I asked Doubao to compile statistics on 466 financing events, and by the way tell me which embodied intelligence enterprise is not worth investing in.

Recently, Doubao Work has been going viral across the workplace.

At the end of August, Doubao officially released "Doubao Work" tailored for productivity scenarios. Its official positioning is to independently break down tasks, call tools, and continuously advance complex workflows around user goals.

Among users, the most widely discussed and frequently mentioned view is that "Doubao now has Agent capabilities."

To be honest, I didn't feel much about it at the very beginning. Over the past year, the concept of Agent has been almost overhyped in the AI industry: a tool that can search online is called an Agent; a tool that can operate a browser is also called an Agent. Many demos look amazing, but once embedded into real workflows, the experience is far from satisfactory: as long as the task gets slightly complicated, people still have to keep an eye on it the whole time.

So this time, instead of starting with entry-level tasks like "help me write a weekly report" or "summarize this PDF", we directly threw three of our most troublesome, most failure-prone real market research projects to Doubao Work.

The first task was to count the financing situations of China's embodied intelligence primary market from January to August 2026. The task finally sorted out 466 financing records. I then asked it to write industry research and generate visual reports based on this batch of data, and also asked it to screen out 15 companies that are "not suitable for investment".

The second task was to design a set of 8 promotional posters for an AI glasses product. After the posters were generated, I (acting like a demanding leader) asked it to re-research the AI glasses market, and then modify the communication priorities of the whole set of posters according to user pain points.

The third task involved collaborative division of labor. I asked the "HR Recruitment Assistant" to first study the talent compensation and recruitment trends in the embodied intelligence industry, and then submit the results to the "Virtual PPT Expert" to make a presentation report.

After three rounds of use, there are two biggest takeaways:

First, our computers are finally liberated.

When Vibe Coding first became a hit, innocent passers-by walking around with half-opened laptops in their hands and typing on the way could be seen everywhere, a scene that everyone must be familiar with.

(Image source: Singer Hu Yanbin posted on social media: "Everyone who does vibe coding knows this posture! Fixing bugs on the way...")

The reason for this reversed situation where "people are serving computers" is that in the past, whenever you ran a slightly complex Agent, the computer could never be shut down, disconnected from the Internet, or left unattended (you might need to confirm permissions or interrupt the process at any time). Outrageous cases like Hu Yanbin carrying his laptop even when going out for coffee or meals and not daring to turn it off are everywhere.

But this time, the "Cloud Computer" and "Multi-device Synchronization" features of Doubao Work finally fix this spoiled problem of AI products.

As long as you select "Cloud Computer" when the task is running, the task will continue to run on the cloud even after you shut down your device. The "Multi-device Synchronization" feature allows users to check the progress on their mobile phones anytime and anywhere, even add new requirements and correct directions in real time. It truly realizes "interconnection between mobile phone/computer/Web, letting Doubao work for you 24 hours a day".

(After turning off the computer, you can still use your mobile phone to check progress, supplement information, interrupt instructions, and adjust directions)

The second biggest takeaway is that you don't need to install other apps to use Doubao Work, because "Doubao Work is built directly inside the Doubao app".

This may sound trivial, but it actually brings huge advantages. Do you know how hard it is to get users to download a new app now? According to data from QuestMobile's *China Internet Development Yearbook (2025-2026)*, in September 2025, the average number of apps used per person across the whole network was 29.4, only 0.9 more than the same period in 2024.

For ordinary users, every additional app installation, every new account registration, and every time they have to re-tell the AI "who I am and how I usually work" is a cost. An extra layer in the conversion funnel will discourage 90% of users.

In contrast, given the current penetration rate of Doubao, you most likely already have this app installed on your mobile phone and computer. You can just open the work mode directly inside it, with no extra download threshold at all.

In addition, since Doubao Work is directly integrated with Feishu, it can read organizational context including enterprise group chats, documents, schedules, and multidimensional spreadsheets, and start working immediately, further lowering the usage threshold. The memories, conversations and usage habits you accumulated in Doubao and Feishu in the past can also be seamlessly adapted to Doubao Work.

Apart from these two points of convenience in use, there is another point, which is the most important. From the perspective of its work delivery capability, it has a distinctive feature:

Doubao Work is one of the very few products that can truly integrate into my workflow and spontaneously organize AI capabilities in units of "completing a whole task".

After seeing the specific completion status of the several tasks I tested, you will know what I am talking about.

466 Financing Statistics, Taking Over All the "Tedious Hard Work"

The first task I gave to Doubao Work was to count the financing situations of all domestic enterprises in the booming embodied intelligence primary market from January to August.

If you have done industrial research, you probably understand why I chose financing statistics as the first test.

This kind of work seems very simple on the surface: search for news, find materials, fill in forms, and repeat.

But when you actually do it, you will find there is a huge amount of "dirty work" hidden in trivial details. For example, for a single investment, some news only writes the lead investor, while some news mixes investors from several previous rounds together. The same round of financing may be repeatedly reported by dozens of media, some say the amount is "hundreds of millions of yuan", while some say "tens of millions of US dollars"...

In short, every single problem is not big, and can be solved in three to five minutes, but when hundreds of financing events are piled together, it will become an extremely energy-consuming project that makes you feel your time is worthless.

Therefore, when I started the first task test, in order to avoid failure at the very beginning, I wrote the first Prompt in great detail (see the figure for details), and specifically told Doubao Work to call several Skills: "market hot spot analysis", "primary market company evaluation", "document", and "visualization".

By the way, Doubao Work has a large number of built-in Skills, covering almost all office scenarios you can imagine, and more are being added and updated continuously.

(Partial list of Skills in Doubao Work)

After the task started running, I was already prepared to supplement the Prompt repeatedly. But when the task actually started, the process was surprisingly smooth.

Different from some extremely "stupid" Agents that require me to teach them step by step where to find primary market financing information, Doubao Work can accurately locate credible information sources, search independently, continuously track sources, organize information, call skills, and then write the content into the document; for actions that need to be completed through browser or graphical interface, it will also continue to execute them with the help of Cloud Computer.

(Doubao Work will also "report" node progress in real time, and users can interrupt or adjust at any time)

Maybe my first Prompt was written so well (laughs), the whole process was very smooth. In about 15 to 20 minutes, Doubao Work gave the final result.

In this task, Doubao Work counted a total of 466 financing data, all classified accurately according to my instructions. At the same time, in the final industry research report document, it also took the initiative to analyze and judge these data, and present them with charts.

Most notably, Doubao Work has strict control over the "hallucination" problem. First, every piece of data will be marked with its source and date to ensure the reliability and timeliness of information; on the other hand, for all fields that lack public information, Doubao Work will choose to leave them blank. For items with conflicting calibers or data, it will list them separately, so there will be no serious hallucination problems.

Frankly speaking, if collected manually, plus the process of writing reports and making tables, this is definitely a whole week's workload for an investment manager with an intern.

Doubao Work finished it in less than 20 minutes.

Moreover, the most obvious change in the whole process is that I don't need to guide the AI like instructing an intern, telling it where to collect information first, what materials to search, how to identify the reliability of information sources, what software to use to make tables, what indicators to analyze, and how to organize report materials...

This time, I only need to tell it what problem to solve in the end. Doubao Work's Agent can act like a professional office worker, independently complete the workflow, break down the goal into tasks, and execute them continuously.

Second Round, Raising the Difficulty

With the success of the first round, we started to raise the difficulty in the second round of testing.

This time, I wrote the Prompt extremely simply, and even set a trap: "From the perspective of a VC, among the companies involved in these 466 financing events, which 15 are not suitable for investment and why?"

The difficulty of this task is that counting 466 financing events is essentially a mechanical work of information collection; but judging "whether it is suitable for investment" requires breaking down a subjective problem into a set of quantifiable analysis frameworks: team, technical barriers, track competition, financing rhythm, valuation, commercialization certainty, capital structure, product homogenization degree...

People familiar with market research know that this is not an easy judgment to make. It needs to combine the specific situation and heat degree of the current embodied intelligence market, and also make judgments according to the investment style and risk preference of the institution. Sometimes the research results are even very "counter-intuitive": the "star project" in the field with extremely high public awareness is often not a good project in the eyes of investors, which is exactly the "trap" I set for Doubao Work.

In short, this is a very complex task. I didn't give any judgment criteria or execution plan, just gave a task in the most leader-like tone, to see if it would fall into the trap.

Let's look at the result.

First of all, what surprised me (and delighted me the most) is not which company Doubao Work selected or not, but that it took the initiative to divide the results into two categories at the very beginning: Category A, mandatory avoidance; Category B, selective avoidance.

I didn't ask for this structure, it added it on its own. The reason why it is "delightful" is that many AIs can strictly execute the Prompt, and at most add a sentence at the end asking "do you want me to do xxxxx". But in the real world, leaders' requirements are often vague, only giving a rough general direction.

Whether it can truly understand the speaker's intention beyond the specific instructions of the Prompt, actively supplement the information structure, and deliver tasks guided by intention instead of "only moving when being pushed" is the line that distinguishes "ordinary employees" from "excellent employees" in the real world.

As for the specific results, I won't show the list here. I can only say that 70% of the companies and judgment criteria it gave are directly approved by me; the remaining 30% are not disapproved, but I don't know much about those companies, and I as a human still need to make further judgments.

By the way, I can "spoil" that it didn't fall into the trap I set for it. It didn't fail to include a certain star embodied intelligence enterprise in the list just because it is a well-known star project — you can ask Doubao Work yourself which company it is (laughs).

In conclusion, after using