Original · Unique Research · 2026-08-19
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, four insight sections, closing reflection, and the full roundtable transcript. All named companies, products, and people are preserved. Customer counts, employee numbers, and revenue figures are speaker claims, not independently verified findings.
AI Industry Observer
AI Customer Service and Human Customer Service on One Stage: The Unvarnished Truth
"Strategy sometimes works this way: what you decide not to do tells you more clearly what you truly want to do."
Five people on stage. One runs an AI customer service company. Another runs a human customer service company with 2,000+ employees.
Logically, these two should be fighting.
But at the FOSHO DAY 2026 roundtable, the AI customer service person Chen Guang said: if an enterprise can only invest AI in one part of the chain, invest in marketing first, not customer service. The human customer service person Wang Gengdong said: 70% of our company's business is already after-sales phone support; pre-sales? Basically handed to AI.
The AI guy tells you not to rush your money into customer service; the human guy voluntarily discloses which positions have already been evacuated. What makes this roundtable fun is this: five people with completely different利益 positions say things that corroborate each other, like a cross-examination.
Let's meet everyone. The host is Unique Research's Wu Wei (吴畏). Five guests: QuickCEP founder Chen Guang (陈光), doing AI customer service and consumer operations, serving 70% of China's top 100 going-global brands; Vinyl (乙烯) founder Andy, doing overseas crowdfunding and advertising, with H1 business growing 2-3x YoY; PingPong China e-commerce GM Zheng Zhikun (郑志坤), the money person; BrandLink (链帮出海) co-founder Wang Gengdong (王耿东), the people person with 2,000+ employees, serving 70% of top smart hardware brands; Deep Edge AI's Albert, the only one on stage who built an agent for physical offline spaces.
Five people, exactly covering a going-global full chain: market research, marketing acquisition, payment fulfillment, after-sales service, plus one offline physical.
If AI Could Only Be Invested in One Link? Nobody Voted for Themselves
Wu Wei opened with a tough question: suppose it's 2026, and an enterprise can only invest AI in one link — which one?
This question is actually tricky. Everyone on stage has their own business; voting for which link is voting for whose business.
Result: nobody voted for themselves.
The customer service guy Chen Guang answered first: marketing. Reason is direct: growth is the first layer; without growth, customer service doesn't have that much to do.
The payment guy Zheng Zhikun also voted for marketing. His view is data: Chinese enterprises have strong supply chains and strong operations; by comparison, marketing capability is still the most obvious shortboard.
Wang Gengdong was honest. He does human customer service, logically should defend the customer service budget most. But he directly showed the books: customer service is only a few percent of operating expenses in many companies, even lower; marketing investment can be ten-plus or twenty-plus percent. If you invest, invest in the link with more operating weight.
Andy is a marketing guy; voting for marketing isn't surprising. Even the shopping guidance robot guy Albert answered: market research and consumer insights — in his words, product development is getting faster and faster; the real difficulty is finding the right market. Roughly speaking, this is also marketing's upstream.
Five people, four and a half voted for marketing. From the perspective of people working the front lines, going-global enterprises in 2026 are short of orders far more than they're short of efficiency. The cost-reduction story has been told for three years; everyone knows in their hearts: saved money is linear, grown money is exponential.
How Customer Service Becomes a Revenue Center: One Talks Context, One Talks Giving Up
The next question was the highest-value part of the roundtable: customer service has always been a cost department — how do you make it create revenue?
Chen Guang's answer is context.
His logic: today's Agents all use roughly similar models underneath; the basic capability gap isn't as large as imagined. What really separates outcomes is context. You use a product long-term; why does it increasingly understand you? Because your context is deposited in it.
For AI customer service to move from "answering questions" to "creating revenue," it must know orders, spending habits, which marketing emails this user clicked, what interactions they've had with the brand. In industry terms: context is all you need. The more complete the context, the more likely the Agent can do shopping guidance, recall, and conversion along the way.
The second thing is more interesting. Chen Guang says the so-called Agent "self-evolution" isn't some mystical word; it's essentially that every conversion round can keep optimizing: bring the marketing automation A/B testing playbook into the Agent, continuously test, learn, retest, and gradually approach the optimal solution at some step.
He also pierced a layer of window paper: many enterprises actually knew they needed refined operations, but the problem was ROI didn't work. Employing a bunch of analysts and operators for very granular actions, the labor cost is too high, not worth it. Now these actions can be handed to AI, and things that previously didn't have good cost-performance suddenly do.
Same question, Wang Gengdong from the human side gave a mirror answer.
His company's business structure has changed significantly in recent years: pre-sales, email, phone all done, now 70% is after-sales phone. Pre-sales, especially text shopping guidance, AI is already very good at — high accuracy processing text, occasional small errors with low risk, so hand it over.
After-sales can't be evacuated, especially extreme complaints and high-ticket products. He gave an example: they serve some home energy storage products, tens of thousands of dollars per unit. When a user has a problem, if customer service can diagnose and solve it on one call, it directly saves an engineer visit. At that point, customer service is itself making money.
Then he said the line I liked most of the session:
"Look at what we voluntarily give up, and to some extent you can see what AI can already do."
They now voluntarily give up some seller-type clients and low-ticket categories. A $15 clothing item with a problem — direct refund is the most economical solution, no need to invest heavy human labor. What they hold on to is top brands, high-ticket, high-complexity products; these brands also earn brand premium and user mindshare money, and between people and the brand there needs to be deeper contact.
A company under AI siege, defining its moat by "what we give up." This thinking is clearer than most AI evangelism.
What AI Can't Replace: Timing, Emotion, and the Blame Nobody Wants to Take
As the roundtable went on, the topic shifted to: what things can AI simply not do?
Andy's answer is one word: timing.
Crowdfunding is an industry particularly suited to accumulating data. Vinyl has been doing it for nearly ten years; every project ends with a stored Report, accumulating complete data on 500+ projects, from hundreds of thousands to nearly ten million dollars. This historical data has now become very valuable AI assets. He puts it plainly: agencies that haven't actually done that many projects can't get this kind of data.
Now they see 1,000+ projects a year but only take 30-40. Clients first fill out a questionnaire, AI directly analyzes the product's crowdfunding ceiling and floor, how to set budget, what to do at each stage, then scores the project. They have an internal line: accepted projects should have a shot at $1M+. The initial screening — AI does it fast and well.
But there's one thing AI can't do: it doesn't know that right now, ten companies are rushing to market with very similar products.
The hardware market changes too fast. One crowdfunding hit comes out, three to five similar products appear in three months, ten-plus within six months. This real-time competitive state simply doesn't exist in public data; it can only be smelled through frontline project experience. So product selection ultimately comes down to three things: innovation, experience, timing. The product itself matters, but whether you enter at the most appropriate time matters equally.
On the marketing side there's also something AI can't handle: emotion.
Andy mentioned influencer negotiation. Bargaining is a highly human process; an influencer has been pushed too far and emotion has visibly shifted, but AI doesn't notice and keeps pressing, and the whole collaboration falls apart. A brand's unique aesthetic, expression, and sense of value aren't solved by simply improving click-through rate.
Zheng Zhikun's answer is the sharpest. When discussing who takes responsibility when AI goes wrong, the money person said:
"If an enterprise puts AI in a scenario today and the first thing it thinks about is who takes the blame if something goes wrong, then that scenario probably shouldn't use AI yet."
His logic is: AI can already do so many things; just low-risk, clear-boundary efficiency scenarios have enough space — no need to put it in a place where responsibility boundaries aren't thought through. Even trading cards, a highly emotional consumption industry, has AI authenticity verification and assisted authentication with clear value. So don't ask which industry can use AI; ask which step has clear efficacy and clear responsibility.
PingPong itself is about to launch an AI foreign exchange product: rating sellers by currency, risk factors, and payment cycle, outputting FX strategy. Why foreign exchange? Because a few points of exchange rate fluctuation a year is a big profit variable for cross-border enterprises, and many finance people don't lack the ability to use forwards and hedging — they lack the courage. Do it well, the boss praises; do it poorly, you take the blame. AI first structures the complex information and gives recommendations; the boss only sets a risk appetite: aggressive, neutral, or conservative.
From payment data to business decision-making, Zheng Zhikun actually hopes AI goes slower and deeper. The reason is so simple it's irrefutable: everything in this industry is about money.
The Robot Guy Actually Doesn't Want Robots to Dance
Albert is the only physical agent person on stage. By current standards, robotics people are the most likely to get carried away, but he was the one pouring cold water.
His shopping guidance robots all move on wheels. Reason is simple: in efficiency terms, wheeled mobility is far more realistic than bipedal in many commercial scenarios.
He did enterprise consulting before, and his judgment on this is hard: robots doing kung fu, dancing — those things you only need to see once a year. What enterprises need long-term is a role that continuously creates value every day.
Their entry point comes from a real pain point: a 2,000-3,000 square meter smart hardware experience center with over a thousand SKUs; every salesperson must understand and introduce them. An employee takes two or three months of training to be barely useful, then leaves. This cost black hole is where robots survive.
On the perennial problem of data integration, Albert's answer is also different. Everyone talks about API, MCP, system integration; he directly retreats to edge: clients build their own local AI servers, sensitive data stays in their own environment as much as possible. Reason is equally simple: large models are strong, but large models don't have your enterprise's own private data. He calls this the moat of the "software 3.0" era: only your own data can train a system that truly understands your business, customers, and organization.
Chen Guang added an industry perspective: much of today's software infrastructure hasn't yet adapted to the Agent era. Working with large clients, huge amounts of time are spent coordinating Salesforce, OMS, ERP system integration — very wasteful. In the future every software vendor must become more Headless, opening capabilities through CLI, API, MCP. The interface is worthless; what's valuable is the capability model behind it that has nailed a specific vertical scenario.
What can be handed to Agents? Chen Guang's line is also clear: processes already固化 into SOP, with safety guardrails and regular evaluation, can be gradually let go. New product launches, new businesses, sudden failures — these high-risk things still need human judgment. Responsibility ultimately lands on the client's decision-maker; the tool gives advice, execution is the enterprise's choice.
Is There an AI Bubble? Five Different Answers
Closing question, one line each: is there an AI bubble now?
Albert: Yes. The biggest bubble is the cognitive gap between everyone's expectations for AI and its real capability boundary. People who don't use it often think AI is omnipotent; people who use it heavily every day increasingly understand where its boundaries are.
Wang Gengdong: Absolutely. Too many people doing it; many directions' thresholds aren't as high as imagined. The ones that ultimately survive are companies that, beyond AI, already have unique business capabilities, industry resources, and irreplaceable assets.
Zheng Zhikun: There's a bubble, but a bubble isn't necessarily bad. Something that can actually be used in real scenarios and create value today isn't a bubble; something with no real use case today is a bubble in the present. Whether it's useful in the future — that's a future question.
Andy: If I must name one, embodied intelligence has the biggest bubble. We've actually tested some projects; experience and usability are far from expectations. Many products look busy but don't have features that can stably enter real use scenarios.
After this line, I wonder what Albert on stage thought — he was the only one doing physical robots. Fortunately, Albert's earlier line "robot dancing is enough to see once a year" had already made his cut for him.
Chen Guang's answer was the most contrarian: AI software has no bubble.
He gave two reasons. First, compute is still scarce; new models get rate-limited as soon as they release, showing supply can't keep up with demand. Second, doing large enterprise deployments, they increasingly feel model capability is already strong enough; the real obstacle is how to reform old systems, connect data, and integrate business processes. Once these bottlenecks are broken through, actual token usage will keep growing. As for embodied intelligence, he says it needs more time to find stable scenarios.
Note a detail: among the five, the only one who said "no bubble" is the one who narrowed it most narrowly. He only defended AI software, deliberately cutting out robots.
Closing Notes
After this roundtable, the biggest takeaway isn't a specific judgment — it's that how these people think about AI has changed.
Two years ago everyone asked "what can AI do." Now these five frontline players answer "what do I dare hand to AI, and what do I firmly not hand over." Andy dares hand initial screening but not timing; Wang Gengdong dares hand pre-sales front but not extreme complaints; Zheng Zhikun dares hand strategy recommendations but not payment instructions. Chen Guang draws SOP-fied processes for Agents and leaves emergencies to humans. Albert is more thorough — he won't even let data leave the client's server room.
Wang Gengdong's line can be the note for the whole session: strategy sometimes works this way; what you decide not to do tells you more clearly what you truly want to do.
This is about his customer service company, but it also sounds like every enterprise in 2026 that wants to use AI.
More Conversation Details
Chen Guang Billy Chen | QuickCEP Founder & CEO
Andy Peng | Vinyl (乙烯) Founder
Zheng Zhikun | PingPong China E-commerce GM
Wang Gengdong | BrandLink (链帮出海) Co-founder
Albert Lee | Deep Edge AI CEO
Host: Wu Wei | Unique Research Founder
If AI Could Only Be Invested in One Link, Where Should It Go?
Wu Wei: Let's do quick introductions first. Each in one sentence — your company and business — then answer one question: suppose it's 2026, and an enterprise can only invest AI in one link; where should it go?
Chen Guang: I'm Chen Guang. QuickCEP mainly serves the customer service link for global brands. We use an AI Agent platform to help global brands with consumer service and operations. We've already served about 70% of China's top 100 going-global brands.
If I had to choose one link, I'd actually say invest in marketing first. Growth is the first layer; after growth comes lots of orders and service demand. Without growth, customer service doesn't have that much to do. Of course, after marketing ramps up, clients arrive, and the customer service and consumer operations behind become very important.
Andy Peng: I'm Andy, founder of Vinyl. We mainly do overseas crowdfunding and independent site advertising, also some marketing-related business. Over the years we've mainly served smart hardware going overseas, plus some innovative new products.
Growth has been very obvious these years. Especially recently the capital market is active, more hardware companies going overseas, many enterprises need to complete 0-to-1 tasks. Our H1 this year compared to same period last year grew about 2-3x.
On AI applications, my own experience is more on the marketing side, so if I could only choose one, I'd also choose marketing. Especially for new products needing seeding, AI plays a big role in market research, content, and early-stage advertising.
Wu Wei: Mr. Zheng, you do payments and see a lot of data. First introduce PingPong, then say where you think AI should be invested.
Zheng Zhikun: I'm Zheng Zhikun from PingPong. PingPong mainly focuses on receiving payments, making payments, and remittance products, helping Chinese enterprises solve overseas payment and capital chain issues.
We serve not just traditional cross-border e-commerce sellers. These past two years we've also served companies like BYD and some F&B going-global enterprises. Many enterprises are already strong domestically, but after truly going overseas, the payment chain is often still missing capability — that's what we solve.
Over the past few years, we've also been deepening into various industries. For example, in 2023 we did an e-bike industry alliance; in 2024 we did one of the earlier AI hardware industry alliances; recently we've also been watching Maker Tools — like 3D printing, where the industry's attention suddenly rose with companies like Bambu Lab.
If asked about next year's new category opportunities, I still like Maker Tools — many iterative products and new companies will emerge. Second is still AI hardware; many companies with strong hardware capability will do another round of product iteration because of AI.
If asked where AI should be invested — for PingPong itself, of course the payment link, because we want to use AI to empower more sellers. But from the client's perspective, I still recommend investing in marketing first. Chinese enterprises have strong supply chains and operations; by comparison, marketing capability is still a relatively obvious shortboard.
Wu Wei: Mr. Wang, you do human customer service — this question is challenging for you. Introduce your company, then say where you think AI is most worth investing.
Wang Gengdong: I'm Wang Gengdong from BrandLink. We do "human," not "artificial intelligence." Founded in 2022, now 2,000+ employees globally. Among top going-global brands, we serve about 50; if only looking at top smart hardware brands, we serve about 70%.
Smart hardware customer service isn't just simple Q&A. Many top brands innovate fast and need to educate consumers, and brand tone and user experience requirements are high, so customer service is essentially part of brand experience.
Will AI replace human customer service? We're of course always watching. So we're also in strategic partnership with AI companies like QuickCEP. Our approach is simple: repetitive things AI can solve go to AI; things AI can't yet solve go to humans. Humans plus AI together provide end-to-end service to brands.
But if asked where enterprises should invest AI budget, I don't think it should all go to customer service. Customer service's share of operating expenses in many companies is actually very low — the whole industry average is maybe a few percent, even lower; marketing investment can be ten-plus or twenty-plus percent. From this angle, you should prioritize links with more operating weight.
Wu Wei: Albert, you're the only one on stage doing physical agents and shopping guidance robots — introduce yourselves.
Albert Lee: I left a big company to start up less than a year ago. Deep Edge AI mainly does physical assistant agent platforms; our most important scenario now is shopping guidance robots. We want agents to jump out of screens and software and interact directly with users offline. Core capability leans toward emotional interaction and the experience-layer "robot brain."
I don't think robots must look particularly human. The most important thing is doing the complete user service chain well: introducing products, identifying members, understanding needs, recommending products, giving payment QR codes, etc. Most of our robots now move on wheels because in efficiency terms, wheeled mobility is more realistic than bipedal in many commercial scenarios.
If asked where AI should be invested now, besides R&D, I think market research and consumer insights are more important. Product development is getting faster; what's truly hard is finding the right market, understanding opportunities, and judging what demand is worth doing. Market insight is the starting point.
Wu Wei: It sounds like everyone has different angles but one common point: AI should first solve the link with larger enterprise value share. Most mentioned marketing; some emphasized market insight, payments, and later customer service and consumer operations.
How Does Customer Service Go From "Cost Center" to "Revenue Center"?
Wu Wei: Many enterprises have always treated customer service as a cost department. How do you make customer service not just answer questions, but also do recommendations, recall, sales conversion, and even create revenue?
Chen Guang: Agents have developed to today; if products use the same underlying model, the basic capability gap may not be as large as imagined. Long-horizon tasks, tool calling — everyone's doing these. What really separates outcome differences is one very key thing: context.
You use a product long-term; why does it increasingly understand you? Because your context is deposited in it. Same for To B Agent solutions.
If you want an AI customer service not just to complete service but also do shopping guidance, conversion, and recall, you must give it enough context. It can't only know after-sales knowledge; it also needs to know orders, user spending habits, what marketing emails this person clicked in the past, what interactions they've had with the brand. The more complete the context, the more the Agent can move from "answering questions" to "creating revenue."
Second, make Agents continuously optimize based on data when executing marketing and consumer operations tasks. Marketing automation always needed continuous A/B testing. Now you can bring this closed loop into the Agent, letting it adjust strategy based on each conversion round.
The so-called "self-evolution" isn't a mystical word; it's essentially that every conversion round can keep optimizing. You keep testing, learning, retesting, and gradually approach the optimal solution at some step.
In the past, many enterprises weren't unaware they needed refined operations; it was that ROI didn't work. You needed many analysts and operators doing granular actions, costs too high. Now if these actions can go to AI, things that previously didn't have good cost-performance may suddenly be worth doing.
So I think two cores: first, give the Agent sufficiently complete consumer context; second, deposit the operations process into continuously optimizable Skills. After the Agent truly understands this user, revenue growth has a foundation.
In Crowdfunding, Can AI Replace Experience?
Wu Wei: Andy, crowdfunding is often the first stop for consumer electronics and innovative products going overseas. From market research, product selection, new product development, to Kickstarter and Indiegogo launch — in this chain, what things are suitable for AI, and what can't AI replace yet?
Andy Peng: Crowdfunding is actually a very vertical industry, also especially suited to accumulating experience. We've been in this industry nearly 10 years; we clearly know what scale a project ultimately reaches, what the data behind it is, what launch timing the product entered at, and how it ultimately achieved that result.
In the past, much of this was in the heads of a few core people. We store a Report after every project, now accumulated to 500+ projects. There are projects of hundreds of thousands of dollars, some of millions, some approaching ten million dollar level.
This historical data has now become very important AI assets. Not top agencies that haven't actually done that many projects can't get this kind of data.
We might see 1,000+ projects a year but actually serve only 30-40, because a project cycle often takes four months. Now we have clients fill out a questionnaire first; AI directly analyzes based on product info: this product's crowdfunding ceiling and floor, how to set budget, what to do at each stage of the cycle, then scores the project and judges whether to take it.
We have our own Benchmark — accepted projects should have a shot at $1M+. AI can first help us complete lots of initial screening.
But this also particularly shows what AI can't replace. It's hard for it to judge "is now a good timing." It can grab online info and compare against our historical projects, but it may not know that right now, ten companies are rushing to market with very similar products.
The hardware market changes fast. One crowdfunding hit, three to five similar products in three months, ten-plus within six months. This real-time competitive state, without frontline project experience, is hard to judge only from public data.
So I think product selection ultimately comes down to a few things: innovation, experience, and timing. Timing is especially important. Early multi-color 3D printing, laser engraving machines, energy storage products, including some new categories now suited for seeding — the product itself matters, but whether you enter at the most appropriate time matters equally.
Can Payment Tools Be Upgraded by AI Into Business Decision Systems?
Wu Wei: Mr. Zheng, PingPong holds the chain of enterprise funds in and out. Can AI further upgrade payment tools into enterprise cash flow management, even business decision systems?
Zheng Zhikun: This question has always been challenging for the payment industry. Sometimes I wonder, if there had been today's strong AI ten years ago, could payment have done many things? But thinking carefully, ten years ago maybe there wasn't today's strong demand either.
Because ten years ago doing cross-border, many enterprises were single market, single brand, single platform. Today it's completely different.
First, multi-market. Many enterprises don't just do Europe and America, but also Southeast Asia, Middle East, more markets.
Second, multi-platform. Besides Amazon, there are TikTok and other channels.
Third, multi-model. Besides shelf e-commerce, there's content e-commerce and DTC.
Multi-market, multi-platform, multi-model stacked together, payment complexity is completely different. Different markets correspond to different currencies, different platforms have different payment cycles, and it also involves finance, tax, exchange rate, and capital management.
Many bosses hope to find one particularly strong finance person who solves all these problems alone, but in reality what truly covers these is often a finance team.
We're about to launch an AI-related foreign exchange product. It will rate sellers by different currencies, different risk factors, different payment cycles, then give corresponding FX strategies to assist finance decisions.
FX risk's impact on enterprise profit is actually very direct. For example, a few points of exchange rate fluctuation a year is already a very big profit variable for many cross-border enterprises. The problem is, forwards, hedging, FX derivatives — many finance people aren't incapable, they're afraid to do it lightly. Do it well, boss praises; do it poorly, you take the blame.
What AI can do here is first structure the complex information, then give strategy recommendations.
I understand AI in enterprises can be split into input layer, middle layer, and decision layer. The boss doesn't need to participate in every action; he only needs to set risk appetite first — aggressive, neutral, or conservative. The subsequent strategy can be generated by the system based on actual business and continue iterating based on historical execution results.
Why Hasn't Human Customer Service Been Replaced by AI?
Wu Wei: Mr. Wang, you serve many consumer electronics and smart hardware brands, across countries and languages. Why are people still in this chain? What work can't AI replace yet?
Wang Gengdong: We can look at changes in business structure.
When we started, we did pre-sales, email customer service, phone customer service. Now about 70% of our business is after-sales phone. Pre-sales is doing less and less because pre-sales, especially text shopping guidance, AI is already very suited. It processes text, accuracy can be high, and even if there's a small error, risk is relatively low.
But after-sales is different, especially extreme complaints and high-ticket products. One simple misjudgment might bring returns and bad reviews; more seriously, it could even become a brand PR event.
What we serve most now are smart hardware with complex products and high ticket prices. Customer service needs to operate the product repeatedly themselves, truly knowing how to troubleshoot and guide users when they have problems.
Our client structure these years is also interesting. Top clients in South China are increasingly concentrated in Nanshan — energy storage, laser engravers, robot vacuums, AI smart hardware, breast pumps, etc. — many are categories with high usage thresholds, complex products, and high ticket prices.
For these top brands, customer service isn't just a cost center. It also delivers brand experience and user mindshare. Many clients tell us directly: satisfaction metrics done well, cost isn't the first priority. The core is user satisfaction.
For example, we now serve some high-value home energy storage products, tens of thousands of dollars per unit. When a user has a problem, if customer service can diagnose and solve it on one call, it directly saves an engineer visit cost. At that point, customer service itself is creating value, not just consuming cost.
Our billing is overall still BPO logic: base labor fee plus performance, and one very core performance metric is satisfaction.
Why Does Offline Need a Shopping Guidance Robot?
Wu Wei: Albert, your shopping guidance robot isn't simply replacing salespeople. In offline retail scenarios, why does it need to exist?
Albert Lee: We actually started not doing shopping guidance robots but edge AI. Edge means minimizing dependence on cloud large models, because cloud models face latency, cost, and stability issues.
Later we kept researching what truly deployable scenarios edge AI has in vertical industries. One client was a large smart hardware experience center, a 2,000-3,000 square meter store with maybe a thousand tech products — robots, drones, voice recorders, smart instruments, etc.
Its biggest pain point is training. A thousand SKUs, every single one needs salespeople to understand and introduce. An employee might train two or three months to be barely useful, then leaves. This cost is very high.
So we combined edge AI, interaction capability, and physical hardware into a retail shopping guidance robot.
I used to do enterprise consulting, so I have a strong judgment on this: what enterprises truly need long-term isn't robots doing kung fu and dancing. Those things are enough to see once a year. What enterprises need is a role that continuously creates value every day.
So our robot does shopping guidance, product introduction, CRM, inventory management, bringing some capabilities from websites and mini-programs into offline physical spaces, while letting it directly interact with customers. We also add some light interactions, like mini-games, so customers want to stay.
Additionally, we're researching high-ticket, low-frequency scenarios, like medical aesthetics, beauty salons, hotels. Like hotel quick check-in — user acceptance of robot intervention might be higher.
Privacy is indeed a problem. Cameras, user recognition — not all scenarios are suitable. So ultimately you still choose scenarios; you can't do everything just because technology can.
How to Integrate Full-Chain Data? What Can Be Handed to Agents? Who Takes Responsibility?
Wu Wei: Finally, back to today's "full chain" theme. From market insight, marketing acquisition, sales conversion, to payment fulfillment and after-sales service, many enterprises' biggest problem now is data not connecting.
How do you think these data should truly connect in the future? Which already-fixed steps can be more confidently handed to Agents? Which things shouldn't be handed to AI? If something goes wrong, should responsibility ultimately rest with the enterprise, the service provider, or the model vendor?
Chen Guang: Software Infrastructure Must First Adapt to Agents
Chen Guang: Recently the FDE concept has been very hot; we actually started building FDE teams last year. The biggest feeling from doing it is that much software infrastructure hasn't truly adapted to the Agent era.
In the future, software vendors at every link need to become more Headless, opening capabilities in ways Agents can more easily call — CLI, API, MCP. This way, whether the client's own Agent or Agents from different service providers can more easily call and connect with each other.
Recently doing several large enterprise clients, after going deep into IT systems, huge amounts of time are spent coordinating Salesforce, OMS, ERP. To fuse consumer-related data into one platform, continuously doing system integration. This work is actually very time-wasting.
If everyone's capabilities can be standardized and called, with Agent documentation, it can complete many integrations itself. One of our focuses for H2 is also opening more capabilities to clients and partners.
I think the real core competitiveness of software vendors in the future isn't the interface. The interface can change from GUI to API, CLI, MCP; what's truly valuable is the capability model behind it: can you make a specific vertical scenario good enough.
As for what can be handed to Agents, my judgment is: processes in the enterprise that are already clearly decided and already固化 into SOP can be gradually let go to Agent execution. Of course, with safety guardrails, Eval, and regular regression evaluation.
But new product launches, new businesses, new failure types, batch product defects that suddenly appear, or other high-risk, sudden situations — I don't recommend letting the Agent decide itself. These things still need human judgment.
Who takes responsibility in the end? I think ultimately the client's own decision-maker takes responsibility. Model vendors and Agent vendors essentially provide tools. Tools can give recommendations, but whether to execute is ultimately the enterprise's own decision.
In Brand Marketing, Humans Still Need to Handle "Emotion" and Differentiation
Wu Wei: Andy, how do you see this from your marketing and advertising side?
Andy Peng: We've also been testing AI automatic advertising. Platforms like Meta are pushing more and more automated advertising features; content can also directly generate AI images, AI videos.
Agency will of course be anxious: one day will ad optimizers be completely replaced by AI? Our team's best optimizers might have personally spent tens of millions in ad spend; if platform automatic advertising is strong enough, is this position still needed?
But currently from our testing, it can't be fully replaced yet.
A core reason is that brands have emotion and differentiation. Truly great brands eventually form their own very unique aesthetic, expression, and sense of value; this part isn't solved by simply improving click-through rate.
We've actually stepped on pitfalls. Some crowdfunding platforms have clear restrictions on AI content; some brands also don't accept agencies using AI materials without confirmation. Now if we want to test AI images, we must first ask the client if we can test. Only with client consent do we run it.
Also influencer marketing negotiation. Bargaining itself is a highly human process. An influencer might already be annoyed by your negotiation, emotion visibly changed, but AI doesn't notice and keeps pressing, and the whole collaboration falls apart.
So every step needs judgment: which parts can be handed to Agents, and where humans must still be. At least at the current stage, in marketing the parts involving brand, emotion, and relationships, humans are still important.
If You're Still Worrying About "Who Takes the Blame," That Scenario Probably Shouldn't Use AI Yet
Zheng Zhikun: We're now doing payment and FX-related products, also talking with ERP vendors about data integration.
Why can't many things connect now? Technology is one aspect, but another very realistic problem is interests and boundaries. Everyone worries: are you entering my domain, am I handing over my core data? This is a natural data and business barrier.
As the market matures, I think everyone will ultimately see more clearly: data integration isn't about who invades whom, but about jointly serving sellers and letting enterprises make better business decisions.
As for the "blame" question, my view is direct: if an enterprise puts AI in a scenario today and the first thing it thinks about is who takes the blame if something goes wrong, then that scenario probably shouldn't use AI yet.
AI can already do so many things. Just in low-risk, clear efficiency scenarios, the space is already big enough; there's no need to put it somewhere with unclear responsibility boundaries from the start.
And many seemingly very "emotional" industries actually have very clear AI scenarios. For example, trading cards — essentially a very strong emotional consumer product — but AI can be used to identify card authenticity and assist appraisal. The value of such scenarios is very clear.
So it's not which industry can use AI, but finding the specific step with clear efficacy and clear responsibility boundaries.
"What We Give Up Shows What AI Can Already Do"
Wang Gengdong: Let me put it another way. Look at what we voluntarily give up, and to some extent you can see what AI can already do.
We now voluntarily give up some seller-type clients and some low-ticket categories. Because these clients compete on operational efficiency and supply chain efficiency, they don't need customer service to convey much emotional value to users. Like a $15 clothing item with a problem — direct refund might be the most economical solution, no need to invest heavy human service.
What do we hold on to? Top brands, high ticket prices, and high-complexity products.
These brands don't just earn product value-for-money money; they also earn brand premium and user mindshare money. At this point, between users and the brand there needs to be deeper contact, and humans still have value.
Strategy sometimes works this way: what you decide not to do tells you more clearly what you truly want to do.
Private Data Will Become Enterprises' Real Moat
Wu Wei: Albert, you have both offline hardware and CRM/member system involvement — how do you connect online and offline data?
Albert Lee: I used to do enterprise consulting and long worked on data, so I'm sensitive to this issue.
Now when we do projects, we often need to connect client CRM and member system data. One of the biggest problems is that clients worry deeply about data security: after connecting, will other clients get it? Will it leak?
So many of our retail projects go back to the edge AI we're good at. Clients can build their own local AI servers, isolate data, and let sensitive information stay in their own environment as much as possible.
Large models are strong now, but large models don't have your enterprise's own private data.
I think in the "software 3.0" era, enterprise private data will become a very important competitive moat. Only your own data can gradually train an intelligent system that truly understands your business, your customers, and your organization.
Is There an AI Bubble Now?
Wu Wei: Finally, one line each: do you think there's an AI bubble now? If so, what's the biggest bubble?
Albert Lee: I think there is, more or less. The biggest bubble is actually the cognitive gap between everyone's expectations for AI and its real capability boundary today.
People who don't really use AI much think AI is great at everything; but people who truly use it heavily every day increasingly understand what it can and can't do. AI is definitely useful, but don't overestimate the capability boundary.
Wang Gengdong: I think there's absolutely a bubble. Too many people doing it now, and many directions' thresholds aren't as high as everyone imagines.
What ultimately survives is still companies that, besides AI, already have unique business capabilities, industry resources, and hard-to-replace assets. AI combined with these is more likely to become long-term competitiveness.
Zheng Zhikun: There's a bubble, but a bubble isn't necessarily bad. When the economy is good, there's naturally a bubble.
My judgment is simple: AI that can actually be used in real scenarios and create value today isn't a bubble; something with no real use case today is a bubble in the present. Of course, it might become useful in the future.
Andy Peng: If I must name one, I think embodied intelligence currently has the biggest bubble. We've actually tested some projects; experience and real usability are far from expectations. Many products look busy but don't have features that can stably enter real use scenarios.
Chen Guang: I say "no bubble," mainly referring to AI software, not including robots. Like AI software, image generation, video generation — these directions, I think today there's still no talk of no space.
One reason is compute is still scarce. Many new models get rate-limited as soon as released, showing supply can't keep up with demand.
Second, doing large enterprise AI deployments, we increasingly feel: now model capability and AI capability are actually already strong enough; the real obstacle is often not the model, but how to reform old systems, how to integrate, how to connect data and business processes.
Once these bottlenecks are gradually broken through, actual token usage will keep growing. So I think AI software still has a lot of space. As for embodied intelligence, I think it still needs more time to find truly stable application scenarios.