Original · Unique Research · 2026-05-14
Editor's note: The first-person interview and its judgments belong to the original Chinese author. This English rendition retains the opening essay, all interview sections, and the complete 10-question Q&A. Founder background, company history, market claims, and technical terms are source or speaker claims, not independently audited findings. Company, personal and work titles are transliterated where official English forms remain unverified. The term "DLA (Document Level Agreement)" is retained as the speaker's coined term.
Unique Awards · Guest Interview
Agent Is Here, and the Entry Point for Documents Is Being Rewritten
The Elimination Standard for AI Office Products Has Changed
"
Generate 10 times and only 1 is usable—that's a toy;
Generate once and 80% is usable—that's a tool;
Generate the final draft with one click—that's what delivery means.
Last year, a user used AI to generate a PPT, then spent three or four hours revising the copy, adjusting the layout, and fixing the design. This year, if the same user encounters this situation, they will give up after just one try.
When Zhou Ze'an told me about this change, he was very calm, but I think there is a rather cruel signal hidden within it: the elimination standard for AI office products has changed.
He is the founder of Biyou Technology, the person behind ChatPPT. He has been in the document space for 13 years—WPS Product Director, Kingsoft Office's Most Valuable Professional (KVP), created Pocket Animation in 2015 (at the time the largest Office plugin by user count in China), acquired by Kingsoft in 2017. He left to start Biyou Technology in 2020, and its core product ChatPPT was one of the earliest Chinese AI PPT generation tools.
This year they did something noteworthy: launched the world's first MCP Server based on document creation, followed by ChatPPT Skill and a CLI refactored version. Behind the document capabilities of major companies like Baidu and Kingsoft, Biyou is one of the underlying service providers.
How does someone who has been immersed in the document field for 13 years view the changes happening in AI office?
From "Can It Work" to "Can It Work Directly"
Zhou Ze'an made a judgment last year: the keyword for 2025 is "landing," and for 2026 it is "delivery."
I asked him, looking back now, has this judgment changed?
"It has become stronger, and more certain."
He said that user behavior in 2025 was essentially curiosity-driven "verification actions": throw a task at AI and see if it can do it. But by 2026, what users are starting to ask is not "can it do this," but "can it actually get this done so I don't have to rework it."
Then he gave a layering I found particularly precise:
"Generate 10 times and only 1 is usable—that's a toy; generate once and 80% is usable—that's a tool; generate the final draft with one click—that's what delivery means."
Most AI office products are currently stuck between "tool" and "delivery." Users are no longer satisfied with the experience of "AI helped but I still have to spend two hours finishing up."
So what exactly are users paying for?
"Before, they paid for novelty; now they pay for certainty."
He put it very directly: "Salespeople buy the probability of closing deals; office workers buy the freedom to leave work on time. Whoever can guarantee the results gets the money."
After hearing this, I thought the phrase "freedom to leave work on time" was quite clever. Ultimately, the ultimate value of AI office products is not making PPTs faster, but making it so you don't have to make PPTs at all.
The Plastic Flower Problem
But there is a paradox here.
AI is increasingly capable of generating beautiful documents—neat structure, standard wording, refined layout. The problem is—if everyone uses AI to generate, then everyone's documents look more or less the same.
Zhou Ze'an has thought deeply about this issue. He has been in the document field for 13 years, seen countless document tools, templates, and methodologies, and finally arrived at a conclusion:
"Information is cheap—it's everywhere online. But viewpoints are expensive—they are your judgments about the business and your insights into the market."
"The essence of a document is a person expressing their own viewpoints, thoughts, and judgments in a standardized format. The format is just the carrier; the viewpoint is the soul."
So is a PPT that is beautiful but has no viewpoint a failure?
"Of course it's a failure."
Then he said something that left a deep impression on me:
"A beautiful document without a viewpoint is like a plastic flower—it looks pretty, but it has no life."
This metaphor is accurate. If you look at the PPTs now being batch-generated by AI, each one taken alone is decent, but put them together and you will find—they look identical. The same logical framework, the same transition sentences, the same "on one hand... on the other hand..."
"If everyone uses AI to express themselves, everyone's viewpoints become mediocre."
This sounds a bit absolute, but what he means is: AI will optimize expression into a "pretty good average." Standard wording, smooth logic, beautiful structure. But when everyone is optimized to roughly the same level, where is the differentiation? Where is the personality? Where is your independent thinking?
When everyone's PPT is AI-generated, a document with viewpoints and edges becomes a scarce item instead.
AI Should Be a Catalyst, Not an Answer Production Line
So what should AI document products do? They can't tell users not to use AI.
Zhou Ze'an's answer is: AI should help users discover that they already have better viewpoints, rather than directly serving up conclusions.
"AI document products cannot become production lines for standard answers."
In ChatPPT's design, there is a deliberately retained step: after AI generates the outline, let the user confirm their core viewpoints and judgments.
"Once this interactive step is removed, users shift from 'document creators' to 'document reviewers'—that is the real disaster."
This design choice is quite counterintuitive. Many competitors wish they could automate the entire process, one button to get everything done. But Zhou Ze'an believes that the division of labor between humans and AI in document creation must be dynamic: AI handles the execution-layer work of structure, materials, layout, and design; humans handle viewpoints, judgments, trade-offs, and ultimate responsibility.
"AI doesn't sign contracts, doesn't make decisions, and isn't responsible for consequences. If a PPT fails to impress investors, the person bears the responsibility."
I followed up: is it possible that AI will also be responsible for results in the future?
He thought for a moment: "The more pragmatic statement is 'joint decision-making, but humans are responsible for the final result.' This situation will not change in the short term."
Agent Is Here, and the Entry Point for Documents Is Being Rewritten
Talking about recent industry changes—Claude Code, Codex, various Agent environments starting to build in document-type Skills—Zhou Ze'an's judgment was more radical than I expected.
"This is not a change in entry points; it fundamentally changes the competitive logic."
His logic is this: in the past, users had to specifically open an AI document tool to make a PPT. Now, in an Agent environment, they can directly invoke a Skill and complete research, analysis, chart generation, and PPT output in one workflow.
"In the future, office work may no longer be 'imprisoned' by formats like PPT, Excel, and Word. Agent workflows—research, analysis, decision-making, output all in one go—documents are just the final deliverable."
Think about it: you need to make an industry research report. In the future, an Agent might automatically collect data, analyze trends, generate charts, and finally output a PPTX or PDF. Throughout the entire process, you don't need the action of "opening Office" at all.
This means the definition of document products has been reshaped—from "tools for making documents" to "providers of professional document capabilities within Agent workflows."
The path Biyou Technology has chosen is clear: it does both the human-facing experience layer (ChatPPT) and the Agent-facing capability layer (MCP Server, Skill, CLI). He said they have refactored almost all of their capabilities.
"From launching the world's first document-creation MCP Server in 2025, to ChatPPT-Skill enterprise and personal editions, to opening the CLI refactored version last month—we have run a complete cycle in the Agent ecosystem."
And he mentioned a detail: including Baidu and Kingsoft, multiple major companies have Biyou's service supporting their document capabilities behind the scenes.
A team of just over a dozen people providing the underlying document engine for major companies—this fact itself says a lot: in the AI era, a "small and specialized" vertical capability layer may have more leverage than a "large and comprehensive" platform.
"All Payment Is Based on Results"
So if an Agent itself can invoke code, templates, and knowledge bases to generate documents, where exactly is the moat of document products?
Zhou Ze'an listed three: scenario depth, data closed loop, and delivery standards.
"Pure generation capability itself is no longer a moat. What truly makes users unable to leave is the know-how you have accumulated in specific scenarios that others don't have."
For example, the PPT structure for financial industry roadshows, the expression style of educational courseware, the standardized requirements for internal enterprise reporting. These are things general models cannot do—they require repeated training, correction, and refinement in vertical scenarios.
Then he said something I think is worth pulling out separately:
"All payment is based on results."
Not paying for features, not paying for efficiency—paying for results. Whether the final document you hand over is usable and good to use is the reason users pay.
Talking about how to measure "stable delivery," he coined a new term: DLA—Document Level Agreement.
"Cloud services look at SLA (Service Level Agreement); in the future, AI office will look at DLA."
DLA contains three core metrics:
Accuracy: whether every number, conclusion, and data citation is true and credible. They have layered a traceability generation system in ChatPPT, so the source of viewpoints on every page can be traced.
Consistency: whether document style, format, and brand tone are unified.
Traceability: users need to know where each piece of content comes from before they dare to use it for reporting or decision-making.
I think the concept of "DLA" is quite interesting. Essentially, it is saying: competition in AI document products is shifting from "who generates faster" to "who delivers more stably."
General Goes Broad, Vertical Goes Deep
How does Zhou Ze'an view industry differentiation in the next three years?
He drew a dividing line: shallow applications are absorbed by platforms, while deep scenarios become larger because of the Agent ecosystem.
General large models and Agent platforms will integrate simple basic generation functions—writing an email, producing a piece of copy, generating a basic report. This is an irreversible trend.
But conversely, the popularization of Agents will make industry-level deep document scenarios more valuable. When an Agent needs to generate a professional bidding document, a compliant and brand-consistent internal report, or an industry research report with strict data requirements—general models cannot handle it.
"General goes broad, vertical goes deep."
The needs of individual users and enterprise users are also completely different.
"Individuals want 'fast'—there's a report tomorrow, just get it done today. Organizations want 'consistent'—brand standards unified, knowledge reusable, content output capacity controllable."
He mentioned that when enterprises do knowledge accumulation, the hardest part is not technology, but whether the enterprise itself can organize its internal knowledge clearly.
"Brand standards are scattered in several places, different departments' report templates are all over the place, and sales scripts have no unified standard. If the enterprise itself cannot clearly state its own document standards, no matter how strong AI is, it cannot help you."
Put simply, AI amplifies an organization's clarity. If you already know clearly what you want, AI helps you improve efficiency a hundredfold; if you are already chaotic, AI will only help you create chaos more efficiently.
From Generation Hallucination to Delivery Reality
Finally I asked him: from one-click generation to Agent workflows, where is the real opportunity for AI document products?
His answer was straightforward:
"From generation hallucination to delivery reality."
Not using AI to generate more documents, but thoroughly compressing the redundant costs and standard chaos in document output. Let every document and every viewpoint be directly delivered in a more professional way.
"Whoever can transform users' implicit expert experience into standardized document assets will hold the key to the next generation of AI office."
After hearing this sentence, I was thinking about a question: when AI truly reaches the day of "one-click delivery" of final drafts, what will happen to those positions that exist because of "making PPTs"?
Zhou Ze'an has probably thought about this question too. He did not avoid it, but his answer points in another direction—those who truly have viewpoints, judgments, and can take responsibility for results will instead become more valuable because of AI.
Because in a world where AI is available to everyone, plastic flowers are everywhere.
What is valuable is the real flower.
Selected Q&A
Q1: From 2025 to 2026, how have users' expectations for AI office changed?
Zhou Ze'an (Founder & CEO, Biyou Technology): In 2025, user behavior was essentially curiosity-driven "verification actions"—throw a task at AI and see if it can do it. In 2026 it has completely changed; users are starting to ask "can it actually get this done so I don't have to rework it." If the generated PPT requires extensive manual revision, users will give up after just one try.
Q2: What exactly are users who are willing to keep paying right now paying for?
Zhou Ze'an: Before, they paid for novelty; now they pay for certainty. Salespeople buy the probability of closing deals; office workers buy the freedom to leave work on time. Whoever can guarantee the results gets the money.
Q3: When does the transition from "trying it out" to "high-frequency use" usually happen for users?
Zhou Ze'an: Two nodes. The first is the "emotional turning point"—the first time a user feels "AI really understands me," seeing that the generated structure is more reasonable and more aligned with industry context than what they thought of themselves. The second, more important one, is "trust verification"—several consecutive deliveries of results that require almost no changes and can be directly used for reporting or sent to clients. High-frequency use corresponds not to surprise, but to stability.
Q4: Why is it said that documents are not just information organization, but viewpoint output?
Zhou Ze'an: I have been in the document space for 13 years, and I have found a common deviation in the industry—many people treat documents as information piling. But the core value of a document has never been the information itself, but the person making judgments behind that information. AI can optimize format, structure, and wording to near perfection, but it can never make that judgment for you.
Q5: How can AI document products avoid turning users into an "average"?
Zhou Ze'an: AI should play the role of a catalyst—using high-quality structure and materials to stimulate users' deep thinking, rather than directly serving up complete conclusions. In ChatPPT's design, we deliberately retain the step of letting users confirm their core viewpoints. Once this interaction is removed, users shift from "document creators" to "document reviewers"—that is the disaster.
Q6: Agent environments like Claude Code and Codex are starting to do documents—what do you think?
Zhou Ze'an: This is not a change at the entry-point level; it fundamentally changes the competitive logic. When users can complete all tasks in one Agent workflow, independent document generation tools become just one segment of the business process. Document products themselves must shift from "creation tools" to "providers of professional document capabilities within Agent workflows."
Q7: If an Agent can generate documents, where is the moat of document products?
Zhou Ze'an: Scenario depth, data closed loop, and delivery standards. Pure generation capability is no longer a moat. What truly makes users unable to leave is the know-how you have accumulated in specific scenarios that others don't have—such as the PPT structure for financial roadshows, the expression style of educational courseware, the standardized requirements for enterprise reporting. All payment is based on results.
Q8: What is the hardest part for enterprises wanting to accumulate brand standards and knowledge into an AI document system?
Zhou Ze'an: The hardest part is not technology, but the enterprise organizing its own internal knowledge clearly. Brand standards are scattered in several places, different departments' report templates are all over the place, and sales scripts have no unified standard. If the enterprise itself cannot clearly state its document standards and expression style, no matter how strong AI is, it cannot help you accumulate.
Q9: You mentioned the concept of DLA—how exactly should it be understood?
Zhou Ze'an: Cloud services look at SLA; in the future, AI office will look at DLA—Document Level Agreement. There are three core metrics: Accuracy—whether every number and conclusion is credible; Consistency—whether style, format, and brand tone are unified; Traceability—users need to know where each piece of content comes from before they dare to use it for decision-making. When these three metrics are stably achieved, AI can be said to have truly "delivered."
Q10: How will AI document products differentiate in the next three years?
Zhou Ze'an: A very clear dividing line—shallow applications are absorbed by platforms, while deep scenarios generate greater value because of the Agent ecosystem. General large models will integrate simple basic generation functions, but when an Agent needs professional bidding materials, compliance reports, or industry research reports, general models cannot handle it. General goes broad, vertical goes deep.