Original · Unique Research · 2026-07-30
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, four insight sections, closing reflection, speaker list, and the full roundtable transcript. Product, user, and company figures are speaker self-reports attributed to the named individuals, not independently audited findings.
AI Industry Observation
My AI Work Entry Point Is That Terminal on My MacBook
"The entry point doesn't matter; pain does. Whoever relieves the pain is the entry point."
The person who said this is Zhu Yue (朱越), Chief FDE at Zhidemai Technology (值得买科技). The company is A-share listed, has over a thousand employees, and uses Feishu company-wide. But he spends 80 to 90 percent of his day in Codex and Claude Code. Feishu is just an IM tool for him, plus Miaoji for meeting notes after calls. Feishu positions itself as an "AI super entry point" and iterates hard, but for a power user like Zhu Yue, this entry point doesn't exist.
That's awkward. Over the past two years, DingTalk, Feishu, and WPS have lined up to embed AI into their tools, all wanting to be "the single entry point." But at the Unique Grand Awards AI workbench roundtable, five practitioners' judgment was nearly unanimous: a unified workbench doesn't hold; the real entry point grows on money-related roles.
On the same panel were Hua Kun (华琨), founder of Wavenote; Zhou Ze'an (周泽安), founder of Biyou Tech (必优科技); Li Jinwei (李锦威), growth lead at OpenDesign; Wang Yipeng (王毅鹏), CGO of ATOA.AI (爱智能); and the host was Unique Research partner Duan Hongyu (段宏宇). One has 14 years in Office documents, one built a 70k-star open-source project in two months, and one has deployed AI workbenches in a thousand-person company.
Your "Workbench" Might Just Be Someone Else's Notepad
Zhu Yue divides people at his company into two types. One is like him — deeply using AI tools since 2023, with the terminal as their entry point. The other is AI novices — HR, finance, middle-platform colleagues who find Doubao and Miaoda embedded in Feishu perfectly usable and prefer an all-in-one place.
Both types are real, but their needs are opposite. So Zhu Yue's judgment is direct: unified entry points like Feishu and DingTalk don't hold for mid-to-large enterprises. Data security and compliance need managing; token economics need calculating — these aren't solved by a one-size-fits-all entry.
The more interesting second half: whether an entry point is even a "product" is itself uncertain. His entry point is the terminal; someone else's might just be a chat input box. As long as it solves problems for users under compliant authorization, that's their workbench. "Whether it's monopolized by one big company, I personally think it's unlikely."
Li Jinwei raised the question one level up. His view: plugins vs. workbenches are both "software forms we imagine in our heads" — the question itself is asked wrong. Two things matter. First, whether the entire enterprise's context is fully shared — far more important than software unification. Second, software must solve the attention problem.
Attention has experimental basis. They had the same Agent do two tasks: one focused only on HTML design, no functional considerations, just looking good; the other needed a complete backend, working features, and full front end. Results were completely different — when attention focused on visuals, delivery was clearly stronger.
So his imagined future is: underlying data shared, front-end workbenches can be diverse. Design has its own dedicated design workbench; other tasks use other venues. Workbenches won't unify; they'll fragment into more vertical forms, because focus produces quality.
Zhou Ze'an added a historical angle. His first product was a plugin with 45 days of organic growth; his star product ChatPPT now has 20+ million registered users and nearly 80 million across all platforms, built on three to four decades of Office foundations. In 2023, wanting to do full-chain AI document generation was impossible; the first generation had to borrow Office's rendering and layout capabilities.
But now? Writing documents is no longer Microsoft's or WPS's exclusive. Claude can write, Trae can write, Codex can write — and even "write with a different flavor." His conclusion: plugins vs. workbenches is an evolutionary process. The AI-era workbench, narrowly understood as a unified software collection, is wrong — it's redefining the entry point and reassembling the capabilities that scenario requires. Three people, one conclusion: stop waiting for a super entry point.
Enterprises Pay for Only Two Things
Hua Kun's view is also sharp. He makes Wavenote — hardware plus software that turns recordings into text — but when discussing office, he pulls the topic straight back to business.
"Office isn't simply document processing or communication; concretely, it's eight hours a day and the various roles in a company. Some of these roles are very much about money." Enterprises spend money for only two reasons: higher revenue or lower cost.
Why is everyone willing to pay for coding? Because they can truly stop hiring or cut headcount. Conversely, when Feishu says "today I organized X things for you, let me be your assistant," Hua Kun's reaction is: "I barely notice, because what it does isn't core to my work."
"This sentence is worth pasting on the wall for everyone doing AI office: users don't lack assistants; they lack things that help them make more money or spend less. Because the pain is real enough, operational inconvenience is trivial. Where the money is, AI is. Whichever tool first makes a particular role feel great to use gets paid, no need to worry about entry points."
Middle Management Isn't Dead; the Megaphone Is Dead
After entry points, Duan Hongyu turned to people. His observation: roles with high barriers and long growth curves make veterans more valuable; roles with low barriers, where AI has accumulated extensive experience, actually give newcomers more opportunities. Only the middle layer is hollowed out. Will middle management become redundant?
Zhu Yue used SEO as an example. Before AIGC, an executor hand-crafted articles, producing a few a day at most; middle managers supervised a group of writers, responsible for traffic. AI first liberated frontline productivity, but management won't disappear — middle managers shift from managing people to managing Agents.
He had a sharp line: in the workplace, some people have ten years of experience but it's really one year repeated ten times. In the past, it was unfair to frontline workers — veterans occupied positions, and having handled a site with 500,000 visits was the ceiling. After AI, newcomers can decompose any site's traffic sources, copy, imitate, and take a shortcut.
An early narrative said AI is like Thanos — a snap and half the people lose value. Zhu Yue thinks that's overstated: it's not people who lose value but work that doesn't need humans, like half of low-efficiency communication. And "executives should actually be more panicked" — not embracing AI at the strategic level harms the company more.
Wang Yipeng reviewed the meaning of "middle management" in traditional organizational structures. In traditional organizations, middle management is a necessary link in the information transmission chain. The past growth path was frontline to middle to top, each leap a high-elimination transition, "one general succeeds while ten thousand bones wither" — you have to burn a lot of firewood to smelt a little steel. AI capability empowerment and the rise of OPCs give former "middle layer" workers more opportunities: they can grow within existing organizations or use market-based "value delivery" to collaborate with their former employers. From this angle, it's not "middle management being eliminated" but "middle management being liberated." What's eliminated is only the "megaphone-type middle manager."
Li Jinwei added a new dimension: mental strength. The dividing line isn't experience or age but creative desire. People with strong mental strength are greatly amplified by AI; those without were already just low-value-creation nodes in the organization. So everyone should ask themselves: do I still have the energy to create in this organization?
That SQLite Story — Every Boss Should Hear It
The round's best detail came from Zhou Ze'an. A failure story.
The company has 20+ million users but only a few dozen employees, so customer service is a major problem. With AI coding, operations staff found off-the-shelf customer service tools unsatisfying and coded their own — it actually worked, the interface was decent, and they stacked operational features on top.
But after launch, hundreds of thousands of UV per day crashed the service immediately; crawlers and bots flooded in. Developers went back to the code and found a "especially basic" mistake: the database wasn't connected to the official database but used Node's built-in SQLite in the web app.
"An experienced person would see immediately that this could never go to production; it's fine for personal tinkering." No one on the team caught this in advance; only someone who had held that position and knew what production-grade means could make this judgment.
"This story explains 'middle layer value' perfectly. AI coding is powerful, but it doesn't know SQLite can't handle hundreds of thousands of UV; the person who knows that is someone who's been burned before. AI-era supervision isn't supervising the process; it's supervising whether what AI wrote survives in production."
Zhou Ze'an added a second point. Bosses make transformation decisions quickly, but delivery downward hits habitual process resistance. The biggest driving force is precisely the experienced person in the middle — who has done frontline work, knows what's feasible, and can give executors a path they can get feedback on.
Hua Kun closed from a management perspective. First judgment: AI makes information transparent, so management span can expand. The classic management rule is "one person can manage at most seven" — he thinks this will change: previously limited by information radius and skill boundary costs, those costs are declining. Second judgment is a middle-management survival guide: stand in the boss's shoes and anticipate his anticipation. The boss's brain is always on business and will certainly use AI for efficiency. If you proactively help him use AI well, he won't eliminate you.
Over the Next Three Years, What Will Be the Most Valuable Enterprise Asset?
Final segment: one-sentence bet — over the next three years, what will become the most important enterprise asset?
Zhu Yue gave two. First, how to replicate internal excellent skills, Agents, and workflows through sandboxes. Second, how to cultivate more FDEs and super individuals. Given a foundational model and an open-source project, someone who can solve a business department's problems — these are what enterprises lack most.
Li Jinwei bet on two capabilities. First, the ability to define problems — what is good aesthetics. Second, the ability to judge good from bad. After building software gets simple, people who can judge whether software can enter production become scarce — the SQLite lesson. This answer is counterintuitive: the more ubiquitous AI becomes, the more expensive the gatekeepers.
Zhou Ze'an's answer is one word: adaptability. The scenario hasn't changed — you still use PPT when presenting, just from Docs to Office to ChatPPT. Solve old problems in new scenarios with better methods.
Hua Kun is the most pragmatic: find a clear demographic and need; capture that and you have revenue, you survive. On the supply side, use three years to build an AI-native productivity team — he adds: this is hard.
Written at the End
When the event ended, I pictured a scene.
Big tech is on stage giving keynotes about super entry points, all-in-one, "let AI organize everything for you." In another room, deep AI users are opening a dark terminal window on their MacBook. A few-dozen-person team gets schooled by SQLite; a Fortune 500 design chief anxiously refreshes open-source projects looking for answers.
The entry point war is a keynote war. The real war is in more specific places: whether two hours a day in a certain role can be eaten by an Agent, whether a piece of code can survive hundreds of thousands of UV, whether a middle manager dares to learn how to cut people.
"Zhu Yue says his entry point is the terminal. Hua Kun says when the pain is real enough, inconvenience is trivial. The entry point doesn't matter; the pain does. Whoever relieves the pain is the entry point."
More Conversation Details
Speakers
Hua Kun, Founder & CEO, Wavenote
Zhou Ze'an, Founder & CEO, Biyou Tech (必优科技)
Li Jinwei, Growth Lead, OpenDesign
Wang Yipeng, CGO, ATOA.AI (爱智能)
Zhu Yue, Chief FDE, Zhidemai Technology (值得买科技)
Host
Duan Hongyu, Partner, Unique Research
Duan Hongyu: Let's start with introductions. Please each introduce your company's main business.
Hua Kun: Hello, I'm Hua Kun from Wavenote. We're in the office space, mainly serving business professionals, turning recordings into text through hardware combined with software. We feel a lot of information is still offline; some users, especially older ones, can't quite figure out software, but if the hardware is simple enough, they'll use it. That's roughly what we do, but behind it is mainly software.
Zhou Ze'an: Hello, I'm Zhou Ze'an from Biyou Tech. We're a company focused on AI document Agents — some overlap with others here, but different. Our team has been rooted in this field for years; I've personally done Office documents for about 14 years, then AI documents since 2020. Our products include AI resumes, AI contracts, and the star project ChatPPT. Now we have about 80 million registered users across all platforms, with ChatPPT at about 20+ million registered users, and over 500 enterprise clients including Baidu, Kingsoft, and Tencent — we're a supplier, sometimes exclusive, for many document services. We're a hidden champion in this specific vertical across ToB and ToC. Where we differ from others defining documents is that everyone is racing on how AI improves efficiency or replaces traditional PPT. We believe AI documents should return to the essential demand of content creation itself — content output and content value presentation. Soon, by the end of this month, we'll launch a global public beta of a new product with a completely new look — a delivery-style document product.
Li Jinwei: Hello, our company is called Zhiqi Xinyuan (智启心源), based in Shanghai. We've always done open-source AI projects; in April this year we launched our latest open-source project called OpenDesign. Our original motivation: as AI models' aesthetic capabilities have grown — especially with Claude Opus generation showing huge aesthetic improvement — and as OpenClaw brought massive local Agent普及 earlier this year, we built OpenDesign. The core is letting anyone with a local Agent — local Codex, OpenClaw, or Claude Code — connect to our workbench. Today's theme is workbench, so it aligns well with what we do. Users can pull into our design workbench and design scene to complete more professional design delivery tasks. This project, in about two months, already has 70,000+ stars, with nearly 400 developers from 30+ countries co-building. By growth rate, among open-source projects reaching this scale, few come from AI startup teams — most come from big companies and known teams. So AI has brought huge change to our team, letting an AI startup like ours compete on the same stage as world-class teams. Thank you.
Wang Yipeng: Hello, I'm Wang Yipeng from Aikeyi Intelligent Technology (爱客易智能科技). Our core product is AI Office, including Maidangxiu AIPPT, plus hardware ecosystem products for AI office scenarios. It's quite relevant to today's topic: office needs are universal but low-barrier, and doing them well is hard — we're continuously working on that.
Zhu Yue: Hello, I'm Zhu Yue from Zhidemai Technology. You may all be male users who've used our product, called Shenme Zhide Mai (什么值得买). But today I'm not here to share about that consumer decision platform; it's about AI workbenches. I personally lead internal development of a product — because Zhidemai is an A-share listed company with over a thousand employees, I lead the internal AI workbench product. I'm also in an FDE role. We feel we need to not only define AI capability boundaries, scenarios, and products, but also use FDE capability to make the workbench work in every business department. So I have some insights and ideas that align with today's theme; I look forward to sharing with everyone.
Unified Workbench or Plugin?
Duan Hongyu: We know big companies like DingTalk, Feishu, and WPS embed AI into their tools, and startups like Li Zong are building independent AI entry points. Through your judgment, will there be a super entry point with an AI workbench as the gateway, or will AI be embedded as plugins across all current office scenarios?
Zhu Yue: From our own company, I can divide people into two types. One is like me, deeply using AI products since 2023. The other doesn't embrace this ecosystem or industry. People like me spend 80–90% of our time, apart from iterating our own internal product, using Codex and Claude Code. Feishu is just an internal IM tool for me, or Miaoji for meeting notes — it supplements my workbench entry with information. Although Feishu now positions itself as an AI super entry point through feature iteration, for users like me, that capability doesn't really exist.
The other type is interesting — AI novices in HR, finance, or commercial middle-platform departments. They find Doubao embedded in Feishu, or tools like Miaoda, very easy to use. Because external tools have high barriers, they prefer experiencing capabilities in an all-in-one place.
So from our perspective, because internal capability iteration involves data security and compliance, and the need for true efficiency, including internal token economics to reduce costs —综合考虑, first, a unified entry like Feishu or DingTalk doesn't hold for mid-to-large enterprises. Second, whether an entry point is even a product is uncertain. My AI work entry is the terminal on my MacBook. For some users, the entry might be a product's chat input box. As long as it solves problems for users under compliant authorization and authentication, it can be called that user's workbench. Whether it's monopolized by one big company — I personally think that's unlikely.
Wang Yipeng: If we speculate about the future, look at history. Office workflows and entry points only formed platform and entry-point concepts after digitalization. Going back 20–30 years, workflows moved through paper documents, driven by staff and organizational structure. After 2000 came the SaaS era, where many things modularized into enterprise decision processes or work collaboration. Looking further ahead, if AI becomes the future technology foundation, new interactions including compliance will form. Current financial paper documents must still be retained — not because they must, but because legal rules require it. In the next era, traditional software platforms will still exist, but new interaction entry points may emerge, not necessarily just plugins. But this is speculation, not a conclusion; let's discuss.
Duan Hongyu: What role is your product in this picture?
Wang Yipeng: Our product has two parts: ToC and ToB. For ToC, it's mainly personal productivity tools. First, many early-adopter users. Earlier I mentioned with Zhou Zong that we're collecting resumes from new graduates and was surprised to see "can use VLOOKUP" listed as a personal skill. So there's a big gap between actual application and our R&D frontier, which needs time to bridge. Second, more pioneers are needed in AI tool usage and AI architecture, including our education system.
Li Jinwei: I look at this as two questions. The first is, as AI develops, what will software itself be and what problems should it solve? Whether plugin or workbench, these are both software forms we imagine. So what should software solve in the future? The second question is, if a future organization is AI-native, how should it receive this? Two points behind these questions: first, whether unification — at the workbench or software layer — is less important. What truly matters is whether everyone's context in the enterprise is fully shared, which matters more than the software layer. So we need to solve the first problem: can our future product sufficiently aggregate all enterprise data, whether in Feishu or local Agent interactions, and let it flow and penetrate every person's every business?
The second problem software must solve is attention. Before building OpenDesign, we did an experiment. We had Codex and Claude Code design the same webpage: one focused only on designing an HTML that looks good, not worrying about functionality; the other needed a complete backend, working features, and full front end. Same two tasks, completely different results. When attention focused on front-end visual presentation, visual delivery was clearly stronger. So Agent attention during task execution greatly affects output quality.
Our view is that future Agents and humans both need a professional-chain collaboration venue — a venue where both human and Agent can focus on a specific professional task. Such a professional workbench can be many forms. For example, when designing with an Agent, there can be a dedicated design platform focused on design; both Agent and human know which information, which context to focus on, what task to deliver. For another task, there's another workbench. So it's more about providing humans and Agents with a sufficiently attentive field for higher-quality delivery. Future workbenches needn't be unified, but as long as underlying data and information are shared, front-end workbenches can be diverse. The workbench's function is letting humans and Agents share a more focused attention field in a professional domain, so it will fragment more vertically.
Zhou Ze'an: Let me callback to the topic of workbench vs. plugin. Our first product — the one that grew 45 days organically — was actually a plugin. Coming from Kingsoft, my earliest projects were resumes and those things, ToB and ToC, essentially plugins using others' interfaces. Later I did independent products, but they weren't truly independent — still plugins. This brings me back to what the host asked: is this a workbench or a plugin? I have deep feelings about this. I think it's a process.
From an office-scenario perspective, calling it plugin or workbench — it's definitely a process. Step one: when we started ChatPPT in 2023, pushing an independent product where AI does full-chain document design was nearly impossible. We could only take Office's 30–40 years of foundation, extract its atomic capabilities, leverage its rendering and layout, combine with strong AI capabilities, and organize it into a first-generation AI generation — that was a critical point.
But now, hearing about document processing this year, it's no longer familiar Microsoft or necessarily WPS. You can see Claude can write, Trae can write, Codex can write, and even write with a different flavor. It's first a process, and in my view an evolutionary process. Long-term, the concept of a workbench in my mind has both the context unification just mentioned and, from our application scenario, will emerge. Because the workflow is completely different. The AI-era workbench, narrowly understood as a unified software collection, is wrong — it's redefining an entry point and re-aggregating the capabilities that scenario requires. At first the workbench may have limited capabilities, but as more creators join, it will definitely form a scope in specific domains. It's not a clean break; it's an evolutionary process.
Hua Kun: I have a slightly different view. Let me state it: personally, short-term unified workbenches don't mean much. Office isn't simply document processing or communication; concretely it's eight hours a day and the various roles in a company. Some roles are very much about money, so enterprises spend money on them. Two dimensions: higher revenue or lower cost. Why is everyone willing to pay for coding? Because they can truly stop hiring, increase workload, or cut people. Since Claude reached that level earlier this year, I've been thinking about how to make R&D more efficient and cut certain roles.
Advertising is another area that needs AI badly — we think about ROI every day. There are many advertising roles, like designers producing better materials more efficiently, all money-related. But Feishu keeps saying "today I organized X things for you, I'm your assistant now, let me do it" — I barely notice, because what it does isn't core to my work. I care about how to optimize ad ROI and how to raise certain roles' ROI.
So if a good tool can greatly improve this, I'll definitely use it — whether installing it or whatever. Some entry points are convenient. I'm doing cross-border e-commerce now with Shopify, Google Ads, Facebook — I think about these every day. But honestly, assistants aren't great; many questions get bad answers. We still use Claude or GPT, paying for the latest versions. Maybe we use Anthropic and OpenAI, constantly topping up. Why? Because although it's not great — I have to screenshot, throw it in, ask many questions — if something automated could review all these images and directly tell me conclusions, I'd love it. But it doesn't matter; because it solves my pain point, because I'm in enough pain, operational inconvenience is trivial.
So you must solve real enterprise problems related to money. A startup picks a small entry point that enterprises care about — truly reducing costs, increasing efficiency, or opening revenue. Big models are doing this too, starting from where the money is. Claude is a huge market in coding because Silicon Valley has lots of people to cut, and China too. In March in Silicon Valley, it was already well done, but OpenAI, Meta, Google, xAI are all in the race because the track is big enough. Anthropic is also competing, starting with financial analysts because that track's people are expensive and worth mining for money. There are countless examples.
To this day, there's no great tool that lets a role say "I can spend half the money, or two people can cut one" — not quite there yet. But enterprises desperately need this; they see the hope and the capability but it still feels bumpy, missing something, so they make do. But you can see it's moving forward. If something works particularly well in a specific role and scenario, enterprises will definitely pay. Long-term, could there be a more unified entry? Maybe, but the prerequisite is that a fixed daily task taking two hours is well-solved by an Agent — that's very meaningful, and enterprises will pay. If startups build something great in these small spots, no need to worry about entry points; they can integrate into big platforms.
Duan Hongyu: So there's expectation for the convenience a super entry point brings, but the prerequisite for landing is that results and quality must be excellent; otherwise enterprises prefer complex paths.
Will Middle Management Be Hollowed Out by AI?
Duan Hongyu: In discussing AI workbenches, enterprise productivity restructuring is actually tied to people. Previous guests also talked about cutting many middle managers. From our three years of summits, the prevailing view is that for roles with high entry barriers and growth difficulty, experienced veterans are increasingly valuable; execution-level newcomer work is largely replaced by AI. For roles with low barriers and limited experience premium, AI has accumulated extensive experience, lowering entry thresholds and giving execution-level newcomers more opportunities. Only the middle layer is hollowed out. I want to ask four guests: will middle management become redundant, or more important in some scenarios? Zhu Zong, please start.
Zhu Yue: Because Unique Research has strong going-global relevance, I'll start from a going-global role. Google has SEO; from an SEO perspective, before AIGC, an executor hand-crafted articles, producing a few a day at most. AIGC liberated the SEO executor's productivity. Middle managers in SEO managed writers, responsible for SEO traffic and data outcomes. At the career top are strategy and marketing strategy setters. AI first solved productivity for frontline and middle layers. Management is still needed, but middle managers shift from managing people to managing Agents. There will still be people executing tasks. So middle management certainly still exists; the concepts of management and doing work still exist, but they offer many emergent-capability opportunities. In the workplace, some people have ten years of experience but it's one year repeated ten times; others genuinely have ten years. In the past, this was unfair to frontline workers, because veteran middle managers occupied positions — having handled a site with 500k or 1M visits was the ceiling, unwilling to explore upward. AI lets newcomers analyze how any site's traffic comes from, decompose PPT traffic sources, learn SEO points, copy, imitate, and take shortcuts. It gives people in this industry more opportunities.
Short-term, middle management still exists but may see more personnel changes — those who aren't up to it step down. Including the earlier narrative about the internet/AI industry being like Thanos — a snap and half lose value. I don't think it's that dramatic, but what loses value isn't people but work that doesn't need humans — half of low-efficiency communication. Not just middle management; executives should be more panicked. If an executive's strategic side doesn't embrace the AI era and drive company transformation, the harm is greater. So this is something everyone needs to think carefully about.
Duan Hongyu: So it sounds like new middle managers, new executives, and new executors will emerge.
Wang Yipeng: Rather than saying middle management is eliminated, it's liberated. In traditional organizational structures, there's an information and task transmission chain from top to bottom; middle management must convey, interpret, back up, etc., making it a critical link. Now with AI empowerment — and even before AI, the e-commerce and internet eras already talked about flat decision-making. So it's not whether the role is called middle management; it's whether the person has grown. The past growth path from frontline to middle to top had high elimination rates at each leap, high growth costs, one general succeeds while ten thousand bones wither. Whether in sales, business, or internal tech, you have to burn a lot of firewood to smelt a little steel.
Now with new technology, people can more freely develop growth space — whether through OPC controlling their own business, or using OPC to form value-chain interactions with their original organization to realize their value — it gives more imaginative space. Many people now doing OPC projects were originally functional modules in enterprises; they voluntarily say they don't need the thick middle ceiling and prefer to deliver results in exchange for growth space. So the term "middle management" may disappear in specific industries, because technology upgrade gives people more choices and more value-exchange opportunities.
Li Jinwei: Let me tell a real customer story. When we launched our product early, a very well-known Fortune 500 consumer electronics company's design department head saw our open-source project and reached out. His first words were: "I feel very anxious because the group requires AI transformation starting from the design department, but I don't know how." He was following various AI projects daily until he found us. We talked and I said: why are you anxious? You should instead free up more energy for professional aesthetic judgment, doing more specialized work.
Looking at design department managers or leaders now, are they doing aesthetic judgment, people management, or project management? AI should gradually shrink the energy spent on the latter two, letting the person truly focus on professional matters. A design head should focus energy and thinking on professional aesthetic judgment, letting judgment guide the enterprise's brand tone and visual identity. That's the first point: freeing up energy previously spent on pure people management and project management.
Second, different people experience AI's impact differently. Rather than experience or age, the bigger difference may be mental strength. In people around me — colleagues — my real feeling is that someone who was already mentally very strong with strong creative desire, even without prior experience in this area — say someone who did SEO but with AI can quickly pick up other growth directions, quickly know how to work with influencers and place ads, quickly compensating. In the past, mastering another professional field and creating in it was hard. So people with creativity and mental strength get greatly amplified. People without strong mental strength easily fall into panic from AI's impact because they don't have strong creative desire. People with lower mental strength may have just been low-value-creation nodes in past organizational processes. So the question to ask yourself is: do I still have enough energy to create in this organization? If yes, AI will definitely be a huge multiplier for you.
Duan Hongyu: So for roles where experience matters less, judgment matters more; for roles where experience matters more, AI solves it well, so management span or reach expands.
Zhou Ze'an: Whether middle management should be reduced or replaced — in my view it's a value-transfer process. In the past, middle management mostly did megaphone work or process management. With AI, bosses need fast decisions and daily response to change, while subordinates have AI and new productivity. Beyond what Zhu Zong mentioned — from supervising people to supervising Agents, which is formal change — the biggest change for middle management is a fundamental shift in value positioning. First, shift from managing processes to managing judgment. Judgment has two aspects: first, having judgment on existing things. Being technical, formerly doing algorithms, I found that this year's coding went from extreme excitement to extreme pessimism to now slightly calmer. At first seeing AI could code — I hadn't coded in years, doing management and entrepreneurship — I thought efficiency would improve instantly. But writing came out poorly, low quality, and couldn't go live.
At a certain stage, although you have judgment, you lack the daily internal friction from facing specific roles — internal friction a boss can't imagine — which requires experienced, judgmental people in the middle. For example, our internal OA was often "we can just code it ourselves." A very real problem: one or two million users, a few dozen employees, how to solve customer service? Previously it was last-gen AI customer service; with new AI, operations staff thought WeChat customer service and others weren't good enough, so they coded one themselves. They copied it; initially the interface and experience looked good. But after deploying, they stacked another feature — putting daily operations and basic operational data, even article publishing, into that system. What looked like a simple customer service tool attracted bot crawlers, and with hundreds of thousands of daily UV, the service crashed immediately. Developers went back to the code and found a very basic mistake: the database wasn't connected to the official database but used Node's built-in SQLite. An experienced person would see this immediately can't go to production; it's fine for personal tinkering. At the time, nobody caught this; only someone who had held that role and knew what industrial-grade or production-grade means could make that first value judgment. Now it's not supervising the process but supervising whether this approach and process can be completed under AI's perspective — that's the first point.
Second, during transformation, bosses make decisions quickly, but the real difficulty is delivering the transformation to subordinates — saying we need to do this technical direction, facing people's resistance in habitual processes where AI coding is unreliable. But often the biggest driving force is the experienced person in the middle — who has done frontline and foundational work, knows what's doable, and can quickly give all executors a result with height, traceability, and observability that gets feedback quickly. So these two things are very important for all middle management. But if you don't want to transform, the consequences are serious — you may be eliminated.
Duan Hongyu: So this kind of middle manager has execution-detail experience that executives may not have.
Hua Kun: Let me approach from another angle. This role is relative — small companies don't need it; large companies managing a thousand people still have middle management. Let me talk about possible changes AI brings to management, and if you're a middle manager now, how to move up — otherwise, not advancing means retreating, and not embracing AI will get you cut.
First, AI makes information transparent; many skills and methods become clearer, and data becomes clearer too. So management span can actually expand. Management theory and HR always said one person manages at most seven. I think this will change. Previously you couldn't manage many people because information radius was insufficient and skill boundaries created costs — but these costs are actually declining.
Second, how to move up: stand in the boss's shoes. The core is that business still matters; the number one's brain is always on business. How to help more and become indispensable? Pay more attention to business. I've managed many middle managers and found few who are very strong — this is something AI may struggle to replace.
Also, use AI well. Empathize: quickly use AI well from the boss's perspective, raise efficiency. First, management span may expand; second, the boss's brain is definitely on using AI for efficiency, so anticipate his anticipation and stand in his shoes to help him get things done — then he won't cut you.
If you can truly help the company build an AI framework — and this inevitably includes learning how to cut people — you can move forward and gradually become indispensable.
Duan Hongyu: If someone's management span expands, someone else's shrinks. What kind of people see their management span expand?
Hua Kun: First, use AI well to raise your own effectiveness. Either manage more people or produce more. Because management isn't about how many people you manage but what value you bring to the company and your team's human efficiency. From the executive perspective, that's what they look at.
One-Sentence Bet: Most Important Enterprise Asset Over the Next Three Years
Duan Hongyu: Time is limited; last question. One-sentence bet. AI has been around for three years; over the next three years, what do you think will become the most important enterprise asset? Zhu Zong, start.
Zhu Yue: When we do internal workbench-style efficiency tasks, we have two red-line principles. First, embrace the open-source ecosystem — I can reveal that this product is built on open source. Second, when productizing or optimizing for business, absolutely never compress model capability for the sake of productization. Any super individual, through open-source projects combined with the strongest model capabilities, can already solve many problems — as long as they can find APIs and MCP themselves. For enterprises, the two most important things may be: first, how to replicate internal excellent skills, Agents, and workflows through sandboxes or better methods. Second, how to cultivate more FDEs or super individuals internally. Regardless of department, right now, given a foundational model and a foundational open-source project, you can solve many problems in a business department with AI. What people currently lack most is this. So part is people-related, part is technology-related — likely the most important asset for enterprises wanting AI transformation.
Wang Yipeng: This question is broad; different enterprises in different states and lifecycle stages have different answers. For startups, the most important thing over these three years is building a team that collaborates well with AI, plus healthy cash flow. These three years, this may be most critical.
Li Jinwei: From this perspective, AI capabilities are the same for every company; everyone can use the same AI. The key difference is people. At the individual level, the core most valuable capabilities are two. First, the ability to correctly define problems. Building a design product requires defining what good aesthetics is, what good design is. So we need people who can define these problems. When hiring, we focus on whether we can recruit top-tier designers, world-class aesthetic designers who can help define these problems. Second, the ability to judge good from bad. Today, when building software itself becomes simple, you need to judge whether software is qualified, whether it can enter the user's production environment. So we focus on hiring another type: QA engineers. Such QA engineers may have seemed less important in the past, but today, when the entire software co-building process becomes very lightweight, the person who holds the final gate on whether software is good or bad is equally precious.
Zhou Ze'an: For a company — especially in the AI era, from our own experience — one important thing is adaptability, which relates to the team. Regardless of what we do, we've evolved from traditional software through 2023–2026 — the assisted phase — to today's native, truly Engine-driven. What we do never changed; I'm very old-fashioned, I've only worked in one industry for all these years, but every conversation I still feel excited. What excites me is the ability to adapt. The scenario hasn't changed — you still use PPT when presenting, just from Docs to Office to now ChatPPT, and the next product may look completely new, but the scenario itself hasn't changed. The first core is adaptation — re-solving old problems in new scenarios with better methods; this is very important. It's also important for the team. When new technology, new capabilities, and new scenarios explode, you must quickly adapt rather than purely resistively clinging to old ways. This itself requires adapting to direction control and adapting to the response after change, maintaining passion.
Hua Kun: Not talking about the abstract three years but concrete 2027, 2028, 2029 — for startups, the most important thing is finding a relatively clear user need and demographic in the AI era, whether ToC or ToB. Because if you capture this group, you have business, revenue, and survival — otherwise the company won't exist in three years. From the supply side, use these three years to build an AI-native productivity team that fully leverages AI tool capabilities and organizational efficiency — this is actually very hard.