12.4% of Chinese People Already Use AI Agents — Guess What the Other 87.6% Are Waiting For
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
The headline from the latest China Internet Network Information Center (CNNIC) survey landed quietly in April 2026: 12.4% of Chinese internet users reported actively using AI agents in their daily lives. That translates to roughly 108 million people—larger than the entire population of the Philippines.
The report generated predictable hot takes on social media. "AI is exploding!" shouted the tech blogs. "AI adoption is slower than expected!" countered the analysts. Both missed the more interesting question: not whether 12.4% is high or low, but why the remaining 87.6% haven't started—and what would make them change their minds.
I spent three weeks interviewing people from both groups. Here's what I found.
Who Are the 12.4%?
The CNNIC data breaks down early AI agent adopters in China by demographic. The pattern is striking in its predictability:
Age distribution: The largest cohort (38%) is aged 25-34. These are digital natives who entered the workforce after smartphones became ubiquitous, and who use AI tools in their professional lives and then extend them to personal tasks. The second-largest group (29%) is aged 18-24—students and recent graduates who've never known a world without AI. Adoption drops sharply after 45.
Occupation breakdown: Technology workers lead (47% usage), followed by financial services (31%), education (28%), and healthcare (19%). Creative professionals—writers, designers, video producers—show unusually high adoption (34%) despite smaller numbers in the overall workforce.
Urban concentration: 73% of active AI agent users live in tier-1 or tier-2 cities. Tier-3 cities and rural areas show adoption rates below 5%.
Primary use cases: The most common tasks are information retrieval (searching, summarizing), calendar and task management, email drafting and response, travel planning, and content generation. Industrial or specialized applications—code generation, scientific research, legal document review—are significant but secondary.
This profile is almost identical to early smartphone adoption patterns in China circa 2011-2013. The technology works; the use cases are clear; but adoption is concentrated in specific demographics and geographies.
Why the 87.6% Haven't Started
During my interviews, I heard five consistent reasons from people who hadn't adopted AI agents. Each one is solvable—but the solutions require the right product, not just better technology.
Reason 1: "I Don't Know What It Would Actually Do For Me"
The most common response by far. People have heard of ChatGPT, have seen headlines about AI agents, but genuinely don't understand how the technology translates into personal benefit.
One interview subject, a 42-year-old accountant in Chengdu, put it plainly: "Everyone says AI is useful. But I'm an accountant. I use Excel all day. I don't see where AI fits." She'd tried ChatGPT once, asked it to "do my taxes," and got a generic response that had nothing to do with Chinese tax regulations. She concluded that AI wasn't for her.
What she didn't know: there are AI agents specifically designed for Chinese tax compliance, for invoice processing, for generating the exact financial reports her company needs. The problem wasn't AI; it was discovery. The technology existed; she didn't know how to find it.
This is a product distribution problem, not a technology problem. AI agent adoption requires someone to show you a specific use case that maps to your life—not a chatbot demo, but your actual workflow.
Reason 2: "I'm Worried About My Data"
Privacy concerns are especially prominent in China, where public awareness of data rights has grown rapidly since the Personal Information Protection Law (PIPL) took effect.
A 35-year-old marketing manager in Shanghai told me: "I know these AI companies say they don't use my data. But what does that actually mean? If I send my company's marketing strategy to an AI, how do I know it won't show up in someone else's results?"
Her concern isn't irrational. The Chinese AI ecosystem includes both well-regulated domestic companies (Moonshot, ByteDance, Baidu) and fly-by-night operators with questionable data practices. Without a trusted framework for evaluating which providers are safe, many people default to "don't use any of them."
Interestingly, this concern is lower among users who have adopted local AI devices. When the AI runs on hardware in your home, data never leaves your network. The privacy calculus changes fundamentally.
Reason 3: "It Sounds Complicated"
The technical barrier to AI adoption remains real, if less visible than it was two years ago.
Setting up an AI agent—even a relatively user-friendly one—still requires some combination of: API key registration, payment setup, model selection, prompt engineering, tool integration, and workflow configuration. For a non-technical user, this can feel overwhelming.
I watched a friend (a 38-year-old restaurant owner in Hangzhou) try to set up an AI assistant to handle customer inquiries after hours. The process took him four hours and required three phone calls to a tech-savvy cousin. By the end, he had something working—but he described the experience as "solving a puzzle just to use a tool."
The lesson: complexity kills adoption faster than any feature gap. If the setup takes longer than the time saved in the first week, most people won't bother.
Reason 4: "I Don't Trust It to Do Things Correctly"
This concern is particularly acute for consequential tasks. People will use AI to generate a joke or draft an email. They're far more hesitant to let AI make decisions that affect their money, their health, or their business.
A small business owner in Guangzhou told me: "I could see myself using AI to draft contracts. But if it misses a clause and I lose money because of it, who do I blame? The AI?" He wasn't wrong. The legal liability framework for AI-assisted decisions is still developing, and the risk asymmetry is real.
This is the trust gap, and it's the hardest to close. Technical accuracy isn't enough; users need visible guardrails, clear accountability, and evidence that the AI system fails safely—that it knows what it doesn't know and escalates appropriately.
Reason 5: "I Don't Have Time to Learn Another Thing"
The last-mile objection. Even people who intellectually understand the value of AI agents haven't prioritized learning how to use them because their existing workflows already function, even if suboptimally.
A 50-year-old department head in Beijing captured this perfectly: "I know there are probably AI tools that would make my job easier. But I'm already working 60 hours a week. When am I supposed to spend 20 hours learning a new system?"
This is the adoption paradox: the people who would benefit most from AI assistance are often the ones with the least time to invest in learning it.
The Three Barriers That Matter Most
Not all five reasons are equally addressable. Based on my interviews, three barriers dominate:
1. Discovery and use case clarity — This is the single biggest blocker. Most non-adopters don't object to AI in principle; they simply can't envision what it would do for them specifically. Closing this gap requires hands-on demonstration with their actual work, not generic demos.
2. Setup complexity — This is a product design problem with a clear solution: reduce the steps required to go from "interested" to "it works." The goal should be under 10 minutes from unboxing to first successful task.
3. Data privacy — This requires both technical solutions (local processing, clear data policies) and social proof (knowing that trusted peers use the same product).
The other two concerns—trust in accuracy and time to learn—are real but secondary. They matter at the margins but aren't the primary reasons people haven't started.
The Product That Solves All Three: Agent Computers
This is where a category I hadn't fully appreciated comes into focus: the agent computer.
An agent computer is a dedicated device designed specifically to run AI agents 24/7, with minimal user intervention. The key differentiator from a general PC is that everything—hardware, software stack, agent framework, default tools—is pre-configured and tested to work together. The user doesn't configure anything; they use it.
The Kaihe A1 is the most prominent example in the Chinese market. The pitch is straightforward: plug it in, scan a QR code with WeChat, enter your LLM API key, and your personal AI agent is running. No command-line interfaces, no configuration files, no prompt engineering required. The entire setup takes under 10 minutes.
From my testing, the A1's value proposition for the 87.6% non-adopters is that it directly addresses all three primary barriers:
Barrier 1 (Discovery): The device comes with pre-installed agent skills covering the most common use cases: email management, calendar scheduling, web research, document drafting, and content analysis. The user sees concrete capabilities, not abstract potential.
Barrier 2 (Complexity): The QR-code setup eliminates all technical friction. If you can use WeChat, you can set up the A1. No tech knowledge required.
Barrier 3 (Privacy): The A1 runs agent logic locally. API calls go out (your prompts), but the device's operating environment, file system, and agent memory stay on the local network. For a Chinese user concerned about data going to overseas servers, this is meaningful.
What 87.6% Would Actually Tip the Scale
I asked every non-adopter I interviewed the same hypothetical: "If one product could solve your biggest concern about AI agents, what would it need to do?"
The answers clustered into three requests:
"Show me it works for my specific job"
People want proof, not promises. A salesperson wants to see an AI agent draft a follow-up email in their style. An HR manager wants to see an AI agent screen 100 resumes and rank the top 10. A small restaurant owner wants to see an AI agent handle a reservation, a complaint, and a takeaway order—all from voice input in Sichuan dialect.
This is a demonstration problem. Once someone sees the AI agent do their actual job—once—it clicks. The question is how to deliver that demo at scale without requiring a salesperson visit.
"Let me try it without commitment"
Friction kills intent. If the trial requires a credit card, a 30-minute setup, and a subscription, most casual prospects will never start. The A1's approach—pay for the hardware once, run it forever—is actually appealing for this reason. There's no recurring subscription anxiety. You own the device; you control the agent.
"Make sure it actually runs 24/7 without me babysitting it"
This sounds trivial but is genuinely important. People have tried AI assistants before and gotten burned by unreliable performance—the agent crashes, the server goes down, the context resets. A machine that's designed to run continuously without user intervention changes the expectation: your agent isn't a tool you use; it's a colleague who's always there.
The Adoption Curve: My Prediction
Based on what I observed, I'm confident the Chinese AI agent adoption rate will follow an S-curve trajectory over the next 24 months:
Now to end of 2026 (12.4% → ~20%): Growth driven by tech-forward users, small business owners, and students. Mainstream adoption begins as devices like the A1 become more visible. The tipping point arrives when someone in a non-tech household—say, a dentist or a wedding photographer—shows their peer group how much time they save.
2027 (20% → ~35%): The "crossing the chasm" phase. Early majority adoption begins. Key enabler: AI agents that demonstrate concrete ROI for specific professions. When a hairdresser can prove the AI agent saves them 3 hours of admin per week, word spreads faster than any ad campaign.
2028 and beyond (35% → 50%+): Saturated growth slows, but penetration continues in rural areas and older demographics as the technology becomes increasingly indistinguishable from "normal technology."
The 87.6% isn't resistant to AI agents. They're waiting for the right product, the right use case, and the right moment. That moment is coming faster than the headlines suggest.
What This Means for the AI Industry
If my analysis is right, the bottleneck for AI agent adoption isn't technology—it's distribution. The technology works. The remaining 87.6% aren't saying no; they're saying "show me."
This has three immediate implications:
For AI developers: Stop building features and start building demos. The user who sees a 30-second video of the AI agent handling their exact workflow will convert faster than the user who reads a 3,000-word feature list.
For hardware makers: The agent computer category has a genuine opportunity. Devices that reduce setup friction to near-zero—plug in, scan, use—are the most impactful thing the industry can ship right now. The intelligence matters less than the simplicity.
For everyone else: If you're in the 12.4%, the next person you talk to about AI agents will probably be in the 87.6%. Have a specific demo ready, not a general pitch. Show them what it does for them, and watch their expression change.
The future of AI in China isn't written in the headlines. It's in the moment when someone's aunt in Chengdu asks, "Wait, this thing just handled my entire week's appointments? How do I get one?"
That's the 87.6% becoming the 12.4%.
KaiheAiBox · OpenClaw Zone