--- title: "The Better AI Makes Documents, the Less Useful Documents Become" author: "Unique Research" sourcePublication: "Unique Research Substack" originalPublishedAt: "2026-04-12T12:01:35+00:00" canonical: "https://ffcap.cn/en/research/src-20260412-03html" source: "https://uniqueresearch.substack.com/p/src-20260412-03html" language: "en" --- # The Better AI Makes Documents, the Less Useful Documents Become _Original · Unique Research · 2026-04-12_ _Editor’s note: The opening first-person commentary belongs to the original Chinese author. This edition preserves that report and all 28 panel speaking turns. The narrative describes five entrepreneurs; the published roster distinguishes four guests from the moderator, who also describes himself as an entrepreneur. Product user counts, financing, company relationships, model capabilities, the “Hire of Human” name and its claimed Claude connection, autonomous on-chain transactions, labor claims and future unemployment are retained as statements by the source or speakers, not independently verified findings. The prototype is an anecdote, not an experiment performed for this article. The P/ARR figure is reported hearsay, not a verified transaction valuation or investment advice. The source distinguishes around 17 million product users from more than 80 million registered AI-document users; neither figure establishes active or paying users. Zhou’s description moves from implementation to delivery, while the moderator reverses those labels in his reply; both formulations are retained. Company and personal names without confirmed official English forms are transliterated provisionally. References to tending a “lobster” retain the source’s OpenClaw-related colloquial language. The original source date is April 12, 2026; relative dates and projections remain relative to the discussion, without an inferred event date._ Unique Research A Man Who Has Spent Thirteen Years Working on Documents Said Something I Kept Thinking About As AI makes formatting, structure, and wording ever more polished, the only thing that truly determines a document’s value is your own judgment. " If everyone uses AI to express themselves, everyone’s views become mediocre. Zhou Ze’an of Biyou Technology has spent thirteen years working on AI documents, with eighty million registered users. At the roundtable, he said: if everyone uses AI to express themselves, everyone’s views become mediocre. I did not quite grasp it at the time. Only after thinking about it for a long time did I realize that he was describing a paradox unfolding right now: we use AI to make documents better and better, but may also be eliminating the reason documents exist. A document conveys your judgment. AI can optimize its formatting, structure, and wording. But if the judgment itself is also “optimized” into some kind of average, whose views does that presentation you hand over actually represent? The question was not fully answered, but it drew a line. For me, that was the point in this roundtable most worth thinking about. I listened to this roundtable in Hangzhou: five entrepreneurs who had worked in intelligent office applications for years. To begin, each summed up 2025 in one word. Pang Dawei said evolution, Zhou Ze’an said implementation, and Ma Liang said the cloud. Gu Chenggang of ATOA.AI paused, then said: alive. The moderator asked: what about 2026? He said: stay alive. Gu Chenggang said these two terms were realistic. There is something about “alive → stay alive” that cannot be performed. Ma Liang of Guiji Jike said something that left me a little stunned. He said that during this year’s Spring Festival, a product called “Hire of Human” appeared in the Claude ecosystem: AI posts tasks, people accept them, complete the work, and report back; a person supplies the money. He then ran his own experiment, letting two AIs assign tasks, sign contracts, and settle payments with each other on-chain. With no human participation, a small economic system ran by itself. He said he planned to open-source it. You can call it a toy. But it is hard to say it does not point toward a possible direction. After that, he added: in the short term, over two or three years, we will inevitably endure the pain of a wave of unemployment, but things will improve after the transition. That was his actual judgment, not something said to comfort you. Investors have recently liked asking a question that makes entrepreneurs very uncomfortable: once foundation models become more capable, will the application layer not simply die? The question is devastating because of its implied logic: what you are building is essentially waiting to die. Pang Dawei did not evade it: you need to understand how far you are from the model’s growth trajectory, how much you can grow yourself, and whether you can make it through this cycle. None of those three questions has a standard answer, and none can be answered with a presentation. But their answers shared an underlying logic: do not compete with models on general-purpose capabilities. You cannot win that race. Your space lies in things models will not handle carefully and have no incentive to handle carefully. Specifically, that means private data, Know-how accumulated over more than a decade in a particular industry, and the hard-to-replicate capabilities created when you turn it into SOPs and then code. Gu Chenggang tells investors directly: my competitive barrier is my users. Large companies serve broad audiences; I serve precisely targeted users. If you recognize that, invest. If not, we will find someone else. There is unusual confidence behind those words—not confidence in technology, but the confidence of genuinely knowing whose problem he is solving and what it is. He therefore does not need to dress himself up as something else. Moderator Tang Minglei shared a detail about his own company: more than ten internal systems, not a cent spent on SaaS, all built by team members born in the 2000s using AI Coding. He had become an entrepreneur after ten years in VC and had been running his business for four months. He said this was a wonderful era. It is hard to call that performed optimism. Someone who has been an entrepreneur for only four months and has built all the company’s internal tools with AI has grounds for saying it. Back to Zhou Ze’an’s remark. AI can do more and more things for you. But there is one thing it cannot do in your place: decide what you really think about something, what you are willing to take responsibility for, and what you believe is right. You can outsource efficiency, but not judgment. This is not an argument that people are more capable than AI. It means that when AI makes everyone’s documents equally beautiful, the only thing that can distinguish your expression from someone else’s is you. More Details from the Conversation Unique Awards · Hangzhou AI WEEK Trends Roundtable Panel “Intelligent Office Work: When AI Takes Over Data Flows, Humans Focus on Creating Value” Guests: Pang Dawei — Founder & CEO, Yuankong AI Gu Chenggang (Gavin) — Founder, Aikeyi · Aizhineng / ATOA.AI Ma Liang — CEO, Guiji Jike Zhou Ze’an — Founder & CEO, Biyou Technology Moderator: Tang Minglei — Co-founder, Panfeng Intelligence, and experienced investor Tang Minglei: Today we are discussing intelligent office applications. All four of you are veterans who have worked in this field for years and have also entered AI. So first, I would like each of you to briefly introduce yourselves. Let us open with this question: use one word to sum up AI in 2025, and one word to look ahead to AI in 2026. We will start with Pang. Pang Dawei: I am Pang Dawei from Yuankong AI. Another relatively well-known product of ours is ChatExcel. The name tells you that we entered AI office work and AI Data through Excel, offering a complete AI-office suite to solve users’ productivity problems. On your second question, my word for ’25 is “evolution.” Things have been iterating rapidly, from month to month and even week to week. Both user needs and AI products’ problem-solving capabilities are evolving quickly. Entrepreneurs and users alike can feel that process. For ’26, my word is “boundary-breaking.” Why that word? After OpenClaw emerged, we released what we describe as China’s first, most usable mini-program version of OpenClaw in early March. We saw an opportunity to break boundaries and have already validated it: people can enter the entire productivity process. So my word for ’26 is “boundary-breaking.” Gu Chenggang: My name is Gu Chenggang, and I founded Aikeyi · Aizhineng, ATOA.AI. Our company mainly develops suite applications in AI Office. We have been doing this work for some time, and we hope people everywhere will no longer need to make PPT presentations. My word for ’25 is “alive,” and for ’26, “stay alive.” That is a deeply felt experience. Ma Liang: I am Ma Liang from Guiji Jike. Our product is an intelligent voice assistant called All-in. On a PC, it has capabilities similar to OpenClaw, but that is not all it does. It can type, process text and files, and perform comprehensive tasks of the sort OpenClaw handles. In short, the interaction is entirely voice-based. If I had to sum up ’25, I would say “the cloud.” The hottest things were certainly those products released in March, including Agents that later ran in cloud sandboxes. But from the models released this January through the emergence of OpenClaw, we found that everything had moved to running locally on devices. So my word for ’26 is “local.” We can discuss the differences in more detail later. Zhou Ze’an: I am Zhou Ze’an, founder of Biyou Technology. First of all, I am someone deeply involved in AI documents and have worked in this industry for about 13 years. Biyou is an AI document-agent application company. Its products include flagship projects in resumes and PPT. Those products have around 17 million users, while registered users across AI documents exceed 80 million. We are something of a hidden champion behind companies such as Kingsoft, Baidu, and Tencent in the specialized, purely native document field. My first word for 2025 is “implementation.” Last year felt different from ’23 because people were actually using the products, whether on the consumer or enterprise side. That was a basic sign of real implementation. This year, I would choose “delivery.” Previously, people were testing the waters. Moving from implementation to delivery, we now clearly see that expectations of results have risen. Technological iteration further intensifies demand for those results. Tang Minglei: That is the shift from final delivery to implementation. Personally, I think AI has evolved considerably in ’25 and ’26. I summarize ’25 with the letter “C”: Chat, Coding, or Creator all reflect progress in AI Capability. In ’26 it may be “R”: Remember, Research, or Run. AI directly runs the task and produces a Result. That is a substantial change. The other day, I posted on Moments that AI is evolving so quickly: might we move from “humans using AI tools” to “humans employing AI staff,” then “AI employing human staff,” and ultimately “AI employing AI staff”? I would like to ask each of you: what is the future of intelligent office work in ’26? Where is the boundary between humans and AI? Will tools really still exist? What should their ultimate form be? Let us start with Pang. Pang Dawei: I use one word for the relationship: “symbiosis.” From a Darwinian and biological-evolutionary perspective, paramecia and ants still exist after five hundred million years. Humans and AI will also coexist symbiotically, including AI employees and AI productivity. Fundamentally, they are in a state of coexistence. Time will determine the boundary. Penetration in many traditional industries and high-stakes settings is still not that high; adoption gradually proceeds over time. Whether APPs and other tools will be eliminated is also a matter of time and use case. In some situations, the tool layer will be subsumed. I agree with Principal Tang about reaching the Result layer: once people believe AI can deliver results, the process ceases to matter. But I think ’26 has another C alongside C and R: “Connect.” The value of Connect will become prominent in ’26. Moving from Connect to Result will change every kind of use case, not just office work, and create limitless possibilities. Gu Chenggang: When I was working on artificial intelligence in ’15, I had a clear view that it might have four stages, corresponding to the four Chinese characters in the term: human, work, intelligence, and capability. Today’s software applications are called tools and sit at the boundary between “work” and “intelligence.” The previous generation consisted of people and work. OpenAI was built by piling up large quantities of data and the labor of many Indian workers. This year will see an explosion of tool applications, and we are about to enter the year of “intelligence.” Ultimately comes “capability,” what we generally call AGI. But I have a definite view: people’s ultimate goal is to enjoy life. Will humans therefore be replaced by AI? Absolutely not. Will we be led by AI? Personally, I believe that will never happen either. Our name, ATOA.AI, actually carries three meanings. At the current stage it means AI to AI. The second stage is Anyone to Anyone: people with AI capabilities help Connect those who do not know AI. The third is Anything to Anything, with AI driving connections among everything. But people will always be the masters of the universe. Tang Minglei: Gu’s remarks may give everyone a greater sense of security: perhaps AI is not evolving quite as fast as we thought, and people can still control much of what needs doing. What is your view, Ma? Ma Liang: Humans employing AI is already happening. API Tokens are quite expensive, and people spend a fair amount on them. As for AI employing people, during the Spring Festival a product called “Hire of Human” appeared in the Claude ecosystem. It posts tasks in the physical world that AI cannot reach. AI posts the task, a person accepts and completes it, and reports back to AI; a person pays. That is already happening. AI employing AI seems very natural to me. Two weeks ago I coded an experimental prototype: AI issued cryptocurrency on-chain, paid other AIs, and signed contracts. One AI assigned work and another accepted and delivered it. A small economic system ran completely autonomously. I plan to open-source it for people to try. Will tools still exist? I think they always will. Large models keep predicting the next Token and can perform planning and logic, but execution requires calling external tools. A model cannot contain every tool. Far into the future, intelligent office work might not require offices at all. Everyone could collaborate online through the combined capabilities of an individual and AI. In the short term, over two or three years, we will inevitably endure the pain of a wave of unemployment. But after the transition, things will improve. Tang Minglei: After the transition, everyone will have become accustomed to it. I also said the other day that we can cover ourselves in AI tools and become ever more competitive, but in the AI era we can also do things unrelated to AI. Starting businesses that teach people with more leisure time how to have fun or spend their holidays could be a good idea too. What do you think, Zhou? Zhou Ze’an: The office-document industry differs from other fields such as design or marketing in one fundamental respect: documents serve people by expressing their views. If AI does everything, the result cannot represent your views. If everyone speaks through AI, everyone’s views become mediocre. That means people and AI must find a balance and create together. You instruct AI to do something, then decide whether its work represents you, incorporate your own views, and produce another version. The content should reflect both personalized needs and the highest level of human decision-making wisdom. We talk about outcome-oriented documents this year because the real result is whether the PPT you take onstage communicates your views to the audience. A beautiful document without a point of view is useless. In the future, people may determine the intermediate process together with AI. For now, humans still take responsibility and take the blame for the result. But one day AI may share that responsibility with people. Tang Minglei: People still have the core function of taking the blame. It cannot simply be signed over to AI in the near term. Let us turn to competitive barriers for AI-application startups. I have recently talked with many investors who worry that an upgrade to a foundation model will flatten applications like a road roller. What do you think is the core barrier for AI applications in intelligent office work: speed, engineering capabilities, the number of templates, or understanding users’ Know-how? Pang? Pang Dawei: Every major company and investor is asking this. Model upgrades will inevitably cover many general-purpose use cases. Particularly after OpenClaw’s release, the Coding war will surely end this year. It is only a matter of time before many use cases are subsumed. The fundamental questions are: how far is what you are building from the model’s growth trajectory? How big can you grow? Can you make it through the cycle? We are entrepreneurs first. We need to consider whether the outcome can be delivered through AI in a way that lets us make it through the cycle. I think competitive advantage still lies in the team’s combined strengths: speed, understanding, and accumulated technology. Gu Chenggang: Almost every company represented by our partners onstage received investment from the state-assets regulator last year. Pang received three rounds, and we have also just received investment from the state-backed camp. Technologically, I do not think small applications have much of a technical barrier. I often tell investors: first, users are my barrier. Large companies cover general users; ours are precisely targeted. Second is the time barrier. Everyone here has spent around three years doing this, while Zhou has spent 13 years. That investment of time is something a big-company team cannot match. If you recognize it as a barrier, invest in me. If not, we will find someone else. Ma Liang: Large companies develop models and also enter applications, which does have a substantial effect on the ecosystem. But there are two things models can never swallow: tools that solve problems in particular domains, and private data that large companies cannot obtain. Big companies will inevitably swallow some startups, but we can keep iterating along the way. A small boat turns more easily. We will certainly be faster than big companies; find the right position and you will be OK. Zhou Ze’an: Investors’ questions have changed. They used to ask whether the audience was too small to scale. Now they ask whether it is sufficiently targeted and whether we can make money from it. Foundation models are the basis of value, and applications are the expression of value. The two cannot be set against each other. I have two principles for building a temporary competitive barrier. First, dare to develop deep insight into a specific use case. Big companies go broad; we go deep, covering only core, high-value users who can convert. Second, completely rebuild the workflow for that scenario. Travel light: do not merely wrap an API; genuinely eliminate a pain point. Turn the team’s accumulated Know-how into SOP skills and encode them in software. If AI understands your business, you can go further than others. Tang Minglei: Very good. Last year, while looking at AI investments at Alibaba, I summarized the barriers for AI products in three phrases: start with speed—UI/UX interaction, financing, and co-creation; stay true to technology—engineering, industry-template fine-tuning, and vertical models; and end with users—data, habits, and mindshare. Only after securing a use case do we have an opportunity to push back. Models and applications move toward each other. Finally, how do you view investment institutions? Will investors still have FOMO in ’25/’26? Please offer some advice on dealing with them. Pang Dawei: In ’25, we raised three rounds in half a year, then spent the following months working on the product and overseas expansion. Investors now care more about where the barriers are and how the company can survive and become bigger and stronger. That raises the demands on startup teams—not just product development, but also vision, technological reserves, and iteration speed, which used to be monthly and is now weekly or even daily. I think sincerity is most important. Communicate more and learn more. Nobody can do everything, and every interaction is an opportunity to grow. Gu Chenggang: The relationship between founders and investors is a lot like dating: if you are good enough, someone will be waiting for you. ’15 was the era of raising money with PPT. ’25 is the era when you need a product, revenue, and users to raise money. Only when you have all three will people pay. Investors value certainty because they are reluctant to spend during a consumption downturn. ’26 may be harder, but it is also the best of times. Ma Liang: Investors have a lot of FOMO. There are so many projects that they only have the capacity to examine one-tenth of them. My advice is to use AI Coding and other methods to put a Demo online quickly and validate it. Complete a small working loop, get some users and revenue, and then seriously seek investors. Your success rate will be much higher. You can no longer raise money entirely on personal connections or reputation. Zhou Ze’an: Investors’ FOMO will definitely continue, because OpenClaw and similar developments have brought new opportunities. My advice is, first, maintain a good mindset and produce data that reassures investors. Second, show that you can genuinely embrace and handle change. If a big company builds what you do, you can do something else. Use your existing data to demonstrate that you can capture the next opportunity. Tang Minglei: People are currently saying internally that P/ARR valuations can still reach 20 times, so seize the moment for FOMO fundraising. Finally, please offer one sincere message to people working in the AI era. Zhou Ze’an: Keep holding on to your insight and feel for each specialized use case. That will be the best remedy as you confront the changes AI brings. Ma Liang: Do not be too anxious. Slow down a little. Tend to your “lobster” when it needs tending, sleep when you need to, and take care of your health. Gu Chenggang: As someone born in the 1970s who has experienced the transition from the mobile internet to artificial intelligence, I think this is the best era. Just go and do it. Pang Dawei: People who want to work in AI must dare to question themselves and keep learning. Accept the shift in paradigms so that you can get through cycles and go further. Tang Minglei: I am a four-month-old entrepreneur; before this, I spent ten years as a VC partner. As someone born in the 1980s, I am excited to have caught this entrepreneurial opportunity in the AI era. Our company has not spent a cent on SaaS. More than ten systems were built entirely by people born in the 2000s using AI Coding. This is a wonderful era. In the AI era, speed is unstoppable, and so is love. I hope everyone will bravely climb their own peak. Thank you! --- Original publication: https://uniqueresearch.substack.com/p/src-20260412-03html On-site reading page: https://ffcap.cn/en/research/src-20260412-03html