Today at the office I did one thing: installed AI agent tools for my coworkers and taught them how to use natural language to get AI to simplify their daily work.
What I had prepared was content like "how to write good prompts" and "what agents can do." After a full day, not a single thing that actually blocked everyone had anything to do with model capability.
Gate one: half the room went down at the installer download
Step one was installing the software on everyone's computers. Every office machine used Baidu or 360 Search. I searched for the tool's name and the top results were all fake download sites sitting in paid ad slots: identical icons, wrong domains, and a bundled installer waiting inside.
The more embarrassing part was me. I'm so used to Google that on a computer where it doesn't work, I genuinely couldn't think of a normal way to search for a moment. It took a beat before I remembered Bing, typed the URL in by hand, and finally reached the official site.
If someone who lives in this industry all day stumbles here, an ordinary user never stood a chance.
The impact of this on me was bigger than I expected. It's 2026, people are already talking about "the year of the agent," yet the very first step for an ordinary Chinese office worker who wants to use AI — finding the real installer — still means crossing a paid-ranking minefield that has existed since 2010.
AI capability is sprinting ahead, but the road to AI is still jammed a decade and a half back.
Gate two: the word "AI" doesn't mean the same thing to us
Once installation was done and we started chatting, I realized that for most of my coworkers, their entire understanding of AI was Doubao.
That's not a put-down. Doubao is the only AI entry point on their phones: ask questions, write a bit of something — good enough, and free. So when I demoed an agent operating the computer by itself, installing drivers, batch-processing files, everyone's first reaction wasn't amazement — it was wariness:
"This thing costs money, right?"
I asked about everyone's price ceiling. The answer was unanimous: anything over 200 yuan a year and they wouldn't pay; free only.
I personally spend several times that figure every month on a handful of AI subscriptions. In that instant I realized: my coworkers and I were using the same word, "AI," but referring to two entirely different species.
For me, it's productivity infrastructure. For them, it's a free, take-it-or-leave-it chat toy.
Neither side is right or wrong. Free Doubao has defined the public's price anchor for AI, and any paid product must first answer: "Where exactly are you better than free — better enough to justify the price?" For an ordinary salaried office worker, until that math works out, willingness to pay is zero.
The ceiling for consumer AI subscriptions in lower-tier markets may be far lower than every business plan claims.
Gate three: fear — and the professionalism behind the fear
When the demo reached the part where the agent installs drivers for a new device, one of the office coworkers asked a question:
"Is it going to delete all my work files?"
I was about to offer a couple of reassuring words, but on reflection, this worry was anything but naive. This is exactly the permission-control problem of agents — a security challenge the entire industry hasn't fully solved. She had never read a single paper, but her intuition hit the bullseye.
My answer was simple: high-risk operations stop and ask you to confirm first; just remember to glance at what it's about to do before you click confirm. That's enough.
When it came to teaching, in the end I used only one sentence:
"Just treat it like making a wish — state the wish clearly."
That sentence worked better than every feature introduction I'd given before.
Then the coworker responsible for paperwork asked another question:
"If everyone uses AI to write materials, won't the output be highly repetitive? I wouldn't dare use it directly."
Another professional question. The homogenization of AI-generated content is a real pain point in every serious use case right now.
My suggestion was to demote AI from "generator" to "processor": formatting, item-by-item checking, and filling in gaps can be delegated to it; the core content should be drafted from your own real accumulation, with AI only polishing and organizing the structure.
Your differentiation always comes from what you have and others don't. AI can't replace that part, and shouldn't.
Closing thoughts
After a full day, my conclusion is this: the last mile of AI adoption has little to do with model capability.
What's actually blocking the road is three things:
- Search engines that surface fake download sites
- Price anchors locked in by free products
- Legitimate fear of losing control
Not one of these can be solved by shipping a stronger model.
There was also an unexpected reward: the best questions of the day came precisely from the people most "afraid" of AI.
Fear of deleted files maps to permission control; fear of repetitive content maps to output homogenization. They don't understand the technology, but their fear gets closer to this industry's real hard problems than the hype of many tech bloggers.
Next time someone tells me "AI adoption is just a matter of time," I'll invite them to an ordinary Chinese office and have them search for an installer on Baidu.