---
title: "AI Customer Service Clients Are Happiest With Presales Conversion, Not Aftersales"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-07-31T10:02:18+00:00"
canonical: "https://ffcap.cn/en/research/src-20260731-01html"
source: "https://uniqueresearch.substack.com/p/src-20260731-01html"
language: "en"
---

# AI Customer Service Clients Are Happiest With Presales Conversion, Not Aftersales

[![](https://substackcdn.com/image/fetch/$s_!pYhF!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4188a545-b3b9-4460-a113-beb5328206b6_2048x1152.png)](https://substackcdn.com/image/fetch/$s_!pYhF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4188a545-b3b9-4460-a113-beb5328206b6_2048x1152.png)

_Original · Unique Research / 非凡产研 · 2026-07-31 · Chinese source: https://mp.weixin.qq.com/s/QR9ecB1\_0JTFDMG8UbwZCg_

_Editor's note: This is a complete English rendition of the source roundtable transcript from the WeChat Official Account. All speaker attributions and company claims are retained as the speakers' own statements. Source images are not processed per task scope._

\---

_Original, Unique Research · 2026-07-31 18:00 · Shanghai_

AI entrepreneurship: small closed loop keeps you alive, large closed loop makes money. Doing half is the same as not doing it.

**AI Industry Observation**

"AI entrepreneurship: small closed loop keeps you alive, large closed loop makes money. Doing half is the same as not doing it."

But a brutal fact is: most AI products, users leave after using; founders hustle hard, and ultimately only earn "demo fees."

Where's the problem?

At a recent roundtable in the Unique Research Awards, five AI entrepreneurs discussed one word—closed loop. AdsGency AI founder Bolbi Liu, Yixuan Technology (宜选科技) SVP Chen Erhang (陈尔航), MulanAI founder Liu Rushan (刘如山), Shulex Marketing VP Wang Lei (王磊), and Guanghe Dongli (光盒动力) founder Zhuo An (卓安) each gave very different answers pointing to the same truth:

"AI entrepreneurship: small closed loop keeps you alive, large closed loop makes money. Doing half is the same as not doing it."

What is a small closed loop? Simply: the user inputs a need, the system outputs a complete result, no need to jump out to another tool.

Sounds basic? But many AI products don't even achieve this.

**Case 1: MulanAI—Why Is 20 Card Draws Worse Than One Correct Result?**

MulanAI founder Liu Rushan told a detail: a user first card-drew 20 times on some large-model platform without getting the desired effect, switched to MulanAI, and got it right on the first try.

"It's not luck; it's lots of engineering optimization and prompt optimization underneath."

This is a small closed loop—reducing user trial-and-error cost, making results predictable.

A more striking one: other canvas products make users click hundreds of nodes one by one; MulanAI did something they "didn't think worth advertising"—one-click run. Click once, hundreds of nodes all run automatically.

"Clients who used other platforms then came over and said 'amazing, you actually have one-click run.' We thought this was just how it should be."

Insight: the small closed loop isn't showing off—it's doing the things users should do, for them.

**Case 2: Shulex—AI Customer Service Isn't a "Q&A Bot," It's a "Conversion Engine"**

Shulex Wang Lei's story is more counterintuitive. The company does AI intelligent customer service; clients are cross-border big sellers like Anker, Aosom, and Zhiou. Initially promised 30% reply rate, 80%+ accuracy; now many clients far exceed this. But the biggest surprise came from an "accidental discovery": visiting a client in Shenzhen, asking the customer service lead "what are you most satisfied with," the answer was: "presales conversion." "Aren't you responsible for aftersales?" "Now AI does nighttime presales, and it does it well!" It turns out cross-border e-commerce customer service boundaries are expanding—from passively answering complaints, to actively承接 presales inquiries and driving conversion.

Insight: the core of a small closed loop isn't feature completeness—it's measurable, closeable value at a business node.

**Case 3: AdsGency AI—From $12 to $2, the Data Closed Loop Is the Moat**

AdsGency AI does full-funnel overseas ad buying. Founder Bolbi Liu gave an example: Pika (the AI video tool) tested multiple vendors simultaneously; AdsGency AI brought CPI (cost per install) from $12 down to $2.

How?

"The data closed loop and the entire business-chain closed loop." From linking ad accounts, backtracking data, budget management to model optimization—the whole chain is connected. On average it helps clients improve CPA (cost per acquisition) about 3x.

Insight: the end of a small closed loop isn't delivering features—it's delivering quantifiable business results.

The small closed loop solves "is the product good to use"; the large closed loop solves "is the client making money."

This is precisely the hardest part of AI entrepreneurship—no matter how great your tool is, if the client's business doesn't grow, your tool has no value.

**Yixuan Technology: B2B's Closed Loop Is Hardest in Connecting "Online" and "Offline"**

Yixuan's Chen Erhang has served foreign-trade independent websites for 16 years; he pointed out a long-standing B2B pain point:

Online promotion, offline conversion, the middle is disconnected.

"Before, foreign-trade enterprises receiving buyers had time-zone and language barriers; inquiries arriving at midnight naturally lost many opportunities. Cold-email outreach was extremely inefficient."

Now AI customer service solves time-zone and language issues; AI marketing CRM manages opportunities; more importantly—feeding offline order data back to online promotion, achieving OCI (Offline Conversion Import) closed loop.

"Two parts of the business chain that were completely disconnected before now have a chance to connect."

Insight: B2B's large closed loop isn't a tech problem—it's the organizational problem of connecting online and offline data.

**Guanghe Dongli: AI Short Drama's Closed Loop Is "Content as Product"**

Guanghe's Zhuo An's path is more complete: self-produce dramas → self-distribute → self-develop AI creation tools → enable more creators.

"Each drama continuously absorbs foreigners' payments, monthly fees, and ad tiers overseas. Output feeds tool tuning; tool tuning produces better works; better works bring better revenue."

Even thinking bigger: in the future, make 100 "Marvel-level" AI short-drama IPs to target the US market; "maybe one of our IPs will just blow up."

Insight: the highest form of the large closed loop is the tool, content, and business flywheel self-rotating, each loop amplifying the last.

MulanAI's Liu Rushan asked the audience: "Is the AI industry an internet model or a manufacturing model?"

Roughly 10% raised hands for internet model; more for manufacturing.

Her judgment is clear: more manufacturing.

If it's manufacturing logic, you can't rely on fundraising, ad buying, and traffic; losing money to圈 users desperately isn't necessarily optimal. Moreover, AI changed productivity, but did it change production relations? No. Unchanged production relations mean enterprises that originally owned industry scenarios and client resources may develop faster after applying AI. What if a startup's tech is world #1? Without enough production relations, best to hold onto a thigh and grow.

This isn't pessimism; it's clear-headedness.

AdsGency AI serves NIO, Alibaba, Aishi; MulanAI partners with a certain Wei company and a certain SaaS invisible champion; Shulex binds top sellers like Anker—essentially all borrowing boats to go overseas, grafting tech onto existing production relations.

Shulex's Wang Lei said something counterintuitive:

"AI tech advancing hugely in half a year may not help business much. The most fundamental is our understanding and practice of client value."

He gave an example of a big Shenzhen seller—clients shifted from pursuing "efficiency value" (cost reduction) to "operational value" (revenue growth). Whoever polishes products around client operational value will grow fast.

Guanghe's Zhuo An confirmed: AI short drama's lighting, character aesthetics, local cultural taboos (like in the Middle East you can't show shoulders or wine glasses)—ultimately it's not the model, it's "people"—people with production experience adjusting references, building standards, doing review.

"Before, film workers carried DSLRs and shot with foreigners, experiencing native communication. Now AI helps, but human aesthetics and native understanding build AI's standards."

"So, the truth of AI entrepreneurship is: tech thresholds are rapidly flattening; the better the model, the higher the product experience rises, but it won't disrupt the landscape; the real moat is industry know-how, client resources, and service depth."

Based on this discussion, I want to distill three actionable conclusions:

**First, pursue the small closed loop first, then the large one.** Don't try to rebuild an industry on day one. First achieve "input-output-measurable" at one business node, making users unable to leave. MulanAI's one-click run, Shulex's presales conversion, AdsGency's 3x optimization—all single-point breakthroughs.

**Second, bind with people who hold production relations.** If you're tech-origin, don't fantasize about product self-growth. Find partners with client resources and industry standing; be their AI enabler, not their replacement. Liu Rushan put it directly: "Get out there and add WeChat; everyone here is a thigh."

**Third, shift from efficiency tools to operational tools.** AI 1.0 is "save you time"; AI 2.0 is "make you more money." Shulex extended from aftersales customer service to presales conversion; Yixuan from promotion to order closed loop; Guanghe from tools to content distribution—the further the value chain extends to the back end, the stronger the irreplaceability.

Companies that raise funding with demos and trade PR for users will find it harder. Companies that truly help clients complete closed loops—small or large—will survive the coming shakeout.

As roundtable host Sang Zhuohao (桑卓豪) said: closed loops come in small and large. The small closed loop is whether a business node delivers something complete; the large closed loop is whether the revenue-growth closed loop can be better achieved through service.

"AI entrepreneurship, doing half is the same as not doing it. Once closed loop, that's when it begins."

\---

**Speakers**

-   Bolbi Liu, Founder & CEO, AdsGency AI
    
-   Chen Erhang (陈尔航), SVP, Yixuan Technology (宜选科技)
    
-   Liu Rushan (刘如山), Founder & CEO, MulanAI
    
-   Wang Lei (王磊), Marketing VP, Shulex
    
-   Zhuo An Jim (卓安), Founder, Guanghe Dongli (光盒动力)
    

**Host**

-   Sang Zhuohao (桑卓豪), Senior Director & AI Lead, Focus Media (分众传媒)
    

**Sang Zhuohao:** Please introduce your business—who you serve, what pain points you solve, how you differ from competitors. Starting with Bolbi.

**Bolbi Liu:** Hello, I'm Bobi. We're an ad operating system startup based in Silicon Valley, close to 3 years since founding end of 2023. As Mr. Sang said, some here come from ad growth or related backgrounds. Traditional ad marketing growth work chains are split—designers, media buyers, data growth devs/engineers; between them data integration speed, understanding, and decision-making are relatively independent without AI, leading to misaligned decision angles. Based on this idea I founded AdsGency AI, completing the entire ad-buying workflow from 0 to 1—this is a big difference from other products. From client targeting, purchase journey, targeted content generation, budget management, ad buying, etc. Current ad channels include Google, Facebook, LinkedIn, Twitter, and OpenAI; basically every reachable overseas platform, even CTV TV ads. Clients are mainly mid-tier enterprise clients or listed companies, especially going-overseas enterprises like NIO, Alibaba, and Aishi outside the door—all our clients.

**Chen Erhang:** Hello, I'm Chen Erhang from Yixuan Technology. The company was founded in 2009, quite long, focused on China's foreign-trade enterprises' independent-site going-global service, committed to building an independent-site-centered marketing ecosystem to help expand overseas markets. On AI, we deeply feel it—the whole business flow from early on, 2017-2018, started using AI; earliest in promotion, specifically ad buying. At the time Google pushed automated buying in China; most was still manual, a role called SEM everyone knows. We were relatively early using AI for buying, around 2018. Later as AI developed, the whole business flow used AI—independent-site building, content marketing, promotion, conversion loops, order acquisition, etc. Now under AI tech drive, every step can use AI, the whole chain AI-driven. Of course each step's AI role and tech maturity differ. AI now is pivotal in our track, not only for our service enterprises, but for Chinese export enterprises daily. That's the current picture.

**Liu Rushan:** Hello, I'm Liu Rushan, founder and CEO of MulanAI.

Now everyone sees more and more AI image, AI video products adopting canvas and workflow product forms. But when we started, this wasn't yet an industry-recognized route.

Per our continuous public search from early 2025, MulanAI is the earliest product globally to combine canvas workflow, natural-language Agent, with AI image and AI video production. You may have heard of ComfyUI. After MulanAI launched, ComfyUI's founding programmer contacted us multiple times, asking why we designed the product this way, team background, next steps. We also saw in the backend several members using ComfyUI's official email registering and trying MulanAI.

This was interesting for us. Because MulanAI at the time was a very small Chinese startup with almost no industry recognition, but the product paradigm we pioneered had already started being noticed and studied by the most professional people globally in this field.

Later, we went to a hardcore AI product competition in Silicon Valley. We had no local resources, didn't know judges. Reportedly 500+ applicants, a dozen projects into live demos; MulanAI won second place.

This result is very important for us. It proved one thing: a small team from China, relying not on brand or resources, only on product innovation, can be recognized on the world's most cutting-edge tech stage.

But MulanAI didn't start from a grand concept, but from a very specific, even painful problem.

We were already doing AI video delivery in 2024. Back then, making an AI video required switching between different models: first generate script, then storyboard, then images and video, download locally, edit, dub, subtitle, composite. Lots of repetitive, mechanical, tedious work. One piece might take weeks, even a month or two.

Through R&D personally doing AI video, we found the pain point and developed MulanAI. Work that used to take half a month or weeks, now users just describe needs in natural language, and the Agent helps build the complete workflow. Users can complete script, storyboard, image generation, video generation, dubbing, music, subtitles, editing, and compositing on one canvas; also one-click run the entire canvas, reusing a proven process repeatedly, expanding from one piece to a hundred, a thousand.

**Wang Lei:** Hello, I'm Wang Lei from Shulex. The company does AI intelligent customer service; clients are mainly large cross-border e-commerce sellers, roughly 1 billion+ RMB. In China, Anker, the biggest 3C seller, was the first client; home-furniture top two like Aosom and Zhiou are also clients; robot companies Kuma and Yuanding too. So many large cross-border sellers chose us. Founded end of 2021, we're the first AI customer service company in the field to commit to client results. Initially promised 30% reply rate, 80% accuracy; now far exceeded. Maybe our slight advantage: focused only on the very niche going-overseas cross-border field, only for large sellers. We can talk more later.

**Zhuo An:** Hello, I'm Zhang Zhuoan, CEO of Guanghe Dongli. An AI film tech company focused on film content creation and short-drama going overseas. The company previously self-built content teams, full process from writer to director to editor, won many creative TVC awards in ad circles. Based on creative characteristics, we've been developing products combining advertising and short drama, called Yangyang Menggu, an AIGC creation platform. On one hand we use it ourselves, on the other provide to peers, facilitating internal group management and efficient AI content generation, providing AI templates to creators. On short-drama going overseas, through AI short drama we solve many overseas commercialization barriers; traditional live-action takes lots of time, high cost, hard to manage. AI short drama breaks barriers, letting Chinese going-overseas short-drama companies and content-team-background AI teams enter, using small cost to leverage overseas commercial growth closed loop. Short drama itself is content merchandise, can be sold to foreigners for payment, monthly fees, ad tiers. One side has self-produced drama, self-distribution teams; the other has self-developed AI short-drama/AI-ad creation products—these two as an AI film company's two characteristics. Roughly introduced.

**Sang Zhuohao:** Today we discuss closed-loop reconstruction. I understand there are small and large closed loops. The small closed loop is whether a business node can deliver something complete, not having to switch out like Mr. Rushan said. The large closed loop involves whether C-end product complete user journeys can be completed; B-end demands higher—whether the revenue-growth closed loop can be better achieved through service. Let's talk about the small closed loop first; I want to know what small closed loops you've achieved in products. Client input, output, how constructed? If you don't mind.

**Bolbi Liu:** We're actually quite ambitious; achieved the business small closed loop quite early. For example, linking ad channel accounts like Google Ads accounts, backtracking data, doing budget management model optimization based on data—completed early. Core goal mainly targeting overseas markets, potentially replacing traditional ad agencies, that's the core business vision. The ad agency industry has many international companies 100+ years old; many workflows can be fully automated and optimized. The enterprise's main vision and ambition is to form a large closed loop, completing the entire ad-buying workflow. Delivery isn't like customer service where you can set KPI by reply; telling clients you'll definitely reach something is high risk. But on average we can achieve about 3x optimization on client CPA or last-touch metrics. For example, software AI companies value signups; pre-signup cost maybe $10 each, can do $3-4; the gap from $3 to $10 is delivery. Simple example: Pika in AI video was roughly Runway level back then; we brought CPI from $12 to $2. Three or four different vendors participated simultaneously; ours performed best. The main reason for best performance was still the data closed loop and the entire business-chain closed loop. Fairly confident in our delivery and output.

**Chen Erhang:** Different business chains have many links. Doing independent-site going-overseas for foreign-trade enterprises, the large closed loop is from the business flow's first link to the last, with the last feeding the first. The small closed loop is middle links promoting each other. For example, content marketing and online promotion promote each other. Before AI tech, efficiency was low, fully manual content, then manual promotion, promotion results then optimizing content; experience固化 in individuals, hard to replicate, low efficiency. Experience accumulated slowly; one person promotion limited, budget limited, experienced data samples limited. Pre-AI era had closed loops but much weaker effect. With AI tech, data—first historical data can accumulate and be called in real time; the platform has big data, any industry's past decade+ service historical data can serve future clients, much more efficient—this is the inter-link small closed loop, AI tech driving efficiency. Another example is the closed loop between online promotion and offline conversion. Before, B2B enterprises' two parts were disconnected—online was online, offline was offline; order conversion almost all offline, hard to connect automatically to online promotion. With AI tech we built a two-link closed loop: offline data, order tracking data feedback to online promotion, called OCI technology, started pushing this most early in B2B China in 2022. At the time AI tech application wasn't deep; now AI tech enters making data feedback to promotion fully automated; this year also used agent tech for higher efficiency. Closed loop formed helps enterprises improve efficiency. From the whole large closed loop, each link's small closed loop makes the large closed loop more efficient, very good.

**Sang Zhuohao:** Clients are partly B and partly C; how to understand closed-loop relationships?

**Liu Rushan:** In the US market, we're more Pro-C. For example, UC Berkeley has already introduced MulanAI into relevant curricula. Such users have strong learning ability and creative willingness, can help us validate product frontier capability and universality.

But in the China market, at this stage we're more B-end result delivery-focused. Because we found that purely teaching users to use tools has a long learning cycle and slow deployment.

So we now use a more pragmatic approach: parts clients temporarily can't learn, we directly help complete, first delivering the final result.

In the delivery process, we continuously consolidate client-repeated needs, easiest error-prone steps, and most human-dependent steps, then productize and automate them into new workflow templates and product capabilities. This forms a large closed loop:

B-end real needs bring delivery scenarios; problems found during delivery drive product iteration; after product capability improves, it lowers next delivery cost, and ultimately lets more clients complete production themselves.

Also many smaller product closed loops.

For example, a client told us: using a certain large-model company's product to generate images, same prompt tried 20+ times, not necessarily getting desired effect; but in MulanAI, results are much more stable.

This isn't because we invented a stronger underlying model, but because MulanAI did lots of engineering optimization underneath—this is a small closed loop: users worry about unstable generation, so we encapsulate operations that needed experts into the product, lowering user trial-and-error cost.

Another client used other canvas products then came to MulanAI; most surprised by our "one-click run." Many canvas tools require users to click one by one, execute step by step—essentially lots of manual operation. But in MulanAI, users can one-click run the entire canvas, letting dozens or even hundreds of nodes complete automatically by established logic.

We didn't even originally treat this as a particularly important promotional point, because in our view, since it's a workflow, it should auto-run. These are all closed loops MulanAI is currently building.

**Wang Lei:** Customer service closed loop is relatively simple: client info comes in, AI intent recognition, then pulls knowledge base to reply,必要时转人工. After operation, AI recognizes, summarizes, reviews all replies, next step changes knowledge base and process. The small closed loop theoretically does this. But objectively it's theoretical; with strong AI ability it can keep learning, but real closed loop relies on service. The company has delivery engineers for each client, guiding clients how to operate the closed loop, enter knowledge base, set processes. Helps client companies internally train AI trainers; problems arise then adjust. In practice, theoretically AI can learn, but actually closed loop must be truly achieved through service. Practice has been relatively effective.

**Zhuo An:** Let me explain roughly. Mainly these two-three years doing short-drama going overseas; after AI 2.0 signed annual framework, feeling big—AI quickly made content explosive growth. Behind it core is still story, native content. Based on platform having a hit-drama engine, search current hit-drama data, copy, hooks, scripts, stand on giants' shoulders to imitate. Each AI content output is consolidated in the creation platform, continuously tuned and used, better effect, becoming future reusable template. Result: short-drama chain process gets simpler and simpler; original-voice dubbing, generating whether face assets are real faces, VFX/lighting whether fits overseas native lighting, whether aesthetic—all recorded by AI short-drama generation, making AI short drama more localized. Since 2024 rated TikTok short-drama production service provider; short-drama going overseas because AI explosion makes each output feed tool tuning, tool tuning produces better works, better works bring better revenue. Self-produced drama, self-distribution, self-produced creation product three-in-one chain forms relatively fast growth; each content merchandise overseas continuously absorbs foreigners' payment or various money. Through capability replication to more OPC creators, provide people, compute, overall distribution service, production陪跑, scripts, letting more domestic creators join overseas AI short drama, enjoy distribution dividends. Also closed loop, growth. Future AI short-drama copyright appreciation, long-chain operation, IP long-line consideration, making 10 or 100 Marvel-level short-drama IPs to target the US market, maybe one IP blows up, also growth future closed loop. Both doing content and tech; each content is also merchandise, core selling point is content, through tech enablement making better content, roughly.

**Sang Zhuohao:** Follow-up question, mentioned how to judge character lighting fitting foreign aesthetics? What data?

**Zhuo An:** Look at American TV lighting effects, preferred color habits, native—like religious beliefs; Southeast Asia, Middle East can't show shoulders or wine glasses. Lighting, local asset output under prompt premise avoid generating non-native things, card-draw count increases, card loss increases, can't push out—pre-doing avoidance, more fitting native foreigner habits. Look at American TV style; Prison Break, Game of Thrones lighting differs from domestic dramas.

**Sang Zhuohao:** So still human judgment, right.

**Zhuo An:** Pre-production art makes reference, find references as multi-parameter input to the system and keep tuning. Whether US-native, Brazil-native, Southeast Asia-native, Middle East-native visual style, script style, asset style. Always doing content; content level has execution ability; before film workers carried DSLRs and shot with foreigners. Experienced native communication, localization; AI short drama also studies traditional film. Currently AI short drama human-machine combined collaborative development; AI helps; human aesthetics, understanding, native-context understanding build AI standards or tools. Roughly.

**Sang Zhuohao:** Zhuo An mentioned both large and small closed loops, mentioned IP future appreciation. Next questions from Mr. Wang Lei. The whole chain—advertising, content, B2B inquiries, sales, customer service, future repurchase—is a complete chain for enterprises especially ToB large enterprises. I don't know what level the closed loop is built at now; are there future expansion plans? If so, what's the biggest blocker?

**Wang Lei:** Still relatively focused on customer service, not that ambitious to do the whole chain well. But found client changes; completely designed products around cross-border e-commerce customer service department's contribution, goals, value results. For example, everyone understands customer service as complaints come in and you answer; cross-border e-commerce has lots of aftersales customer service, Amazon's the big head. Last week went to a big Shenzhen seller, mechanism link; asked the customer service lead what they're most satisfied with—answer was especially unexpected: most satisfied with presales conversion. Very unexpected; isn't customer service aftersales? They said now also does presales inquiry conversion. Asked how they know AI conversion; set redemption coupons, all AI conversions visible. Asked human vs. AI conversion ratio; said no humans anymore, all AI doing it. Answering the host's question: focused on cross-border customer service, large sellers' very niche market; diversification is whatever the customer service system needs now or in future, adjust. Before didn't need presales; two-three years doing aftersales; now needs presales, presales承接 service also does well. As clients grow and needs extend, extend in that direction, but don't cross fields—won't go from customer service to marketing, stay in本行.

**Sang Zhuohao:** Could it develop into a longer-chain closed loop in the future?

**Liu Rushan:** MulanAI's initial product conception was letting users not leave the canvas, completing the full process from idea, script, storyboard to image, video, dubbing, editing, compositing. This is the product-level closed loop. But from an AI company's endgame, we also need to consider: First question: is the AI industry ultimately more like internet, or more like manufacturing? I asked everyone on site; thinking AI more like internet model was about one-tenth; most think it may be closer to manufacturing model.

This question has no absolute right or wrong, but my current judgment is: AI, especially industrial AI, is closer to manufacturing. A typical internet-company logic is first pursue user scale and market share, then through continuous fundraising support growth, ultimately build platform effects. But in industrial AI, traffic and tech leadership alone aren't enough. Second question: AI changed productivity, but has it changed production relations? At least at this stage, I think not.

AI can greatly improve production efficiency, but enterprises' original client relationships, sales channels, industry resources, and trust systems don't immediately disappear because a new tech appears. Many traditional enterprises holding lots of clients and industry resources can take off once they apply AI well; while many AI startups hold tech but lack client relationships and real industrial scenarios.

Even if a startup's tech temporarily leads globally, without entering the industry chain or building client relationships, it's hard to achieve scaled growth through tech advantage alone. So MulanAI's longer future closed loop isn't just "make the product stronger"; it must grow together with companies that hold client relationships and industry resources.

**Sang Zhuohao:** How does Mr. Chen understand the closed-loop question? The large closed-loop question.

**Chen Erhang:** B2B business differs greatly from B2C. B2C can basically form a complete loop online; the business chain can all be done online, naturally closed online. B2B business chain splits into online and offline; online is mainly marketing/promotion, conversion is all offline, long-term hard to fully connect or form a large closed loop—that's the status quo. Trying to connect from online to offline; clients long-term because the two parts are hard to effectively connect, impacting efficiency—promotion efficiency, conversion efficiency. Previously insurmountable physical difficulties, like foreign-trade buyer reception time zones and language problems; previously lead conversion had big losses. Buyers coming at midnight naturally lost many opportunities; especially pre-IM era, cold-email outreach communication efficiency was very low. Now IM usage improved, but time-zone problem remains. With AI, providing AI customer service tech solves part of time-zone and language problems—AI doesn't have these two natural barriers; lead loss greatly reduced for clients. Especially AI interaction level improving fast, basically doing preliminary AI buyer communication screening—pain points enterprises couldn't solve before are very valuable. Previously lengthy negotiation in lead conversion; now using AI-built marketing CRM for clients, the whole process of lead management, also feeding back to promotion. The previously completely disconnected two parts of online and offline business chain now have a chance to connect, making B2B business form a closed loop. AI tech sees opportunities and is practicing. Building the large closed loop pushes from online to offline, business front-end steps push backward, step by step back. Find points to match in the whole client service cycle chain.

**Bolbi Liu:** Won't do any workflow outside ad promotion, no customer service, no sales. Future expansion direction may try after accumulating enough clients and corresponding user data to build an AI-era ad network, AD network, like Google network, Open AD network. Now more optimized tools, but enterprises with network capability in ad workflow have more right or profitability possibility. Possible direction but won't jump out of paid-growth performance-advertising work field.

**Sang Zhuohao:** Everyone imagine the next 3 months or half year; which AI tech breakthrough especially benefits the industry? How will the industry change?

**Bolbi Liu:** Think it's AI testing/evaluation, even including memory, agent memory, currently not fully solved on the market. Building agents is easy; many concepts say AI company or similar concept, difficulty is more in AI architecture infra; tech breakthrough helps our type of enterprise a lot.

**Chen Erhang:** Currently see the next phase is using agent tech well. Agents from second half of last year to this year have been hot, followed fast; talking about business links, all links have agent entry and application, tasted sweetness—efficiency clearly improved, clients interested, big potential. In the future period, better deploy agent tech in business flow, truly self-training process, training efficiency, development speed improved, push business forward. Mainly focus on this.

**Sang Zhuohao:** Understood. Mr. Rushan, seedance 2.5 comes out, seedance 3 comes out, will it completely disrupt the industry?

**Liu Rushan:** At least for MulanAI it's not disruption; it will amplify our value instead.

Models solve "can one shot be generated better"; MulanAI solves "how to stably complete a complete video task."

This is a bit like today's AI coding. Underlying large models get stronger, but everyone still needs Cursor, Claude Code, Codex. Because what users need isn't calling a model once, but understanding the task, managing context, breaking down steps, calling tools, finding errors, and ultimately completing the work. What MulanAI undertakes in the AI video field is a similar role. This is why Claude Code's #1 person Liu Xiaopai and ByteDance's AGENT product managers all consider MulanAI the Cursor of video. In short, model companies provide increasingly strong engines; what MulanAI wants to make is the whole car, even a continuously running intelligent factory. A production system that completes delivery; rather, shallow applications that haven't formed workflows, no client scenarios, no result closed loops.

We actually very much look forward to domestic large models continuing to improve. Companies like DeepSeek driving model capability improvement and cost reduction will benefit the entire application layer. Because MulanAI's intelligence level isn't closed; it evolves as underlying models evolve.

**Wang Lei:** Viewpoint may differ. AI tech advancing hugely in half a year may not help business much. But during the process, deep understanding of client value will help greatly. A big Shenzhen seller on presales conversion saw the market change—cross-border clients from efficiency value to operational value. How to polish products around client operational value, think it through and produce results, will have very fast development. Tech will accelerate, but fundamentally it's the client; now tech is actually good enough; need to further dig client needs in co-creation with clients, from cost end to operational end.

**Zhuo An:** Mainly say one point. Film and short-drama production has lots of derivative value. Now making a product called Yingying Zhiwang, film product-placement website; each drama's reveal points are frame-extracted and seamlessly placed brands. Recently Stephen Chow's Shaolin Soccer placed Wanglaoji; before, boss dramas placed Rolls-Royce, appearing immediately. Can it place Xiaomi cars, Huawei cars? R&D product encountered: after placement frame-extraction fusion it's a bit突兀 from the original, connected not original; frame extraction is definitely a bit突兀. Future C-Dance 2.5 or 3 tech better; each placement more seamlessly generated with original lighting. Provide to more brand clients, creator service; each drama has many points placing various things, bringing creators more value, increased revenue, increased brand exposure. Believe future better tech, also in R&D, stay tuned.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260731-01html
On-site reading page: https://ffcap.cn/en/research/src-20260731-01html