Original · Unique Research · 2026-06-05
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, full panel, and all named speaking turns. Product-selection, content-production, fulfillment, customer, revenue, efficiency and market figures are source or speaker claims, not independently audited findings. Company, personal and work titles are transliterated where official English forms remain unverified.
AI Industry Observation
A Seller of Customer-Service AI Tells Sellers on Stage: Don't Worry About My Product First
AI provides general-purpose capabilities; the moat comes from understanding specific customers and from data assets
"
Get product selection and marketing right first—we come last.
Chen Guang sells AI customer service. QuickCEP serves several hundred cross-border brands. His advice from the stage was: get product selection and marketing right first—we come last.
He explained it with a 2024 experience helping a company build a visual agent: at the time, using a certain domestic model, "the效果 was just terrible." The same thing is trivial today. Do it too early, and you test an immature technology, burning not just budget but also an opportunity that could have waited for a better moment.
This was a cross-border e-commerce roundtable. Three guests represented product selection, content production and fulfillment respectively, all building products with AI. But on when to use AI and at which node to apply it, the three gave different calibrations.
Product Selection: 13 Weeks Compressed to 4 Hours, but Success Rate Matters More Than Speed
LynxAI founder Zhang Keyi gave a figure: in the past, the product-selection cycle for fast-moving consumer goods and consumer electronics was generally 13 weeks; now the data-acquisition and decision-makingaspect can be compressed to 4 hours.
But he immediately added: it is not that the cycle has shortened, but that the success rate is rising.
They have 7 clients they have worked with for over 12 months. Each has no fewer than 3 products for which AI provided leads and assisted product innovation, after which the products exceeded their project-launch sales targets. The cycle has not shortened dramatically, but the hit rate is climbing, and because the logic leaves a trail and data accumulates in structured form, this improvement is reviewable and cumulative.
He also offered a three-layer framework for evaluating AI recommendations: the bottom layer is whether there is structured factual data; the middle layer is whether AI's reasoning logic is clear, turning from a black box into a white box; the top layer is what judgment AI made based on that logic. Only when all three layers are visible can you judge whether AI's recommendation is worth adopting and where human intervention is needed.
Another principle worth recording is what he called the "80/20": even if you discover a deeply unmet user need, the probability is that you will first do 20% innovation and 80% inheritance. A product that is too new has uncertain market acceptance; over-innovation imposes extra education costs on both channels and users. The first two generations of products get the market working, then the innovation ratio increases later. Win rate comes first.
Content: 60 Points Is Already Free—What Are You Competing On?
Linkfox product lead Henk offered a judgment: the production of cheap素材 has disappeared. Anyone can produce a 60-point image or video.
What does this mean? Sellers are no longer competing on who can produce content, but on whose content more accurately hits the target user.
He gave three sources of differentiation.
First is the depth of customer understanding. For the same product, one seller's customers buy it as a gift, while another's use it in a household daily-use scenario. Different target scenarios mean different image angles, copy emphasis and pain-point excavation—the content is already拉开 distance. A seller who only pushes a generic product positioning ends up in a price war.
Second is data assets. He cited the data change of a refined-listing seller: initially, using AI to generate images, about 20% could be used directly; after feeding their own SKU data, negative reviews and Q&A feedback into the system, the directly-usable ratio exceeded 50%, with another 30% requiring fine-tuning. The degree of data accumulation directly determines the output quality of the tool. This moat is built with time and data, not by switching tools.
Third is iteration speed. AI lowers the production threshold, but the closed-loop speed from production to testing and from testing to adjustment is what matters. For the same product, a week later the two sellers may be presenting completely different things to consumers.
He also flagged an easily overlooked detail: a single-point AI tool is very efficient, but when strung into a real workflow, you find that the designer's work is liberated while the operations coordinator's time is stretched, and overall personnel efficiency may not improve. So whether the generated image looks realistic is not the key question. Whether the content can be directly called by listings and ads, how compliance risk is managed, and how quality is maintained when running at scale—these together are real production efficiency. Don't just stare at a single generated image; treat content production as a system.
Fulfillment: It Can Be Done, but It Comes Last
What Chen Guang said deserves展开: it is not that AI has no value in the fulfillmentaspect, but that the value has preconditions.
He described what thisaspect can do now: Voice agents are no longer just answering phone calls and questions—during a call they directly call ERP APIs and operate order systems, closing the loop end to end; Visual agents can identify a SKU from a single image and complete a diagnosis; for a certain phone brand's TikTok Shop returns and exchanges in Latin America, an agent already automatically judges and processes them, without a human needing to confirm again.
These capabilities were at the "concept stage" in early 2025; by 2026 they can be commercially deployed.
But his advice remains: get product selection right first, get marketing going, and wait until your customer-service team has dozens of people and a monthly payroll of over a million before optimizing thisaspect.
The logic is straightforward: good fulfillment is the foundation for not losing; getting product selection right and marketing right is why you win. Attention is limited—put it first where the outcome is decided.
He also offered another judgment: for certain scenarios, waiting for the right moment matters more than rushing in early. Something that can be easily deployed today was laborious, expensive and below expectations two years ago. Models are evolving rapidly; timing is worth more than being first.
Three Links, One Conclusion
In product selection, AI helps you scan supply-demand gaps, but which one is your opportunity depends on whether your supply chain and team capabilities match. In content, AI lets everyone produce 60-point素材, but the gap lies in who understands their target customers better, whose data is more complete, and who iterates faster. In fulfillment, AI can already automate end to end, but when to invest matters more than what to invest in.
The three links are saying the same thing: AI provides general-purpose capabilities. It lowers the barrier to entry, but it does not provide a moat. A moat has only one source—understanding your specific customers, and the data assets accumulated on the basis of that understanding.
This, AI cannot do.
More Conversation Details
Guests:
QuickCEP — Founder — Chen Guang
Linkfox — Product Lead — Henk
LynxAI — Founder — Zhang Keyi
Moderator: Youyoutu Technology — Partner and CMO — Yu Jing (Scarlett)
How AI Is Reshaping Cross-Border E-Commerce Product Selection, Live Streaming and Fulfillment
Yu Jing: Welcome to the final Panel of the day. Listening from the first session this morning until now, I realize I may be the first woman to take the stage today. In AI and going-global, we very much need and look forward to more women founders joining us.
Today's conference theme is excellently designed—from the early discussions of products and macro concepts, to the previous Panel on content and influencers, to our session which is very hands-on. We mainly discuss: how cross-border e-commerce is reshaping product-selection logic, live-stream selling and cross-border fulfillment.
At this current juncture, these three links have already undergone enormous change. In the past, people might only care about "which product will go viral," "where the traffic is," and "how to ship more cheaply." But now, we care more about how a closed-loop system driven jointly by AI, content creators, supply chains and user operations can run efficiently.
The three guests invited today恰好 represent different perspectives along this chain. First, please briefly introduce yourselves and talk about what kind of cross-border sellers and brands your products are serving.
Chen Guang: Hello everyone, I am Chen Guang, founder and CEO of QuickCEP. We are an AI Agent product serving global brands in customer service and consumer operations. We currently serve several hundred mid-to-large brands, primarily Chinese going-global brands and large cross-border sellers. Thank you.
Henk: Hello everyone, I am Henk, product lead at Linkfox. Linkfox was founded relatively early—when AI first broke out in 2022, we entered from multimodal (images and video), mainly helping cross-border sellers generate product images and videos. Now we are also expanding into upstreamaspect such as listing research and listing publication. Currently Linkfox serves about 1 million merchants. Welcome everyone to try and exchange ideas.
Zhang Keyi: Hello everyone, I am Zhang Keyi, founder and CEO of LynxAI. In our view, product selection is actually one dimension of product innovation. LynxAI's predecessor was an intelligent product-manager agent; now it has evolved into an intelligent decision-making platform for enterprise product innovation, more deeply integrated with enterprise digitalization.
On the business side, we use data to judge which direction of products is at the pre-outbreak stage or has already broken out. If we have time to prepare, we do deep product innovation; if we need to quickly seize an opportunity, we do efficient supply-demand matching. Product innovation is a value-creation process: the direction must be right, find consumers with real demand, see what alternative solutions are in front of them, then evaluate what additional value our product and fulfillment capabilities can create on top of these existing solutions. This determines whether we should do it—heavy has heavy's choice, light has light's product selection.
Yu Jing: Thank you, Mr. Zhang. Mr. Zhang comes from Anker Innovations, with very rich experience across the full hardware and product chain. Later you can share more with us about your "AI product manager" concept.
Core Topic 1: How Is Product-Selection Logic Being Reshaped by AI and Data?
Yu Jing: Let's start with product-selection logic. In the past, product selection relied quite heavily on the personal experience of cross-border bosses or product managers—looking at platform rankings, watching competitor moves, or looking up at the supply chain. After AI transforms the workflow, what has changed in the underlying logic of product selection? Whether a product can be done—what dimensions should be evaluated today?
Zhang Keyi: To tie back to the theme, we divide product innovation into three dimensions:
Product managers primarily do value creation, the premise being to find target consumers;
Supply-chain managers (broadly including full-chain customer experience) do value delivery;
Operations / marketing / brand do value communication, which in our perspective can be approximately equated to traditional "product selection."
From a product-selection perspective, the market is always dynamically producing supply-demand imbalances, so there is always room for quick-reaction arbitrage. In the AI era, the essence of product selection is to capture this faster and better supply-demand matching logic through massive data analysis.
In the past, the ceiling on how much data a human brain could process per unit of time was far lower than AI's. Now, using AI, you can process large amounts of data in a very short time and make opportunistic captures. Then, matching with our existing deterministic conditions—China's strong supply chain, excellent engineers, and the founder's understanding of products—you can quickly satisfy unmet market opportunities on channels such as Amazon, TikTok, or independent-site private domains. From an ROI perspective, AI exerts an enormous leverage effect in this chain.
Yu Jing: Is there a specific number for this efficiency cycle? For example, how long did it take with traditional methods in the past versus with AI now?
Zhang Keyi: This is one person's view. LynxAI works relatively deeply with brand clients, currently serving 24 clients. If we look only at the product-selection dimension, taking common fast-moving consumer goods or high-turnover consumer electronics as examples, the traditional product-selection cycle in the past generally required 13 weeks (about three months). Now, purely from the cycle of data acquisition, analysis to outputting initial decision conclusions, it can theoretically be shortened to 4 hours. Of course, this is only efficiency at the data and decision level; the enterprise's final decision and execution must return to the real physical world.
Core Topic 2: New Changes and Differentiation in Live-Stream Selling and Content Conversion
Yu Jing: 4 hours is indeed astonishing. Next, let me ask Henk. Linkfox focuses on AI-generated product images and listings. In the past, what were the most common mistakes sellers made with product images? If sellers want to use AI for product images as daily foundational work today, what specific issues should they focus on more?
Henk: I think there is a kind of "positioning mismatch" in how sellers and we practitioners understand AI tools. As a service provider, in previous years we focused on Stable Diffusion, Midjourney, and more recently researched Image-to-Image technology that releases prompt pressure, keeping up with the technology frontier all day. But for most sellers, the real penetration rate and depth of use of AI are not that high.
The day before yesterday I went to teach a friend who is a mid-sized Amazon seller how to use Codex's Computer Use feature, and I found they are still纠结: is this image realistic? Is the subject clean? Does it look like a template? They put a lot of energy into the technical details of "how the image was generated and whether it looks real."
But if we return to real business operations, we find that as the core information carrier in a listing, images really need attention on these four points:
Efficiency and output: Rather than focusing on how it is generated, focus on how it is used. For example, we have clients who generate 1,000 images a day; they care more about how many can go directly into the asset library, be quickly used by listings and ads, and comply with platform rules. In today's technology democratization, generating images is not hard; using them efficiently is.
Timeliness and rhythm: In the past, sellers were used to scheduling and waiting for designers to retouch images. But doing bulk or refined listing in China requires extremely fast product testing and chasing holiday trends. The old work paradigm simply cannot keep up with the current business rhythm; there must be new requirements for image quantity, quality and tool fit.
Overall personnel efficiency: Many tools have very strong single-point functions, such as local elimination and keyframe modification. But back in the workflow, if it involves multi-image fusion, multi-SKU batch image generation and other scaled operations, a single product cannot flow through, and humans still have to keep intervening and adjusting. The result is that designers are liberated but operations time is stretched, and the overall organizational efficiency has not improved. We need to treat content output as a "content production system," linked with product libraries, ads and listings.
Compliance and copyright zero risk.
Yu Jing: What Henk shared is indeed very practical干货. Many tool providers may not yet fully fit sellers' actual work scenarios, resulting in overall flow efficiency not reaching maximum. This is indeed a pain point that product development needs to solve.
Next, let me ask Mr. Chen. When people discuss product selection, they usually value the early product-creation stage. But after a product launches, there is a large amount of interaction data from customer service, reviews and email inquiries. How should brands reverse-mine the next round of product and product-selection opportunities from these backend user data?
Chen Guang: Our QuickCEP product mainly helps brands do AI shopping guidance, after-sales, customer-service reception and social-media fan interaction management after they have traffic and sales.
In the early days when model capabilities were not strong enough, clients would periodically ask us to export data, which they would take back and manually organize, mining directions for iterating new product features or finding derivative-product opportunities. This is a classic methodology.
With the development of Agents this year, we launched an "Analysis Agent." Every company analyzes data from a different perspective; they can turn their own analysis logic into a "Skill" and configure it into the Agent. After that, AI automatically transcribes LiveChat, emails and phone recordings every day, and抓取 all Reviews and Comments from major overseas e-commerce platforms, then runs the analysis according to this Skill. The frequency of such insights and feedback becomes very fast, greatly improving the adjustment efficiency of supply chains, product design and operational processes.
Yu Jing: This is very similar to our experience at Youyoutu doing content e-commerce. After接入 a large amount of merchants' review-section data, we found there are many product-improvement details hidden inside. Attributing this data and feeding it back to sellers has enormous value. Mr. Chen, have you observed how brands' R&D teams generally use this kind of user-experience and complaint data to drive the next round of innovation?
Chen Guang: As a service provider, we are relatively far from their product R&D departments. We mainly interface with customer service, operations and IT departments; after providing the data, it flows and is used internally by them.
Yu Jing: Understood. Let's move to the second section: live-stream selling. Not all products are suitable for live-stream selling. Mr. Zhang, how do you judge whether a product is suitable for live streaming? Can it be evaluated using AI or data?
Zhang Keyi: Judging whether a product is suitable for live-stream selling requires looking at two dimensions:
Content dimension: determines dissemination. Whether quality content fits platform rules and has built-in dissemination properties.
Product dimension: determines conversion.
In my solution perspective, only when the product is right and has the potential to convert is it worth disseminating (this refers specifically to selling, not pure brand-awareness or fan-building operations). Under the premise that the product is right, we then evaluate whether it is suitable for live-stream selling or for other scenarios.
When we do entirely new product innovation, we also fully consider dissemination fit. Internally we have an "80/20 principle": even if we discover an extremely deep, unmet user need, to ensure win rate and dissemination breadth, we will most likely choose to do 20% product innovation and inherit 80% of mature design. Those more cutting-edge innovations we place in second- and third-generation products to realize gradually.
Yu Jing: Very much agree. Just like our short-video marketing, products with strong visual contrast that can show Before/After (such as beauty and skincare) naturally tend to go viral in live streams and short videos. Considering the subsequent sales channels and visual presentation at the early product-selection and product-innovation stage is the most efficient.
Henk, you mentioned that the普及 of AI tools has brought technology democratization. When all sellers can use AI to batch-generate image and video content, content will inevitably face homogenization. How should sellers体现 differentiation at this point?
Henk: Homogenization is certain. Just like our company internally now uses Cursor or Claude for coding and marketing. People sometimes joke that in the end everyone becomes a Prompt engineer. Using the same model and similar prompts, the results are inevitably more or less the same. In visual and video content, this homogenization is amplified even more.
In the past, differentiation depended on designers' aesthetics, image-generation speed and operational rhythm. Now, AI has flattened the threshold—a 60-point, quite good-looking asset can be easily batch-produced. When consumers develop visual fatigue, it forces us to seek deeper differentiation. We believe that after everyone's starting line has been raised by AI, differentiation mainly comes from three places:
Understanding user scenarios: Selling the same product, you target the gifting scenario, I target the household practical scenario. Different understanding of customers directly projects onto content copy, image angles and pain-point提炼, thereby producing variants. If you only rigidly push the product itself, you will end up in price involution.
Accumulation of private data assets: OpenClaw is especially hot this year, and many large sellers are actively talking to us about how to get SOPs running and how to接入 APIs. They realize that feeding the enterprise's own private data (such as historical dimensions, specific negative reviews, Q&A feedback, etc.) directionally to AI systems or models, letting AI produce brand-specific content based on this, is the new moat.
Speed of data iteration: Everyone gives one instruction and generates a batch of images—that is homogenization. But after I list the product, click data, organic traffic and review data are feeding back every day. If I can maintain an extremely high rhythm of iteration, constantly修正 content based on data, a week later the message we convey will be completely different.
Closely follow users, use new tools to re-wash your private data assets, and maintain rapid iteration—this is the core of maintaining differentiation.
Yu Jing: AI does the foundational work, but the "Skills" and dataaccumulated by different brands are different, so AI ultimately is also personalized for each. Also, after AI improves efficiency, people can put their energy into heavier, deeper creativity and strategy, which can also bring differentiation.
Mr. Chen, front-end live streaming brings enormous traffic, but many going-global brands cannot承接 it on the back end—slow customer-service response, insufficient multilingual capability, leading to user churn. What role can AI Agents play in the post-live-stream chain (shopping guidance, customer service, repurchase conversion)?
Chen Guang: First, we do not yet directly接入 real-time interaction in TikTok live-stream comment sections; this is something we are evaluating and will do in the future.
But the "post-chain" is precisely what we are best at. For example, for some high-ticket, long-conversion-cycle categories that mainly build brand influence in live streams, the purpose of live streaming is actually "lead capture" (obtaining emails, WhatsApp, etc.). Subsequently, we need to continuously reach out based on different user profiles. We help brands接入 the world's mainstream post-chain reach channels (email, SMS, phone, as well as overseas locally-preferred WhatsApp, Zalo, Line, etc.), interacting in the way local consumers are most accustomed to—open rates and interaction rates are much higher. Through CDP (Customer Data Platform) combined with AI, personalized content is automatically generated for different users' interests, running automated reach flows.
Also regarding live streams and short videos, I recommend that brands, regardless of whether their products are suitable for live-stream selling, should do short videos and live streams. Because this is not just a one-time sale, but alsoaccumulated content assets. Content from live streams on TikTok or other platforms, and from influencer demonstrations, can be clipped by AI or manually into highlight moments. We provide a short-video conversion tool that can directly embed these clipped videos into the homepage, category pages and product detail pages of independent sites, letting influencers dynamically展示 products on the website.
Many of our clients have seen very明显 conversion-rate improvements after using it. For example, in the wig category, one client calculated that after switching to dynamic video, the overall daily conversion rate improved by 13%, because wigs very much depend on the dynamic effect after wearing. Another example: 3D printers, AI PCs and other tech products that require geek demonstrations, as well as actual usage scenarios for furniture. Live-stream selling is not just a front-end traffic entry point, but also fuel that continuously generates content assets and improves full-chain conversion rates.
Core Topic 3: How Can Cross-Border Fulfillment Shift from Back-End Delivery to Front-End Growth?
Yu Jing: This shows us that live streaming is a full-chain asset amplifier. The final Topic is cross-border fulfillment. In the past, fulfillment meant logistics, warehousing and customs clearance, but today the boundaries of fulfillment seem to have been redefined—it covers customer service, after-sales, returns and exchanges, directly affecting front-end growth and user experience. Mr. Chen, how do you view the generalization of fulfillment boundaries?
Chen Guang: Fulfillment in the atomic physical world (warehousing, courier delivery, etc.) is limited by objective physical conditions, and efficiency improvement has its objective upper limit—we cannot help with that part. What we do is, on top of the physical fulfillment chain, improve information transparency, efficiency and customer perceived satisfaction.
Two examples. We serve many overseas warehouses and cross-border logistics companies. They use our Voice Agent to make small calls to overseas consumers before courier delivery, doing delivery notifications or delivery confirmations. This is very common in China (for example, SF Express and JD.com ask before delivery whether to leave it at the door or in a locker), but in many places in the United States it is not common. We use AI to help them do this high-concurrency notification, and package delivery success rates and user satisfaction have improved significantly. At the same time, we provide privacy-number calling through AI relay, protecting the privacy compliance of couriers and American consumers.
For brands, we use AI Agents to接入 automated judgment of after-sales returns and exchanges. By setting up Skills, AI can directly judge whether to refund or exchange for the user according to rules, without the need for layer-by-layer manual Check approval as before. Recently we helped a certain phone brand接入 a fully automated return Skill for its TikTok Shop in the Latin American market, and most return-and-exchange orders can be directly and automatically processed by AI. For large brands, consumers' perceived experience is very good, and after processing it automatically triggers an email or WhatsApp for a satisfaction follow-up. The entire SOP is connected as one.
Yu Jing: AI Agents turn庞杂 back-end services into standardized, automated processes, not only raising satisfaction but also bringing incremental growth to the front end.
Mr. Zhang, from a product-strategy perspective, fulfillment capability used to be considered only at the very back end. Should it now be前置 to the product-selection and R&D stage? For example, for some large-item or extremely-high-return-rate categories, although they sell well, the back-end fulfillment is too heavy—should we not touch them?
Zhang Keyi: When an enterprise has the capability and conditions, the earlier fulfillment is considered, the better—on one hand seizing opportunities, on the other avoiding risks. This is actually the concept of CEM (Full-Chain Customer Experience Management).
In this process, the origin point remains: define who the user is, what needs they have in what scenario, and achieve the right product matching the right customer. I emphasize this because: managing customer experience, sometimes "not disturbing" is the best experience. If certain customers are unrelated to the core value of my product and service, I should not force them into my circle.
Among the right customers, different people have different expectations for experience. Some users are more pragmatic—they only care about the purchase cost of payment and to what extent the product solves their problem. At this point, the front-end core is "say it as it is" and do a good job of expectation management.
So returning to the product side, the premise of fulfillment is "value delivery," and the premise of delivery is "the product itself must have real value." Find the right user, come up with a right and valuable product, then do value delivery—this is the directional premise.
Yu Jing: Precisely find users, do a good job of expectation management and foundational value. Henk, from a tool and operations perspective, before a product goes on sale, how can sellers judge fulfillment risk through data? How can back-end data such as negative reviews, return reasons and after-sales keywords reverse-feed and optimize front-end listings, image assets and product descriptions?
Henk: Linkfox does not directly涉足 back-end logistics fulfillment services, but for data fed back from the back end, the product mainly plays a role in risk warning and information completion.
We correlate abnormal review data such as damage and returns in negative reviews with sellers' inbound goods and operating reports. In the future, when sellers want to develop variants for this product or rewrite listing content, this data becomes a ready-made risk-assessment basis.
Specifically for listing optimization, it mainly depends on keywords in Q&A and negative reviews. For example, consumers frequently ask about size and specifications in Q&A, or complain about a poor fit somewhere in negative reviews—we guide sellers to integrate this information into the front-end listing. In images and copy, add multi-angle explanations and supplement more detailed size comparison tables. We regard data fed back from the fulfillment end as a risk control point for initiating the next product listing or operation.
Summary: Looking Ahead to 2026 and Beyond
Yu Jing: Thank you very much to the three guests for their wonderful sharing. It is already one o'clock, and there are still so many audience members坚守 below. Finally, please each give sellers and brands the most core advice in one sentence from your respective fields (Mr. Zhang on product selection, Henk on content and live streaming, Mr. Chen on cross-border fulfillment).
Zhang Keyi: On product selection and innovation, I suggest everyone take two steps:
Step one, resolutely and quickly use AI tools to build the enterprise's capability to acquire and process massive data and identify potential product-selection opportunities. This technology has already begun to rapidly普及 at the brand end and is a highly deterministic first step.
Step two, among the many leads recommended by AI, identify which are truly yours and which real opportunities your R&D and fulfillment capabilities can match.
When manually judging and adopting AI's conclusions, I suggest looking at three layers:
Bottom layer: Is there structured real factual data support?
Middle layer: What are the data logic and business logic that AI runs on? Turn the black box into a white box.
Top layer: Based on these logics and data, what reasoning and choices did AI make?
Seeing these three layers clearly makes human-machine interaction more effective, and allows distinguishing which are suitable for fast-in-fast-out light arbitrage opportunities and which are brand-innovation opportunities worth deep cultivation and planning.
Yu Jing: Mr. Zhang, by the way, regarding the 4-hour product-selection plan mentioned earlier, how are the actual sales and hit rates in the physical world? Is there data you can share?
Zhang Keyi: Case by case. Looking at the limited data from our 7 brand clients we have worked with for over a year, from AI identifying leads, product planning, manufacturing delivery to launch, the new-product period takes roughly 6 to 12 months. During this year-plus period, each client has had no fewer than 3 products for which AI provided leads and assisted product innovation that successfully "went viral."
Here, "went viral" has a quantitative standard: the product's actual sales result exceeded the sales-target forecast we made at the new-product launch stage.
But I want to emphasize that the R&D, manufacturing and delivery cycle in the physical world is currently not greatly shortened; AI efficiency improvement is mainly at the virtual data and decision end. The biggest change AI brings is not simply shortening the physical cycle, but a significant improvement in project-launch success rate, as well as full-process traceability, reviewability and structured data accumulation.
Henk (content and live-stream perspective): Content is already extremely competitive now, because the underlying logic has changed—the era of "cheap asset production" has disappeared, and any individual can use AI to easily generate passing-grade content.
So I suggest: don't fear or be anxious about tools; quickly integrate generative tools, live-stream口播 and other content-production capabilities with your own private data (Skills), and融入 them into the workflow.
We just did a test last week at a large refined-listing seller. They process about 3,000 images a day. If directly using native official models (such as nano banana pro, etc.), only about 20% of images can be used directly, and over 45% require manual fine-tuning. But because they embraced it quickly, built their own data middle platform, and integrated the enterprise's own historical assets and Skills with the workflow for targeted training, by last week the ratio of directly-usable AI images had already exceeded half (50%+), and the remaining 30% still requires fine-tuning, but overall personnel efficiency has undergone a qualitative leap. Tools are advancing rapidly every day; as long as you have your own data, every new tool you integrate can immediately translate into business results.
Chen Guang: Everyone, don't be anxious about the fulfillment and serviceaspect, and don't FOMO (fear of missing out) too early.
I suggest everyone first follow Mr. Keyi and Mr. Henk, define the product well, do a good job with traffic, marketing and content assets, and first get the盘子 selling. Once the business has a certain scale—for example, the customer-service team grows to dozens of people with a monthly payroll of over a million—then come back and find the most professional AI service providers on the market to organize your service and fulfillment system.
Because honestly, fulfillment services can help you control costs and control negative reviews, but they are not the essence that determines whether your brand explodes and succeeds from 0 to 1. And sometimes doing it too early makes you vulnerable to being backstabbed by technology iteration. In 2024 we helped a company build a visual agent with the architecture at that time, and the效果 was quite unsatisfactory; but today, with the rapid enhancement of large models and Agent architecture, many commercial scenarios that two years ago were extremely costly and could not be deployed are all visibly deployable today.
Customer service in the past could only call data to answer questions; today's Voice Agent can, during a phone call, end-to-end call APIs to directly close the loop and operate your ERP, order systems and e-commerce platform interfaces; a visual agent can see an image a user sends and directly diagnose which SKU it is. AI is developing rapidly, and customer service and back-end operations will in the future become a very stable, cost-controllable and quality-controllable center for you. Everyone put your energy into good products and good marketing; leave back-end stability to professional vendors like us.
Yu Jing: Very practical advice! Many times people worry they cannot keep up with technology, but actually sometimes if you wait one more wave, you will find that technology itself has already lowered the threshold very low—no need to over-FOMO.