Workers begin to "perform" AI efficiency gains: skill stacking and distorted assessments

**Key points**

**1. AI efficiency assessments have devolved into reporting performances**

After companies shifted from Token usage to return-on-investment assessments, employees need to show "AI content" and efficiency gains in their weekly reports; a Visier survey shows that among 1,000 U.S. full-time employees, 48% admitted to exaggerating their frequency of AI use or their AI expertise.

**2. Rigid metrics spawn large numbers of useless Skills**

Some companies have folded AI usage rates into performance reviews, deducting 20% for falling short of 70%; to complete the task, employees pile up Skills and cover scenarios, but the results are mostly for show and of limited practical help.

**3. AI does not necessarily boost efficiency; instead it adds review work and extra tasks**

**4. What management wants is a company-wide posture of AI adoption**

Some leaders pursue end-to-end SOPs and departmental AI transformation, using AI to cut training costs and reduce reliance on individual experience; employees are still using AI, but they have developed misgivings about the management logic it serves. 2026-10-24 10:42 Office workers begin to "perform" AI efficiency gains Dingjiao One Quick Overview

This article is from the WeChat public account: Dingjiao One, authors: Jin Yufan, Wang Hanxing, Lei Jing, Wang Lu, Chen Dan, editor: Wang Hanxing, cover image from: AI-generated

With just over two months left in 2026, big-tech employees are about to get through a year dominated by AI.

At the start of the year, Jensen Huang's remark that "if an engineer with a $500,000 annual salary consumes less than $250,000 worth of Tokens in a year, it would deeply trouble me" set off the Tokenmaxxing wave. Everyone was using AI and burning Tokens; it seemed that if you used too few Tokens, your work was not up to standard.

By June, big companies found that Tokens were getting a bit too expensive to burn, and began to focus instead on the return on AI use. This did not reduce big-tech employees' AI FOMO (fear of missing out); instead, it made it harder for employees to "please" their bosses through AI use. They had to prove that AI was helping them work more efficiently, that what they delivered had to contain AI, and that every Token they consumed was producing value.

To that end, some wrote "having AI transcribe half an hour of the boss's meeting recording into text" as "building a structured voice information workflow," and it was even rated an excellent AI application case; some deliberately sought out unimportant work scenarios, turned them into Skills, and packaged them under impressive-sounding names to show the boss; and some big-tech PMs (product managers) used AI to do the work of RDs (research and development engineers), then had to ask their RD colleagues to redo it afterward.

This phenomenon is now increasingly common in workplaces at home and abroad. Not long ago, a survey involving the business intelligence company Visier showed that among 1,000 U.S. full-time employees surveyed, 48% admitted to exaggerating their frequency of AI use or their AI expertise to colleagues or superiors.

Li Zhao, born in 1992, Hangzhou, product operations at an e-commerce company

March to May was truly intense. Our group had a usage leaderboard, and admin exported the data and sent it to the group chat every week. Back then, using AI wasn't for getting work done; it was to avoid falling behind on the leaderboard. I asked the girl ranked first how she did it, and she said she had AI write her full-year plan before bed and never even looked at the result. I copied her: as soon as I got to my desk, I had AI run "competitor monitoring," then scrolled short videos and checked entertainment gossip myself, and that counted as getting by.

In June the wind changed: no more looking at usage, only at output. Weekly reports had to include "AI efficiency cases," with data. I started figuring out how to pad the numbers. The most ridiculous time, I wrote "having AI transcribe half an hour of the boss's meeting recording into text" as "building a structured voice information workflow," and this item was later rated an excellent case. Another time, early this month, I was drafting a channel price-adjustment notice—originally a ten-minute job—but to give my weekly report some AI content, I had AI draft it. It made up a lot of data for me, so in the end I dug out last year's Double 11 notice, revised it, and submitted that. But in the weekly report I still wrote "AI-assisted drafting."

Padding the numbers is padding the numbers, but the work still has to get done, and I gradually worked out a method. Two monitors at my desk: AI running on the left, my work on the right. For proposal-type work, I throw the brief in and wait for a framework, without repeatedly tweaking the prompts—whatever it asks, I answer. While it runs the first section, I write the requirements for the second; once the second is running, I go back and check the first section's results, and while I'm at it finish refining the third part, spinning like that for hours. AI really can boost efficiency. Last Thursday, with more than forty customer service complaints, I had it sort them by issue and it produced a table in ten minutes. And meeting minutes: right after a meeting, while everyone was just getting back to their desks, I had already sent the minutes to the group chat.

But AI doesn't run on a fixed schedule—it could be thirty seconds or an hour. I tried doing something else in the meantime, but halfway through an edit it would finish, so I'd switch back to check the result, then switch back again, and often forget where I was. Eventually I just scrolled my phone, watching the progress while scrolling. During that period I clearly felt my focus declining; switching screens back and forth all day left me very tired, yet I couldn't say where the tiredness came from. I had to categorize my work: when doing A, I resolutely don't use AI to do B; small tasks like polishing I save up and handle in a concentrated one- or two-hour block each day.

2. The demo I wrote was unusable, but my boss was very satisfied

Li Mo, 27, Beijing, product manager at a big internet company

I'm a product manager at a big internet company in Beijing, 27 this year, and I've worked at this company for two years.

At the start of the year, the company began requiring everyone to use AI and write Skills. The R&D colleagues were the first to be affected, and then it spread to every role. Actually, I had used AI on and off before—writing copy, organizing requirements—it was at the "give it a try" stage.

Around March or April this year, the leader of our group talked to me, saying that PMs who can't write code will soon be eliminated, and required all of us to learn to use AI to write code, to solve simple requirements ourselves as much as possible, and not to go to R&D for everything.

After that, I did actually write some small demos and did some small UI designs. To be honest, none of that stuff was really usable; when it was handed to the R&D colleagues, they still had to redo it—so no matter how busy I looked writing it, in the end they were the ones who cleaned it up. But the leader was very satisfied. During that period I clearly felt that reporting and delivering became easier than before; whether the thing could actually land was secondary—the posture of "you are using AI" was established first.

After June, our group changed a lot, with significant adjustments on both frontend and backend, and my reporting line also changed. The new leader refined the assessment of AI output and efficiency gains a lot; the old way of casually writing something myself and handing it to R&D colleagues to fix could no longer pass.

So I switched to a different kind of "AI efficiency." I wrote the steps of requirement feedback, market research, and product framework into Skills—actually just some very simple page reading and organizing—letting AI handle each stage, and the deliverables were basically AI direct output. There wasn't much of real value in it; the "performance" element was heavier.

First, a lot of this work doesn't need AI at all; doing it myself wouldn't take much time. Second, many of the requirements AI produced were actually hard for colleagues to push forward once they got them; they often found logic errors, and we had to meet again to go over them. On the surface the process was running, but the actual efficiency wasn't necessarily higher than before.

Still, doing it this way had benefits: the Leader felt we were very busy, the AI workflow was set up, the process data looked good, and the reports looked good.

My own judgment on this is that AI can indeed improve efficiency, especially in the two areas of writing code and writing documents—it really saves time—but it doesn't always improve efficiency; forcing it on actually creates chaos. And personally, I think the biggest problem is that once the Leader thinks your efficiency has improved, they want you to take on other people's work; the dividend of efficiency gains doesn't land on you, and instead becomes the basis for adding more work—that's a losing proposition.

3. Less than 70% AI usage rate means a performance deduction—we competed our way into a pile of Skills nobody uses

Xiaonan, post-95s, Shenzhen, event planning

Later, AI usage rate was even included in performance assessment: if the usage rate didn't reach 70%, performance would be deducted by 20%.

At first, everyone thought it was no big deal—after all, AI really can improve our efficiency. Making an event planning proposal used to take at least 3 days. Now with AI, a first draft can be completed within a day. And for miscellaneous work like competitor analysis and PPT making, AI can help too, really saving a lot of time.

But gradually, things started to feel off. The company launched an internal AI competition, comparing who could use AI to cover more work scenarios, build more workflows, and create more Skills. As long as you could prove you'd used AI to connect some work process, it counted as an achievement.

At first, what people made was genuinely useful—for example, Skills related to event planning could reduce some repetitive, trivial work.

But valuable scenarios are limited after all, and to complete the tasks, everyone started competing on quantity. Some people specifically looked for unimportant work scenarios, made them into Skills, and packaged them with impressive-sounding names; others used AI to make a mini-program and, when reporting, showed how they had solved some problem and saved how much time.

I feel everyone gradually formed a subtle tacit understanding. Compared with whether AI actually helped, being able to show visible results mattered more. After all, how many Skills you made and how many work scenarios you covered are easy to "fake," but how much time was actually saved and how many real problems were solved are not so easy to measure.

One colleague used AI to quickly generate an event proposal, paired it with a polished PPT, and took it to report to the boss. The boss was very satisfied after seeing it, thought the efficiency was especially high, and immediately told the people below to push forward according to the proposal. But when it came to execution, everyone discovered there were quite a few holes in the proposal, and many practical issues hadn't been considered. In the end, it still had to be supplemented and adjusted bit by bit by people.

What frustrated me even more was that the existence of AI also made the number of revisions to proposals increase. Before, once a planning proposal was done, the leader would look it over, give some comments, and after revisions it could move forward. Now the leader, after looking at it themselves, also hands it to AI for review. AI often comes up with a pile of revision suggestions—some genuinely valuable, but some not in line with the actual situation—and we have to check, explain, and revise them one by one.

I find it quite contradictory. The boss wants AI to help the company save time, but we have to spend extra time participating in competitions to prove we're really using AI. After finally using AI to improve efficiency, we have to take on more review, revision, and delivery tasks. Seen this way, using AI has actually become a bad deal.

Wang Min, post-95s, Shanxi, sales at a software development company

Five months after joining this company, I've seen what "AI not adapting to local conditions" means.

The company's main business is software development. It has been around for over a decade and has about 100 employees. The boss is in his sixties, and his understanding of AI comes basically from marketing accounts. Last year he proposed an AI transformation, and at the start of this year it was directly upgraded to an annual strategy. After more than half a year, some business lines have doubled their output after using AI, but more often than not, it's absurd.

So everyone started using every trick in the book to "scrape together AI." For example, the day before yesterday I built a skill, yesterday I did a round of data cleaning, and today I did a full-lifecycle plan for a client from 0 to 1, with the AI share ratcheting up from 30% all the way to 80%. Whether the AI actually did any work doesn't matter—as long as it "participated."

Writing weekly reports used to be for the boss to read; now it has become a loop process. I send my thoughts to the AI, the AI writes the weekly report, I check it once and send it to the boss, the boss uses AI to raise a few questions and sends them to me, I send them to the AI, and so on and so forth. The thing I use AI for most every week is having AI make up what I did with AI this week.

He thinks employees typing away at their computers making plans all day is slacking off, and always says you can't produce results by typing—you need to make more calls and communicate more. So an AI customer management system went live: when a call is dialed it records automatically, the AI listens in, summarizes, gives suggestions, and generates a daily report when the call ends.

Now employees on the phone at the office look like they're "on camera." Say one wrong thing and the backend logs it; if your tone is off, the AI gives suggestions; if the client conversation doesn't go well, the daily report is automatically flagged in red. Whenever he has a moment, the boss scrolls through the app replaying communication records, and if he hears a sentence he doesn't like, he screenshots it and throws it straight into the big group chat. Every time we see a group message pop up, our backs go cold.

One day the boss came out of a meeting and said he wanted to build an internal AI bot, with employees at the core of the AI transformation, to solve document sharing and improve efficiency. An HR colleague followed an online tutorial, uploaded all the attendance, performance, and payroll tables, and had it built in a day. Then someone in the company group chat asked about salaries, and the AI just dumped out the details.

That set off an explosion. Some people asked and found out that people in the same role were paid differently, some found their performance scores had been marked down, and others went through the entire company's payroll, turning up three names who were still being paid but had never once shown up for work. To everyone's surprise, the boss didn't publicly criticize HR at the next day's summary meeting—instead he praised her for her high execution efficiency. But after the meeting, the boss contacted the AI company's technical staff to add permission restrictions, and the system went back online.

AI is clearly a good thing, yet the more we use it, the more off it feels. We originally hoped it would save effort and boost efficiency, but instead we have to put on a show for it. In the end, what's exposed isn't a problem with AI. The payroll leak was the company lacking basic common sense about data permissions. Employees' performative work is the boss treating AI as a new surveillance tool. AI amplifies management itself: if processes are smooth, it amplifies efficiency; if there's a lot of suspicion, it amplifies internal friction.

Once I understood this, I actually felt relieved. Even without assessments, I still can't do without AI. Making PPTs, writing planning documents, organizing files—these miscellaneous chores really do save me a lot of effort. But anything that requires professional judgment, I absolutely will not hand over to it.

Lin Zhou, post-00s, Guangzhou, software R&D engineer

A few days ago, a colleague I didn't know suddenly reached out to me, asking how to use a plugin I had published on the internal plugin marketplace. I was actually pretty happy. After all, this was something I'd spent time researching, and someone actively using it meant it really did solve a practical problem.

But my +1 (direct supervisor) didn't seem satisfied with just that.

I'm a bit confused. The tools I figured out myself really are useful, but everyone's work content and habits are different—why insist on using the same set of processes? What's even stranger is that the veteran employees in the group who've worked there five-plus years don't seem anxious. One veteran told me privately that he actually doesn't care.

Later, I figured it out. The core competitiveness of veteran employees lies in business experience and professional knowledge. To them, AI is just an auxiliary tool—use it however is convenient, since it isn't tied to performance anyway. What's really tied to AI is the +2 who called for the department's AI transformation, the +1 who actively responds, and the unlucky new employees who get assigned the specific tasks—like me.

This also explains why +1 is so fixated on building SOPs.

Recently a few outsourced colleagues joined the department. I have a vague feeling that once this version's development is finished, they'll probably leave. When the next project starts, new people will come in again. With staff turnover this frequent, the ideal management approach is naturally to standardize workflows so new people can get up to speed without much training. People can leave; the skills stay.

I think the direction of AI development is personalization and customization. In the past, people adapted to tools; now tools should adapt to people, and everyone can build their own workflow according to their needs. But +1 seems to think AI can turn all business into SOPs. To put it bluntly, I think half of his idea is wishful thinking and half is performance. How much efficiency is actually improved may not be the most important thing—what matters is letting superiors see that the whole department is undergoing AI transformation.

It's just that I originally thought I was helping colleagues reduce repetitive work, and only later realized I might also be helping the company reduce its dependence on employees. What's more subtle is that this doesn't conflict with my personal interests—it may even benefit me. I can learn new technologies, accumulate project experience, and gain real sense of achievement.

I still enjoy doing AI efficiency work; it's just that after suddenly seeing through management's possible motives, I feel a layer of unease in my heart. Now I finally understand why some people say that if your moral standards are too high, you're not suited for this company. The most uncomfortable thing isn't being forced to do something bad—it's clearly doing something you like, only to find it also serves a logic you don't agree with.

(At the request of the interviewees, Li Zhao, Li Mo, Xiao Nan, Wang Min, and Lin Zhou are pseudonyms in this article.) Channel: Society & Culture

570 articles published on Huxiu. This content is authorized by the author for publication; the views represent the author only and do not represent Huxiu's position. If you have objections or complaints about this article, please contact tougao@huxiu.com. Everyone who is changing the world and wants to change the world is on the Huxiu APP[]