---
title: "Work Agent in the Real World: What Happens When You Hand Work to AI — and Still End Up Busy?"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-09-11T15:01:14+00:00"
canonical: "https://ffcap.cn/en/research/work-agent-in-the-real-world-what"
source: "https://uniqueresearch.substack.com/p/work-agent-in-the-real-world-what"
language: "en"
---

# Work Agent in the Real World: What Happens When You Hand Work to AI — and Still End Up Busy?

[![](https://substackcdn.com/image/fetch/$s_!NF3l!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab807b46-f10a-49fb-8671-0a5a88d2bd3d_2048x1152.png)](https://substackcdn.com/image/fetch/$s_!NF3l!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab807b46-f10a-49fb-8671-0a5a88d2bd3d_2048x1152.png)

Agents can already search for information, build spreadsheets, and write reports. But how much of what they produce is actually usable? How much revision do humans still need to do? And is the monthly cost worth it?

Unique Research asked 500+ users these concrete questions. Now we’re publishing all the survey data.

* * *

What ignited the debate over “can AI actually do work?” was a red lobster — OpenClaw. It started as a weekend project in November 2025. By early March 2026, its GitHub stars had crossed 250,000. On August 26, GitHub’s official figure was close to 388,000.

It let many people see for the first time that AI could read files, call tools, run commands, and continue completing tasks in the background.

But this fire burned fast and faded fast. Unique Research data shows that from April to August 2026, OpenClaw’s official website monthly visits dropped from 14.2 million to 2.77 million — an 80% decline. Estimated active users fell from 6.31 million to 1.41 million — a 78% decline. August was still down 18% and 17% respectively compared to July.

If you only watch this lobster, the story looks a lot like a hype cycle fading.

* * *

But in the same dataset, another force is amplifying.

The domestic site WorkBuddy reached 14.09 million visits in August, up 135% month-over-month, rising to #8 on the domestic visit ranking. Estimated active users reached 3.76 million, up 123% month-over-month, rising to #9 on the domestic active ranking.

The newly launched Qwen Office reached 1.09 million visits and an estimated 410,000 active users, with month-over-month growth of 552% and 298% respectively.

Among the 17 Work Agent and Skill platforms in the August specialty ranking, 4 grew month-over-month, 12 declined, and 1 was new to the ranking.

(Web traffic only covers browser access — it doesn’t see desktop clients, mobile apps, or usage embedded in office software. It also can’t be used to conclude that users migrated from one product to another. Source: Unique Research.)

* * *

Beyond the hot and cold of individual products, what’s more obvious is the expansion of the entire category.

At the beginning of the year, people were still arguing about “can AI actually do work?” By mid-year, desktops were already crowded with Work Agents claiming to finish your work for you.

OpenAI launched the Codex desktop app in February, released ChatGPT Work in July, and then put Chat, Work, and Codex into the same new desktop client.

Domestically, WorkBuddy launched in March and by August had entered Unique Research’s domestic AI Web top 10. Qwen Office integrated QoderWork, MuleRun, and Wukong into a new entry point in August. Doubao Work connected with Feishu and supports cloud-based continued execution. Kimi Work put up to 300 agents running in parallel into a local workspace. Baidu Dazi, AutoClaw, Nano Work, and WPS Lingxi also entered from their respective strong entry points.

Suddenly there are many products, but feature lists can’t answer the more critical question: have users actually started using them? What work do they hand over to agents? Can the results go directly into deliverables? Are workflows being reused? And are they willing to pay?

With these questions, Unique Research launched this Work Agent survey, asking users to talk about their own usage: what tasks they’ve already handed to AI, where they still have to make changes themselves, and how much they’ve spent.

Among the survey participants, 88.4% had used a Work Agent in the last 30 days. The following answers about tasks, results, and payment mainly come from these users.

* * *

Among respondents who had used a Work Agent in the last 30 days, 79.8% said they had used it every week in the past four weeks. That means, for this group, agents have entered weekly work.

Of course, “used every week” and “used every day” are quite different. Someone might only open it when writing a report, while someone else might use it daily — this question didn’t further subdivide.

What’s more worth looking at next: what are they repeatedly opening the agent to do?

* * *

71.9% of users had agents search, research, or summarize information. 68.4% had them write or revise documents and reports. 50.9% used them for spreadsheets, data, or charts. 49.1% for PPT. 43.9% for writing code, debugging, or building applications.

These tasks aren’t unfamiliar. For example, to write an industry report, you first find materials, put data from different sources into a spreadsheet, and then organize it into an article. An agent can take over a chain of these steps, and the human then checks and revises.

For people who haven’t used one, you can start with a document or spreadsheet at hand. For product people, you can also make the entry point more specific: when a user comes with “help me organize these materials,” can they get something usable as quickly as possible?

* * *

76.3% of users had agents read or modify local files. 65.8% had them operate browsers or web pages.

So-called “can actually do things” becomes very concrete here: open the file you give it, find information on a webpage, write the results back into a document.

72.8% had used Skills, plugins, or tool calls. The name sounds a bit technical, but it can be understood as: giving the agent a method for doing things, or letting it call appropriate software tools.

Two other capabilities were used by close to half of users. One was remembering preferences and previous materials so you don’t have to repeat yourself next time — 48.2%. The other was letting tasks continue running in the cloud or execute at a scheduled time — 49.1%.

Cross-application execution or multi-person collaboration was at 23.7%.

* * *

48.2% of users said that after revision, most of the agent’s results were usable. 28.1% said most results could be used or delivered directly. 15.8% only used a small portion of them. 7.9% mainly used them for ideas.

“Help me produce a first draft” is already a real experience for a considerable number of people. But from first draft to delivery, there’s often still your own checkpoint to pass.

Take a report as an example: the article is written, but you still have to check whether the data was copied correctly, whether the arguments support the conclusions, and whether the tone suits the reader. Counting this step gets closer to the real work process.

* * *

After the agent finishes, how long does the human stay busy? The survey didn’t record minutes, but it asked what level of follow-up work people usually have to take on.

43.0% said they only need to handle a few key issues. 23.7% mainly do final checks. 7.0% basically need no extra handling.

On the other side, 21.9% need frequent corrections or additions. 4.4% have to redo most of it themselves.

“Check it once and it’s ready” and “revise back and forth for half a day” both get lumped together as “used AI.” These experiences are very different.

If a product wants to know whether users are actually saving time, it has to keep asking: where exactly do people get stuck? Is it incomplete materials, misunderstood requirements, or results that are hard to verify?

* * *

Suppose you write a weekly report every week. The first time, you tell the agent how to organize materials, what format to use, and what to focus on. Next week, do you have to say all that again? That’s what “reuse” is about.

37.7% of users have already saved templates, Skills, or automated tasks for repeated use. 17.5% have formed fixed repeat workflows — that is, they have a method they follow each time. 33.3% occasionally repeat similar tasks. 11.4% basically start from scratch every time.

For products, one place to make things smoother is: after a task is done, help users save the format, requirements, and approach from this time. Next time with different materials, they can pick up where they left off.

How much time this actually saves depends on the specific task.

* * *

43.9% of users were bothered by fast quota consumption or high prices. 36.0% felt execution was slow or took too long to wait. 22.8% found verifying and checking results tedious.

This calculation has to be done completely. Buying a subscription is one cost; additional quotas for long tasks might be another. Waiting for it to finish and then spending time checking is also a cost.

So a good agent needs to make a few small things clear: roughly how much quota will this take? What step is it on? Which numbers have sources, and which places still need human confirmation? When users can understand these, they can decide whether to keep waiting or take over themselves.

* * *

At the beginning of the survey, no product hints were given — people were just asked to write down the first name that came to mind.

ChatGPT/Codex was mentioned by 49.0% of respondents. WorkBuddy by 46.9%. Claude-related products by 21.9%.

The list after that is also interesting: there are office tools, programming tools, and products that can run automated tasks. In respondents’ first reactions, “Agent” has already been packed into different kinds of tools.

One person can write several names. The percentages here mean “how many people thought of it,” not “how many people are using it.”

* * *

Asked which product they used most often in the last 30 days: ChatGPT desktop (Codex) at 39.5%, WorkBuddy at 34.2%, Claude Cowork at 6.1%.

This table is more like each respondent’s “frequently used toolbar.” One person might have tried several, but here they’re selecting the one they use most often.

To understand why a product stays, you have to keep asking what users actually do with it every day, and which tasks would be awkward with a different tool.

* * *

31.7% first learned about or started using Work Agents through tech media, official accounts, or content creators. 30.9% because they were already using related models or other products. 23.6% from recommendations by friends, colleagues, or classmates.

For people who haven’t used one yet, “it has dozens of capabilities” may not be as easy to understand as a complete operation process: what materials were given, how it did it, and what was finally obtained. When doing product introductions, it’s worth explaining these details thoroughly. Also show readers which places need human revision, so they know what to expect when they try it themselves.

* * *

42.1% of users felt they could switch to another one the same day. 24.6% needed a few days to adapt and migrate. 17.5% needed to significantly adjust existing workflows or materials. 13.2% found it hard to find a replacement product that could complete the same tasks. 2.6% weren’t sure yet.

Switching tools — some people find it easy, some find it troublesome. This is how people answered “if it shut down.”

To understand where the trouble is, you have to ask further: what exactly is stored inside that has to be redone when moving?

* * *

This follow-up question was only asked of those who felt switching would take days of adaptation, major adjustments, or that it was hard to find a replacement. Among this group:

-   65.1% felt task history, memory, or personal context was hardest to move
    
-   38.1% selected saved workflows, templates, Skills, or scheduled tasks
    
-   34.9% selected usage habits and trust
    

The difficulty, put simply, is the materials you’ve given before, how far the discussion got, and the personal preferences it already knows. Switching software isn’t the hassle — having to re-teach it everything you’ve already taught it might be.

For example, for the same report, a new tool might need to relearn your desired format, commonly used materials, and expression habits.

For product teams, what’s worth refining isn’t just “remember more.” Whether users can see what it remembered, remove outdated content, and take important materials with them also affects daily use.

* * *

For their most commonly used product:

-   16.7% pay 100–199 RMB directly
    
-   11.4% pay 200–499 RMB
    
-   9.6% pay 1,500 RMB or more
    

Quite a few people don’t pay this separately:

-   15.8% have it covered by someone else, their company, school, or team
    
-   8.8% use an existing model subscription or API (for example, if you’ve already bought a model service and then connect it to a tool, you might not need to pay the tool separately)
    
-   14.0% have never paid directly
    
-   5.3% paid before but don’t now
    

These situations need to be looked at separately. Someone whose company pays and someone who tried it and didn’t renew both didn’t pay out of pocket this month, but the reasons behind it could be completely different.

* * *

This time, instead of looking at current bills, the question was: based on the help you’re getting now, if you paid yourself going forward, what’s the maximum you’d be willing to spend each month?

-   20.2% chose 100–199 RMB
    
-   14.0% chose 1,500 RMB or more
    
-   8.8% are temporarily unwilling to pay
    
-   7.9% haven’t thought it through
    

Some people are willing to spend over a thousand, others don’t want to spend a cent. To understand this question, you have to keep asking what they use the agent for. How often a task comes up, how much trouble it is to do yourself, and how long it takes to fix mistakes all affect the “is it worth it” judgment.

Willingness to pay and whether someone will actually buy are still two different things. This set of answers is better used to追问 usage scenarios; pricing still needs actual purchases to verify.

* * *

If you haven’t used an agent yet, you don’t need to learn all the terminology first. Take a document at hand and give it a task that’s clear and checkable — that’s enough to start.

After it produces a first draft, look at what you still have to revise. Next time you do the same thing, see whether the format, requirements, and commonly used materials can be saved to reduce repeated preparation.

For product people, these few calculations from users are worth following: how much quota a task uses, how long you wait, and how much time is spent on final checks and rework. New features can attract people to try, but these small daily things determine whether it feels smooth to use.

When a person encounters the same work next week, will they naturally open the agent, or feel that doing it themselves is easier? The answer is hidden in the work left for them after the last task was done.

* * *

Unique Research conducted an online questionnaire on Work Agent usage in personal work, targeting people with different levels of AI experience — including recent users, as well as those who have used before, have only heard of, or don’t yet know about these products. The research focuses on product awareness, specific tasks, result adoption, human handoff, repeated use, usage barriers, and personal payment — not enterprise procurement or deployment surveys.

This article includes valid responses submitted between August 12 and September 9, 2026. Respondents voluntarily completed the questionnaire online, making it a non-probability, self-selected sample. The questionnaire first collected unprompted product recall through open questions, then explained the Work Agent concept, and showed applicable questions based on usage experience.

Each chart notes the question type and statistical population. Single-choice questions show the composition of each option. Multiple-choice questions allow one person to select multiple options, so the sum of percentages can exceed 100%. Open-question product mention rates use the people who filled in that question as the base. Percentages are rounded to one decimal place, which may cause single-choice totals to deviate slightly from 100%.

Respondents are mainly knowledge workers and people with some AI experience. Their occupations and AI experience are shown below. All usage, delivery, and amounts are self-reported by respondents and have not been verified against product logs, deliverable quality testing, or payment records. The report does not represent all users in China, nor is it used to estimate market share or prove causal relationships.

Product, operations, marketing, or sales: 31.0%. Research, consulting, or investment: 19.4%. Software development, data, or other technical work: 17.1%.

This occupational structure helps understand the knowledge-work characteristics of these responses.

49.6% self-assess that they can design workflows, agents, automation, or use APIs. 27.1% can proficiently handle files and complex tasks.

The following are respondents’ self-descriptions of their experience, not ability levels verified through exams or product logs. API can be understood as an interface that lets programs call models or tools.

* * *

**Full-text fidelity gate passed: 0 substantive omissions**

**Delta public-web standing authorization granted 2026-09-05**

---

Original publication: https://uniqueresearch.substack.com/p/work-agent-in-the-real-world-what
On-site reading page: https://ffcap.cn/en/research/work-agent-in-the-real-world-what