A DingTalk notification, 9:14 on a Monday
At 9:14 on a Monday morning, the DingTalk group pinged. HR had dropped a notice into a 50-person admin channel. Next Wednesday, 2 to 5 p.m., the company had hired an outside instructor for a session called “AI Office Efficiency Training.” Two people per department. Names due Friday.
DingTalk is Alibaba’s workplace messaging app — think Slack, but with a stronger grip on the Chinese office. Yu Hang, an administrative specialist at a mid-sized e-commerce company in Hangzhou, read the notice twice. He’s 28 and has been in the job four years. His first reaction wasn’t excitement. Wednesday afternoon was already blocked for quarterly supplier reconciliation. He typed back, “Is admin sending me?” Forty minutes later, HR replied privately: put your name down, we’ll sort out the calendar.
He forwarded the notice to an eight-person chat. One colleague asked whether it covered ChatGPT. One asked whether there was a test. A third sent three characters — 又要学, “study again” — and nobody replied.
That mix of curiosity, calendar math and mild fatigue is a fair snapshot of how AI training lands inside a lot of Chinese companies right now. It isn’t a launch event. It’s a calendar item.
Why companies are suddenly paying for this
The manager’s logic, as Yu Hang understood it, was plain: competitors were drafting product descriptions and answering customer messages faster.
His company’s customer-service team had been told to shorten response times. Finance was buried in expense reports. “The boss kept saying 降本增效 — cut cost, raise efficiency,” Yu Hang said. “AI looked like the cheapest lever.”
Two years ago, sessions like this were optional, closer to a perk. Now some companies fold them into HR’s own performance targets. Training providers sell package deals, and local governments in several provinces subsidize part of the cost. The important detail is what companies are actually buying. Not a deep dive into how language models work — an office-tools class.
Inside a three-hour session
The first forty minutes are registration. People create accounts, log in, download a browser plug-in, and hit a wall or two with two-factor verification. The instructor walks the room through it step by step, because half the attendees are on company laptops with restrictions.
Then the demonstrations begin, in roughly this order. Meeting minutes: a 25-minute recorded call gets pasted into a tool, which returns a summary and a list of action items. Weekly reports: bullet points in, a first draft out. Spreadsheets: sort this, flag the anomaly. Slides: type a topic, get an outline, adjust the layout. Email: rewrite this for an unhappy customer, but keep it polite.
The instructor deliberately shows a failure. He asks the tool to chart last quarter’s sales and summarize the trend. The summary gets the direction wrong. He uses the mistake to make his point: check the output, every time.
By the end, the room has split roughly into three groups — those already using one of these tools on their phones, those typing notes by hand, and those checking their work chat under the table.
They teach prompt templates, not neural networks
What gets covered: prompt structure (role, task, context, format), office scenarios, data-security rules, copyright and fact-checking basics. What almost never comes up: how a transformer works, why models hallucinate at the architecture level, what fine-tuning costs. Anyone hoping to understand the machinery leaves disappointed.
The data-security section is the one people actually remember. Never paste customer names, ID numbers, contract terms or internal financials into a public AI tool. At Yu Hang’s company, that is a strict red line. Some firms go further and block the public websites outright, running a licensed or locally hosted version inside their own network.
Fact-checking is the recurring warning. Instructors show examples of AI inventing a policy number or misquoting a statistic. The rule they give is blunt: treat every AI output as a draft from an intern who is confident but sometimes wrong.
Mostly, though, what’s taught is a set of templates — “summarize this thread politely,” “compare these two vendor quotes,” “turn these notes into a one-page brief.” Users save them, reuse them, share them in work chats. That folder of prompts is the real product of the training.
Different desks, different homework
Not everyone learns the same thing. The course is usually tailored by role. Admin and HR get résumé screening, interview-note organizing, policy Q&A drafts. Sales gets customer-message drafting, competitor summaries, proposal outlines. Finance gets spreadsheet analysis and invoice classification, with a standing warning not to upload raw numbers. Operations gets SOP writing and data cleanup. Programmers get coding assistants — a separate class entirely.
Age matters less than habit. In one session Yu Hang attended, the fastest person at the keyboard was a 46-year-old finance manager who had been using AI to draft Excel formulas for months. The slowest was a 24-year-old who had only used it to write cover letters and assumed it was a toy.
After the training: what actually changes
Some companies run a test at the end. Some hand out a certificate. A few organize internal competitions — best AI-generated proposal, most time saved in a week. The energy in those contests fades fast.
Back at the desk, the change is uneven. Yu Hang now uses AI to summarize supplier email threads and draft meeting minutes. On a heavy day, he estimates it saves him about 40 minutes. His finance colleague uses it to explain formulas, not to do the accounting. Sales drafts messages with it, then rewrites most of them.
The tasks that stuck all have a human review step built in — summarization, first drafts, formatting, translation. The tasks that didn’t are the ones where being wrong is expensive, or where the data legally can’t leave the building. That gap between demo and desk is the most common complaint you hear from employees. The class makes it look seamless. The actual work has friction.
Who pays, and who’s nervous
Companies cover most of the bill. Some local governments subsidize digital-skills training, and platforms hand out free seats to pull firms onto paid tiers. A lot of Chinese office workers have now had their first taste of these tools without paying anything themselves.
The nervousness is real but rarely voiced in the room. In the small chat afterward, a colleague asked Yu Hang a version of the question everyone carries: if I get good at this, am I training my replacement? His answer was practical. The tool can draft the summary, but someone still has to know which supplier is being difficult, which number looks off, and which manager needs a different tone. That’s still his job. Whether it stays his job in three years is a different question — and he knows it.
Back at the desk
On the Thursday after the training, Yu Hang opened the browser tab with the AI tool, pasted in a 14-message supplier thread, and asked for a summary in three bullet points. The summary was mostly right. He fixed two lines by hand, then sent it up.
The training didn’t make his job disappear. It added a step: check the machine’s work before it leaves your desk. That step takes five minutes. His boss assumes it saves half an hour, and nobody in the room is in a hurry to correct him.
The bigger change is quieter. Yu Hang now writes his own emails differently — shorter, more structured — because he’s been reading machine-generated text all week. No instructor taught him that. The training just handed him a new mirror for how he writes.
Start the discussion at forum.chinacomes.com