is a good thing.

Why? Because the LLM war is over. Now it’s the agent war.

Think about it. In 2024-2025, how many companies in China were building LLMs? Dozens. And now? Only 5-7 really matter: DeepSeek, Qwen, Ernie Bot, GLM, Hunyuan, Kimi, and Doubao.

The rest were either acquired, pivoted, or died.

This is normal. LLMs are a classic “winner-takes-all” market—if your tech is slightly worse than competitors, users won’t choose you. If users don’t choose you, you get less data. Less data means your model gets worse. It’s a vicious cycle.

So the smarter startups are now thinking: I won’t build LLMs. I’ll use them.

But here’s the question: what do you build with LLMs?

In 2025, many people were doing “wrapper apps”—calling GPT or Claude APIs to build chat apps or knowledge bases. That model is basically dead now because why would users pay for your wrapper when they can use ChatGPT directly?

So the 2026 startup direction is clear: vertical agents.

What’s a vertical agent?

I’m not talking about generic “legal consultation bots” or “medical diagnosis bots.” A real vertical agent is one that optimizes all three core Agent capabilities—tool calling, multi-step reasoning, and context management—for a specific niche scenario.

For example: there’s a startup building a “contract review agent” specifically for commercial contract review.

This agent does things like:

1. Automatically identifying risky clauses (based on legal knowledge base)
2. Comparing clause differences across similar contracts (multi-document reasoning)
3. Generating modification suggestions (based on historical cases)
4. Tracking modification progress (task management + tool calling)

These four steps each require deep customization. You can’t just call GPT and be done with it.

And the value of this agent is: it can genuinely replace part of a lawyer’s work, not just “assist.” Users are willing to pay for that.

Let me give you a counterexample: those generic “AI legal advisors” are basically all struggling. Why? Because when users actually have legal problems, they still need to find a lawyer. What can your AI do? Just give some reference opinions.

That’s the difference between “vertical” and “generic.”

So here’s the question: what kind of vertical agents have a chance to survive?

My personal judgment: scenario + data + closed loop.

• Scenario: The scenario you serve must be high-frequency and essential. Not something people use once a year.
• Data: You have exclusive data for this scenario. Data that others don’t have or can’t easily access.
• Closed loop: Your agent can actually solve the problem, not just “assist” but “replace.”

All three are essential.

For example: medical agents. The scenario is essential, but data is hard to get (hospitals don’t open their data), and achieving a closed loop is difficult (medical liability issues). So most medical agents are still in the “assist” stage—none can truly replace doctors.

But legal agents are different. Many legal documents are public, so you can train models; plus, contract review has relatively clear liability (lawyers still sign off in the end), so agents can achieve “quasi-replacement.”

Another direction: education agents. Essential scenario, easy to get data (questions, knowledge points, learning paths are all public), and relatively simple closed loop (auto-grade homework, recommend exercises). That’s why there are many companies building education agents now.

But the problem with education agents is: competition is fierce. Everyone can think of this direction—how do you ensure what you build is better than others?

This brings us to another key point: hardware ecosystem.

Look, companies with hardware ecosystems—Huawei, Xiaomi, ByteDance (Doubao)—have natural advantages when building agents.

Why? Because agents need to interact with users’ real environments. Your agent wants to order food delivery for users? You need access to delivery platforms. Want to book flights? You need access to booking systems. Want to control smart home devices? You need access to the devices.

Without a hardware ecosystem, you have to negotiate partnerships one by one. Can you pull that off? It’s hard.

So startups building agents now either attach themselves to a big platform (WeChat, ByteDance, Huawei) or build pure software tools (contract review, code generation). Those in the middle basically can’t survive.

Honestly, I think in the next two years, the agent market will go through a major shakeout. The only ones that will survive are companies that achieve “true replacement” in a specific niche.

As for those startups still “building LLMs,” honestly, I don’t see much opportunity anymore.

Unless—you have enough money to burn.