Original · Unique Research · 2026-07-09
Editor's note: This is the original Chinese author's interview with Zhao Maojun and its framing. This English rendition retains the full text in source order, including all 15 Q&A items. The interviewee's product claims, market figures, and forecasts are his self-reports, attributed to him and not independently verified. Person, company, and product names are preserved as source attributions.
AI Industry Observation
Exclusive interview with Zhao Maojun, CEO of Guangchen Tongxing (光尘同行): making AI into a "fiduciary" the consumer world has never seen.
The Smarter the Algorithm, the Higher the Return Rate?
Ten years ago, women's wear e-commerce return rates were under 15%. Offline stores were even lower then, 5% to 10%; try-then-buy was taken for granted. Ten years later, the industry-average return rate has soared to 60%, and in livestream rooms 70% to 90% is normal; merchants joke that this is "nine returns out of ten buys."
In 2024 alone, over 40 top women's wear online stores closed. On the 2.33-trillion-yuan online clothing market with a conservative 20-30% return rate, five to six hundred billion yuan of clothing spins empty every year between "buy, return, buy again" — and over half the return reasons are achingly simple: wrong size, wrong fit on the body, item not as shown.
Here's the puzzle. Over the past decade, recommendation algorithms got smarter and smarter. Personalized recommendation, collaborative filtering, deep learning, large models — the tech stack kept upgrading, yet returns rose two to three times. Why?
Plainly, the whole stack optimizes "make you buy," and not one line of code optimizes "make you buy right."
You scroll to a livestream, the host shouts "three, two, one, link is up," you impulse-buy — that's "make you buy." But you put it on and find it widens your hips, the color is off, it's nothing like you imagined — that's "bought wrong." Half the cost of buying wrong is borne by the merchant, half by the user, none by the platform. So this pain wasn't unseen; it was exempted by the business model.
China's internet made "browsing" world number one, but left "getting it done" blank. You want to try that dress — the process is: open the app, search keywords, scroll reviews, watch livestream, compare prices, order, wait for delivery, unbox and try, doesn't fit, request return, wait for pickup. Four or five apps, dozens of jumps, dragged over two or three days.
A long chain was never because technology couldn't do it; it's because the business model didn't want to. Every extra minute a user stays in the chain, the platform earns a cent more — "browsing" sells ads; "getting it done" doesn't. So every generation of tech upgrade optimized "make you buy," not "make you buy right."
Until someone decided: I'll do this. That someone is Zhao Maojun (赵茂俊), founder of Guangchen Tongxing. Many haven't heard of the company, but its product "Dajie" (搭介) — an AI life-and-consumption Agent — is quietly rewriting the rules. The team is interesting too: Zhao is a serial entrepreneur; CTO Lyu Zhixing (吕志兴) has Alibaba Cloud and Microsoft background; fashion partner Zhen Ziqi (甄子祺) is from Parsons and did celebrity wardrobe custom. Three people from completely different worlds came together to do something the consumer world has never done.
"In one sentence: the thing is done."
After the Words "I Want to Try," Someone Handles What Follows
Imagine an ordinary scene. You scroll to a dress, you're smitten, but unsure. In the past, you'd reserve, go out, arrive at the store, queue, try, hesitate, go home and think again. Now you just say "I want to try" — the store delivers this dress, with two matched outfits, right to your home. Keep what fits; take away what doesn't on the spot. You didn't move; the thing got done.
This is what Dajie does. Not another e-commerce app, not another recommendation algorithm, but connecting the stores and services around you one by one into your "one-sentence service." Within three kilometers of your home are hundreds of stores and over a thousand services, yet you use no more than five. It's not that supply doesn't exist; it's that supply is "invisible" to you. Between you and the shop downstairs isn't 300 meters but an entry. Dajie wants to be that entry.
But here's a question: life services are so huge — why start with outfits? Because outfit pain is sharpest. On the supply side, small-order quick response pushes new-arrival speed to the limit; top fast fashion launches hundreds of thousands of styles a year, while one person's clothing decisions in a year number only a few hundred — the scissors gap between supply explosion and human decision bandwidth can only be closed by AI. Goods consumption has peaked; the nearly 20-trillion service consumption is the growth engine. Search solved "can find," never "choose right" — and the first dividend of the AI entry naturally belongs to the most consultation-dependent, non-standard categories.
Search for an iPhone, compare specs and you know which to buy — standard goods need no consultation. But buying a coat, your shoulder width, skin tone, body shape, occasion, budget, even today's mood all affect the decision. These variables are too soft, too personal, too dependent on person-to-person inquiry; traditional internet simply can't hold them.
The direction was designed, but the path was pushed by users. Dajie's team wanted life services from day one, but didn't expect users to run ahead. Someone photographed a wedding invitation asking what to wear, then asked how to arrange around the hotel; someone asked about commuting wear and snapped a business-trip suitcase; someone photographed her wardrobe asking how to match, then chatted into what to wear for weekend parent-child activities, how to prepare for parents staying over. Demand spills past scenario boundaries; users are always more honest than the roadmap.
E-commerce Worked Twenty Years and Never "Saw" You
E-commerce worked twenty years and accumulated all "product data," no "human data." The platform knows every garment's clicks, sales, return rate, but doesn't know any buyer's shoulder width, skin tone, or what's already in her wardrobe. So-called personalized recommendation is collaborative filtering underneath — "what people similar to you bought" — essentially statistics covering up ignorance of the individual.
Multimodal AI lets machines "see a specific person" for the first time. Language is a lossy compression format humans invented. It's hard to describe your face shape, skin tone, and build in words, but one photo says it all — and the photo even holds information you can't put into words: warm or cool skin undertone, shoulder-neck line, style habits you haven't noticed. Photos don't lie. In words, people overstate budget, inflate size, beautify preferences; a photo hands over the real you. The upper bound of judgment quality depends on input honesty — multimodal isn't just more convenient input, it's more honest input.
But "seeing" is only step one; "understanding" is the hard part. Large models are born with mediocre taste. This isn't my conclusion; it's mechanically destined — the model's training objective is "most probable," while the essence of aesthetics is "positioned deviation." The average of all web images is exactly the look no one wants to wear out. What to do? Two things. First, elevate "soft" variables to decision variables: occasion, mood, budget, identity expression, in Dajie's system aren't tags but inputs at the same level as body data. Second, inject "above-average" human standards into the model. AI handles scale; humans handle taste. Outfit data that looks good and is validated by sale can't be bought across the web; it's accumulated order by order. The more accumulated, the more taste compounds.
Here a buyer-shop example is telling. A boutique women's wear buyer shop had years of loyal customers in its private domain. After joining Dajie, when new arrivals came, loyal customers received not model images but renders of this new item on themselves, with a judgment — why it fits, how to match, what occasion. Want it, order in one sentence; unsure, reserve doorstep try-on, take away what doesn't fit on the spot. Result: the store's monthly transaction grew from about 100,000 to about 140,000, up 40% — without spending a cent on ads, without adding a single follower; all growth came from loyal customers already lying in WeChat.
Behind this is a key mechanism: explainability. Online apparel's average conversion is only about 1.4% — a hundred people looked, ninety-something didn't buy. Not because they didn't want it, but because they didn't dare decide. AI says "this dress suits you," and no one dares believe; AI says "your shoulder line is narrow; this dropped-shoulder cut will correct your proportions," and the user dares order. Explanation is trust. The bigger the thing AI wants to do for a person, the clearer it must first speak.
Ultimately, Dajie doesn't just "let you buy," but on the basis of "seeing you," turns "buying right" from mysticism into explainable systems engineering. When AI can for the first time honestly see a specific person, build trust through explanation, and make each order smarter than the last through stacking scale and taste — the absurd loop of "nine returns out of ten buys" can finally break. And the starting point of breaking it is one ordinary sentence: "I want to try." Fine — after that sentence, someone handles what follows.
The Consumer "Fiduciary": a Role That Never Existed
"'Who vouches for the buyer' is commerce's oldest job." Four generations of trust intermediaries, each deteriorating. Acquaintance society relied on shopkeepers and regulars. Department-store era relied on buyers — the store sent people to Europe to select goods, and what came back was your taste. Brand era relied on logos — a brand is essentially pre-packaged trust; you recognize the swoosh and skip thinking. Platform era relied on ratings and reviews — then ratings were polluted by fake orders, and livestream rooms hung "family" on everyone's lips. Plainly, borrowing acquaintances' language, overdrawing acquaintances' trust. "Nine returns out of ten buys" is the bill for that overdraft.
Why does each generation rot? Whoever pays, it behaves like them. Search is paid by advertisers, so it serves the highest bidder; shoppers take merchant commissions, their income comes from you buying, not you buying right; consultants charge fees but don't bear outcomes. No exceptions.
So Dajie wants a fifth kind — a role that never existed in consumption: the fiduciary. Doctors can't prescribe you medicine for kickbacks; fund managers can't use your money to buy their own positions. Medicine and finance long ago invented "fiduciary duty," yet in daily consumption no rule forbids a shopper from calling the highest-commission item the best for you. When everyone has their own Agent, "does it owe me fiduciary duty" becomes the user's first question in choosing AI. Dajie wants to build the answer ahead of time.
A fiduciary isn't a stronger shopper; it's another species. The shopper "recommends" for you; the fiduciary "bears" for you. A one-character difference, an entirely different essence. The shopper's bottom line is "not illegal"; the fiduciary's bottom line is "not betrayal."
Landing this concept in the product are two hard things. One is the dynamic UI engine. Domestic apps monitored by the Ministry of Industry and Information Technology peaked at over four million, now down to about 2.6 million — an ordinary person has dozens installed on their phone but opens no more than ten daily. Several million capabilities compete for a sub-ten-seat entry; that's the App era's dead knot. The dynamic UI engine attacks this knot. In the past, for a service to be used, it first had to build an app, then buy traffic and beg for downloads — two mountains, interface and customer acquisition, crushed countless good services. In Dajie, the interface is generated live by AI: you want to try on, it grows a try-on interface; you want to return, it grows a return interface. From "a thousand people, a thousand faces" to "a thousand tasks, a thousand faces" — services no longer need their own interface, only their own capability.
The other is the relay station, solving the "dialect" problem. Hundreds of merchants' and service providers' systems each speak their own — inventory, payment, fulfillment logic all differ; the relay station translates them into one language AI understands, then uniformly manages permissions, risk control, billing, and settlement. Metaphor: the App era was everyone building their own generator; Dajie is building the power grid.
"After this road is built, the first to run are 'old customers,' not 'new customers.'" That buyer shop whose monthly transaction went from 100,000 to 140,000 — without spending a cent on ads, all growth came from loyal customers already lying in WeChat. What's validated isn't acquisition power but the power to revitalize the existing base. Those ignored people lying in private domains were actually waiting for a trustworthy "person" to vouch for them. AI isn't grabbing traffic; it's repairing an overdrawn trust relationship.
From Selling Tools to Selling Results: We Took the "Evil Button" Apart
With the role set, the business model follows clearly. Dajie's pricing is direct: 8-12% on attributed transactions, nothing on returns. "Why not charge subscriptions first? Many AI products today charge an 'anxiety tax' — selling per-seat subscriptions, whether you use it or not." Moving from selling tools to selling results is the pricing migration of the whole AI industry; Dajie just jumped first in the consumption scenario.
Nearly every two-sided platform has walked the same arc: court users early, tilt to the B side after scale, more ads, recommendations less and less in your favor. It's not that some company turned bad; the model is destined — as long as the platform sells user attention, the user sooner or later turns from customer to goods. Dajie's answer isn't a promise of attitude but structural design.
"The most credible promise isn't 'we won't be evil,' but 'we took the evil button apart.'"
Concretely, three hard bars. First, no impression inventory — no selling ad slots; income only takes a cut on transaction results, nothing on returns. This means Dajie can never show a scene where whoever pays more ranks higher. Second, a brand's dedicated domain holds only that brand, no competitors — in your Nike dedicated domain, you never see Adidas. Third, AI keeps the dissuasion right — when AI judges something doesn't suit you, it directly says "don't buy." Each wrong order dissuaded saves the brand a loss. You may lose one order, but in exchange the user asks you again next time.
For the ad model, user trust is a one-time consumable; burn it and acquire another batch. For Dajie, trust is the only means of production — the thicker the profile, the more accurate the dissuasion, the more valuable cross-scenario reuse; any one trust-harm is asset impairment. "Switching cost trends to zero in the Agent era; whoever betrays the user, the user replaces it in one sentence — 'betraying users' has for the first time turned from a profit lever into commercial suicide."
North Star: Weekly Effective Decisions Per Capita
"Metrics aren't measuring tools, but shaping tools — whatever you assess, the product grows into." The PV era bred clickbait; the "DAU × time" era bred infinite scroll, autoplay, and algorithmic feeding. Every ill of the information feed is that metric's child. So Dajie watches one number: weekly effective decisions per capita — users come with a real question, get a judgment, and take the next action: order, save, reserve, or come again. Not DAU, not token consumption.
"AI that tells stories with DAU treats itself as media; AI that tells stories with token consumption treats electricity bills as revenue — we've burned lots of tokens too, but I know clearly: that's a cost sheet, not a report card. AI applications' next-generation report card has only one form: how many things it got done for people."
We Are Racing the Schedule
What Dajie is doing may have a narrower time window than imagined. The founder sees AI in "three steps": first learn to speak (Q&A), then learn to do (handle people's affairs in the digital world), finally be able to act (robots entering the physical world). We're now in the shift from step one to step two. Technically, intent understanding and service calling are good enough; what really sticks is the psychological threshold — "I don't dare let AI spend money for me." Much like back then "I don't dare bind my bank card to my phone." "Give a coordinate: we're roughly at mobile payments' 2013." What happened to that threshold? Mobile payments took about five years from "don't dare bind the card" to over 80% of people unable to live without it. Psychological thresholds collapse overnight — the first group tastes "the thing got done," and the rest follow as a group.
So on "reversibility," Dajie is especially strict. "Agent trust isn't stacked by accuracy; it's designed by reversibility. Reversible automation deserves to spend money for people — without reversibility, 99% accuracy shouldn't touch the user's wallet." The order of entry scenes shows this caution. Three screening criteria: high frequency, visually verifiable results, low error cost. Image consumption first, then gifts, home soft furnishings, local dining and travel. The essence is "trust-limit management" — AI trust is like a credit card; the limit must start small and high-frequency; swipe a big amount up front and the card freezes.
For consumer brands, what truly matters isn't "new traffic" but "the way of being selected has changed." China's hundreds of billions in yearly internet ads are essentially paying for "being seen"; in the Agent era, budget migrates to paying for "being selected." "I'll predict: within five years, a new line item appears on brand budgets called 'Agent usability' — letting AI find you, call you, verify you." The goal isn't to let people see more, but to let people stop looking — handing the burden of choice to machines, returning time to life.
Five years ago you carried a wallet going out. Today you've almost forgotten what a wallet is. Five years later, you may forget "browsing" too — not consuming less, but no longer worrying about consumption. Once the thing is done, AI should retreat to where it can't be seen.
More Conversation Details
Q1: Without an official intro, how would you explain Dajie to someone who doesn't know what you do?
Zhao Maojun: In one sentence: things in life get done when the user says one sentence. For example, you want to try a newly arrived dress; the old flow was reserve, go out, arrive, queue, try, hesitate, go home and think; now you just say "I want to try," and the store delivers this dress with two matched outfits right to your home, keep what fits, take away what doesn't on the spot. You didn't move; the thing got done. Dajie connects the stores, services, even long-tail apps around the user one by one into the user's "one-sentence service."
Q2: What pain does "getting life tasks done in one sentence" really solve?
Zhao Maojun: Users never lack information; they lack merging a string of actions into one delivery. Want to put together an interview outfit today; you may have to go to content platforms for guides, e-commerce for price comparison, stores for try-on, return courier booking — four or five apps, dozens of jumps, two or three days. China's internet made "browsing" world number one but left "getting it done" blank. Sharper, this blank always existed not because technology couldn't but because the business model didn't want to. "Browsing" sells ads; "getting it done" doesn't. Every extra minute in the chain, the platform earns a cent more. So shrinking the chain to one step is fated not for incumbents in the chain but for a new species outside it. Dajie earns money from "getting it done," not "staying"; the revenue structure lets it dare cut the chain.
Q3: Why start with outfits, fashion, and image consumption?
Zhao Maojun: Because apparel is a textbook "recommendation gets smarter but results get worse" scene. Ten years ago, women's wear e-commerce returns were under 15%, offline stores only 5% to 10%; today industry-average returns can hit 60%, livestream even 70% to 90%, merchants call it "nine returns out of ten." In 2024 alone, over 40 top women's wear stores closed. On the 2.33-trillion online clothing market with a conservative 20-30% return rate, five to six hundred billion yuan of clothing spins empty between buy-return-buy. Over half the return reasons are simple: wrong size, wrong fit, item not as shown. A truth no one says aloud: over the past decade recommendation algorithms got smarter, but returns rose two to three times, because the whole stack optimizes "make you buy," not truly "make you buy right." Big firms also have AI try-on, but usually a feature on their own shelf, not a service loop for tens of thousands of small and mid stores. The hard part is connecting tens of thousands of stores' product libraries, inventory, new arrivals, and fulfillment systems, and managing doorstep delivery, returns, and result attribution. These are long-tail, trivial, high-front-investment dirty jobs. Big firms lack not technology but the reason to bend down and do it.
Q4: How did it go from "AI outfit decisions" to a "life-services Agent"?
Zhao Maojun: The direction was designed; the path grew. Dajie wanted life services from day one, but didn't expect users to push it. Someone photographed a wedding invitation asking what to wear, then asked how to arrange around the hotel; someone asked about commuting wear then snapped a business-trip suitcase for AI to judge matching. Demand spills past scenario boundaries; users are always more honest than the roadmap.
Q5: Why is multimodal so important in life-consumption scenarios?
Zhao Maojun: Because e-commerce worked twenty years and accumulated all "product data," not "human data." The platform knows every garment's clicks, sales, returns, but not the buyer's shoulder width, skin tone, build, or what's in her wardrobe. So-called personalized recommendation is still collaborative filtering underneath — "what people similar to you bought." Essentially statistics covering ignorance of the individual. Multimodal lets machines truly "see a specific person" for the first time. Language is lossy compression; it's hard to describe your face, skin tone, build in words, but one photo basically says it all. The photo even holds info users can't say: skin undertone, shoulder-neck line, style habits. Making users type descriptions was making users work for AI; multimodal flips the order: the human just raises the phone, understanding is AI's job. Another key point: photos don't lie. In words people overstate budget, inflate size, beautify preferences; the photo hands over the real you. The upper bound of judgment depends on input honesty. Multimodal isn't just more convenient input, it's more honest input.
Q6: Consumption decisions involve taste, emotion, occasion, budget, identity expression; how does AI understand these "soft needs"?
Zhao Maojun: Large models are born with mediocre taste. It's mechanical: the training objective is "most probable," while taste's essence is "positioned deviation." The average of all web images is often the look no one wants to wear out. So soft needs are solved two ways. First, elevate "soft variables" to decision variables. Occasion, emotion, budget, identity expression, in Dajie's system aren't tags but inputs at the same level as body data. The same black dress is two entirely different problems "meeting a client" vs "meeting an ex." Can't solve this difference and there's no understanding consumption. Second, inject "above-average" human standards into the model. The last point settles into a data moat. What's truly valuable isn "good-looking" data but "good-looking and sale-validated" data. AI handles scale, humans handle taste, data makes taste compound.
Q7: You mentioned Dajie stresses "explainability." Why is explanation so important?
Zhao Maojun: Explanation is the new trust currency. Online apparel's average conversion is only about 1.4%; a hundred looked, ninety-something didn't buy. Often not because they don't want it but because they don't dare decide. E-commerce twenty years pushed "information display" to the extreme; detail pages get longer and longer, but never solved "information adjudication": a hundred specs laid out, none tells you "what this means to you." AI says "this dress suits you," and the user may not dare believe; but AI says "your shoulder line is narrow; this dropped-shoulder cut will correct your proportions," and the user acts. Explanation isn't a paper metric; it's conversion itself. In principle, explanation depth follows amount and irreversibility: 9.9-yuan socks need no explanation; 8,000-yuan pearls must explain origin, nacre layer, and why this one. In the Agent era, explanation also becomes the basis for authorization. Why does the user authorize AI to spend for them? Because every time it explains "why." The bigger the thing AI wants to do for a person, the clearer it must first speak.
Q8: Is a good life-consumption Agent more like a search engine, a shopper, a life assistant, or a personal advisor?
Zhao Maojun: None. Dajie wants the fifth role that never existed in consumption: the fiduciary. "Who vouches for the buyer" is commerce's oldest job. Acquaintance society relied on shopkeepers and regulars; department stores on buyers; brands on logos; platforms on ratings and reviews. Later ratings were polluted by fake orders, livestream rooms hung "family" on everyone, borrowing acquaintances' language, overdraws acquaintances' trust. "Nine returns out of ten buys" is that overdraft bill. Why does every trust intermediary deteriorate? In one sentence: whoever pays, it behaves like them. Search is paid by advertisers, so it serves the highest bidder; shoppers take merchant commissions, income from you buying, not you buying right; consultants charge fees but don't bear outcomes. None structurally must be responsible for the user's result. Medicine and finance long invented "fiduciary duty": doctors can't prescribe for kickbacks; fund managers can't use your money for their own positions. But in daily consumption, no rule forbids a shopper calling the highest-commission item the best for you. Dajie wants to answer ahead a question: when everyone has their own Agent, who is it responsible to? The answer should be that it must be responsible for the user's result.
Q9: You mentioned a "dynamic UI engine" and "AI service relay station"; explain in plain words?
Zhao Maojun: Both, essentially, are building roads for the "long tail." First a reality: domestic apps monitored by MIIT peaked at over four million, now down to about 2.6 million. An ordinary person has dozens installed but opens no more than ten daily. Several million capabilities fight for a sub-ten-seat entry; that's the App era's dead knot. The dynamic UI engine solves the "interface" problem. In the past, for a service to be used, it first had to build an app, then buy traffic and beg downloads. Two mountains, interface and acquisition, crushed countless good services. In Dajie, the interface is generated live by AI: you want to try on, it grows a try-on interface; you want to return, it grows a return interface. From "a thousand people, a thousand faces" to "a thousand tasks, a thousand faces," services need no own interface, only own capability. The AI service relay station solves the "dialect" problem. Hundreds of merchants' and providers' systems each speak their own; the relay station translates them into one language AI understands, then uniformly manages permissions, risk control, billing, and settlement. Metaphor: the App era was everyone building their own generator; Dajie wants to build the grid.
Q10: Why choose Alibaba's domestic Qwen models?
Zhao Maojun: Four reasons all hold, but with order. First is multimodal vision. Qwen's VL series is first-tier at "understanding a life photo." Second is Chinese consumption context. The cultural meaning behind words like "slimming," "commute," "mom vibe" — domestic models have home-court advantage. Third is cost. Dajie consumed over 100 million tokens in its first two months online; the inference-cost curve directly decides whether consumer-grade AI dares open to free users. This is a life-or-death line, not an optimization item. Fourth is compliance. The underlying model has passed CAC filing; for a consumer product this is a hard threshold, not a bonus. Choosing a model is like choosing a partner, not who scores highest today, but who keeps evolving with you in three years.
Q11: If you later cover dining, travel, local services, gift recommendations, image management, and more, what's hardest?
Zhao Maojun: Hardest isn't the model but the supply side. Model capability is buyable; the service network isn't. Dajie calculated that in shelf-style apparel across all categories alone, there are 30,000 qualifying brands and stores nationally. To make their product libraries, inventory, and fulfillment systems "speak one language," so each service result can be returned, attributed, and reused by the next decision, there's no shortcut but negotiating one by one and fixing interface by interface. Especially "attribution." Who actually brought a transaction is e-commerce's blackest box. Dajie requiring every order attributable means actively choosing the hardest engineering. But only with attribution does "charging by result" hold. Task decomposition and experience consistency have engineering solutions; only service onboarding is brute-force work. Dajie bets that few are willing to do brute-force work.
Q12: How do you see the judgment that "AI doesn't replace apps but re-orchestrates app capabilities"?
Zhao Maojun: Apps won't die short-term, but the app "interface" will. Domestic app count fell from a peak of over four million to about 2.6 million in a few years. As a form, app count has peaked. When demand is executed by an Agent, "opening an app" increasingly looks like today's "going to the service hall": the capability is still behind the counter, but no one queues anymore. To every app this is the same cruel question: when users no longer open you, what's left? The answer can only be the service capability itself. Whoever first dissembles itself into callable capability has a seat in the next era.
Q13: In which model will Dajie's business model break through first?
Zhao Maojun: Transaction commission breaks through first. Dajie has validated taking 8% to 12% on attributed transactions, nothing on returns. A store with about 100,000 monthly attributed transactions gives Dajie under 10,000. Why not subscriptions first? Because many AI products today charge an "anxiety tax": per-seat subscriptions, whether used or not. But overseas AI customer service charging per "solved problem" has appeared. From selling tools to selling results is the whole AI industry's pricing migration; Dajie just jumped first in consumption. Currently, Dajie's key push is brand-private-domain AI try-on transactions: users complete the full loop of "self render — lead capture — order — try-on — return/exchange" within the brand's private domain, and Dajie takes commission by transaction result. First pilots focus on community boutique stores and buyer shops, because they natively have same-city same-day home fulfillment. Since launch, Dajie has zero-spend accrued users over 10,000, with 30 first brand pilots in progress. The second category being expanded is high-consultation-density jewelry, especially pearls, because SKUs are complex, non-standard attributes are strong, and pre-sale naturally needs multi-round professional Q&A.
Q14: Serving C-end users and B-end providers at once, how do you balance both sides?
Zhao Maojun: History's answer to "how to balance both sides" is: no one ever really did. Nearly every two-sided platform walked the same arc: court users early, tilt to B after scale. More ads, recommendations less in your favor. It's not that some company suddenly turned bad; the model is destined: as long as the platform sells user attention, users sooner or later turn from customer to goods. So Dajie doesn't make attitude promises but structural design. The most credible promise isn't "we won't be evil" but "we took the evil button apart." First, no impression inventory. Dajie doesn't sell ad slots; income only takes attributed transaction commission, nothing on returns. Sell a wrong order and the brand loses, Dajie loses too. Second, a brand's dedicated domain pushes only that brand, no competitors. Dajie architecturally has no ability to "reallocate user attention to higher bidders." Third, AI keeps the dissuasion right. Brands may frown the first time, but after seeing return rates and repeat purchases change, they'll understand the value. Over half of industry returns are size/fit issues; one return loses over ten yuan on courier, shipping insurance, packaging alone. Each wrong order dissuaded saves the brand a loss once. The deeper difference is on the balance sheet. For the ad model, user trust is a consumable; burn it and replace the batch. For Dajie, trust is the only means of production. The thicker the profile, the more accurate the dissuasion, the more valuable cross-scenario reuse; any one trust-harm is asset impairment. Switching cost trends to zero in the Agent era; whoever betrays the user, the user replaces it in one sentence. "Betraying users" has for the first time turned from a profit lever into commercial suicide.
Q15: What's Dajie's most important North Star, and what does it say about how you understand the future of AI consumer apps?
Zhao Maojun: Dajie watches one number: weekly effective decisions per capita. Users come with a real question, get a judgment, and take the next action: order, save, reserve, or come again. That's an effective decision. Metrics aren't measuring tools but shaping tools. The PV era bred clickbait; the "DAU × time" era bred infinite scroll, autoplay, and algorithmic feeding. Every ill of the feed is that metric's child. The internet didn't suddenly turn bad; it was trained bad by its own North Star. AI telling stories with DAU treats itself as media; AI telling stories with token consumption treats electricity bills as revenue. AI applications' next-generation report card has only one form: how many things it got done for people. So Zhao judges consumer AI is shifting from content production to decision agency. AI first learns to speak, then to do, finally to act. We're now in the shift from step one to step two. Technically, intent understanding and service calling are enough; what truly sticks is the psychological threshold — users still don't dare let AI spend for them. Much like mobile payments' 2013. Back then people didn't dare bind cards either, but over about five years mobile payments moved from "don't dare bind" to over 80% unable to live without it. This time the window may be shorter, because models, payments, and delivery are all ready infrastructure; only trust is missing. And trust isn't stacked by accuracy; it's designed by reversibility. Dajie starts from the simplest link: returns and exchanges. Every step explainable, result verifiable, action reversible. Reversible automation deserves to spend money for people. The life scenarios best reconstructed first by AI Agents must also meet three criteria: high frequency, visually verifiable results, low error cost. In this order, first image consumption, including apparel, beauty, accessories; then gifts, home soft furnishings, local dining and travel. The essence is "trust-limit management." AI trust is like a credit card; the limit must start small and high-frequency. Letting AI pick a house or diagnose illness up front is overdraw of the whole industry's trust. For brands, providers, and vertical apps, the AI life-service entry isn't just new traffic but "the way of being selected has changed." Budget used to pay for "being seen"; the Agent era migrates to paying for "being selected." Fulfillment quality, return rates, interface response speed — these back-end metrics become new ad slots. Within five years, a new line item may appear on brand budgets called "Agent usability": letting AI find you, call you, verify you. That day, half of ad spend flows to interfaces and fulfillment, not traffic. What really changes life consumption isn't recommending more products but letting people stop looking — handing the burden of choice to machines, returning time to life. Once the thing is done, AI should retreat to where it can't be seen.