Original · Unique Research / 非凡产研 · 2026-07-21 · Chinese source: https://view.inews.qq.com/a/20260721A0AKZE00
Editor's note: This is a complete English rendition of the source roundtable transcript. Speaker attributions, predictions, and company claims are retained as the speakers' own statements. Source images are not processed per task scope.
---
AI Industry Observation
Everyone has an AI now—why hasn't productivity gone up? The problem isn't the tool; it's the "alignment."
"POC tests whether the model is smart; production tests whether the enterprise is clear-headed."
One customs declaration task: a skilled worker spends 20 to 30+ minutes buried in it, and it's especially error-prone. Switch to an Agent, and dozens of tasks come in at once, all processed in ten-plus minutes.
With numbers like these, you'd think the project is a lock, right?
NoDesk AI actually built such an import/export customs declaration Agent. Co-founder Wang Fang (王仿) told me: frontline employees photograph export goods with their phones; the backend must extract dozens of fields from piles of images, Word files, and Excel sheets, then do standardization, weight verification, and finally generate customs declarations, packing lists, and contracts. The Agent does this work quickly and cleanly.
But the story has only just begun.
In the POC phase, the test photos the client provided were crystal clear. Once actually live, they discovered at least 10% of photos were so blurry you couldn't even make out a person.
More tricky is the customs scenario: one wrong character can mean a serious incident.
So the Agent must learn one thing: if it can't see clearly, it should honestly say it can't see clearly and hand the work back to a human, rather than forcing a guess.
"Auxiliary tools at 80% accuracy already wow people. But on the production line, if accuracy can't reach 90%+, it usually can't go live—one error equals one incident."
So what about the remaining 10%? Humans and Agents back each other up. Their approach is called "AI First": high-confidence tasks run automatically on AI, with every node visible to humans throughout; when the Agent isn't sure, it proactively hands off to a person. The final backstop is always the business staff.
Plainly, POC is just a low-cost mutual probing—the client can simultaneously test several vendors, and the vendor's investment is limited. Once in production, what's tested is something else entirely: value, accountability, people, process.
This may be the most cutting line of the entire roundtable: the biggest blocker for many projects is that business value was never defined before project initiation.
Wang Fang has seen too many such projects. In 2023 and 2024, when enterprises did AI, they loved picking admin assistants, HR Q&A, and internal knowledge bases. The effect could be tuned beautifully, but eventually the boss realized the whole company uses it a handful of times a year—and naturally won't keep investing.
Now AI is entering core scenarios like customer service; bosses feel there's value, but business departments start pushing back: what happens when lots of issues emerge after launch? When the large model makes an error, who bears the risk?
So before initiation, you need to argue not just value, but also the accountability mechanism. Missing that, no matter how pretty the POC is, the project can't enter production.
Here's another counterintuitive one.
Ouraca co-founder Zhang Qiming (张栖铭) reviewed his own team: when AI Coding first became popular, product used AI to write docs, design used AI to make images, engineers used AI to write code—every role seemed a bit faster, but overall development progress didn't improve much.
Break it apart, and the longest time sink wasn't execution—it was alignment.
"I need to explain my thinking completely to you; you understand it, then pass it to the next layer. Every layer waits for the upstream to write the doc clearly before starting." Execution speed went up, but the communication friction in the middle didn't decrease at all.
Later they changed how they work: hire programmers with startup experience who understand the business, so one person can take something from start to finish; managers get off their high horse and stop micromanaging. They also wrote many Skills as backstops: if you don't know what's next, run a Skill; first do competitor research, find benchmarks, then add your own understanding. The result? A ten-person full-time team where each engineer writes tens of thousands of lines of code per week.
"The tool didn't change; the organizational method changed, and output changed."
Wang Ting (王婷) of Linghe Shuzhi (灵核数智) serves manufacturing clients. She added a commonly overlooked blocker: people.
Many manufacturing enterprises still live on Excel; "表哥表姐" (spreadsheet jockeys) are everywhere, data lives offline, approvals go through paper documents. Deploying an Agent is only the first step; the real challenge is whether employees will work with it.
She gave an example. A client's spreadsheet format changed—with RPA, the spreadsheet breaks and dies the moment it changes; with AI, the AI also needs to know what happened in the business. This is where employees need to tell the Agent: this is the same old client, something changed, here's how you should adjust.
Sounds like a one-sentence thing. But many employees simply can't describe their own work clearly—can't articulate what changed or what they need the Agent to do.
Their solution is to cultivate an internal role called "AI Refiner" (AI炼化师)—not externally dispatched, but internally trained. The Refiner first articulates the workflow clearly, and AI directly generates the SOP; when business changes, the Refiner discovers the change, describes it, and directs the Agent to adjust.
Once this system runs, almost every role, every person at their company has at least three Agents. Take Wang Ting herself: before meeting a client, the Agent generates an interview outline; after the interview, the notes go to the Agent, which then outputs client requirements, solution, and technical breakdown—what goes to R&D is clear at a glance.
Results are concrete: in one case, the Agent helped a client eliminate 3 positions, corresponding to about 1.5 million RMB in labor cost. From sales order entry, order forecasting, to production scheduling, procurement, and shipping—the entire chain is running.
"It's not 'business plus AI'; it's 'AI plus business.'"
After people and organization, the most practical question: money.
Wang Fang's pricing is candid: knowledge-base products can be standardized, subscription monthly or yearly; but knowledge governance, internal system integration, requires implementation fees. One project is typically 300,000 to 500,000 RMB.
You might ask: how do enterprises measure whether these hundreds of thousands are worth it?
Some math is easy: e-commerce videos cost 500 RMB manually, 300 with AI; customer service, moderation roles with dozens or hundreds of people—how many fewer people, how many fewer hires after Agent launch is clear; during 618 and Double 11, you used to hire a batch of temporary contractors; now you can hire fewer or none.
Frankly, most customers measure ROI in one way: how much labor cost was saved. This also explains why Linghe's "saved 3 people, 1.5 million" case is so persuasive—the more direct the number, the easier the project to push.
Finally, a direction-selection story, quite interesting.
Shenyong Intelligent (深涌智能) didn't start in finance—it started in Infra: compute, models, data, all hard labor. After DeepSeek went viral in 2025, every industry started thinking about deploying large models internally. They dabbled in manufacturing too, and found the foundation too thin: OCR, spreadsheet digitization, knowledge base construction—all unavoidable dirty work. The boss pays 500k or 1M and must calculate how many people are saved—spending is painful, projects push slowly.
Later they pulled back the Infra capabilities they'd been selling externally and started trading themselves. The experience was completely different: finance has high digital maturity, and it's especially easy to evaluate.
CEO Huang Kecheng's (黄可铖) own words: "In letters there's no best, but in combat there is." When trading returns are on the table and good enough, other arguments disappear automatically. Now they batch-communicate with private equity funds in Hong Kong and Singapore; they don't start with tech, they start with results; if results are stunning, everything else is easy. Customers even ask back: beyond trading, can you do digital employees?
They now focus on Asia-Pacific and the Middle East—Singapore, Japan, Hong Kong. These markets' customers are more willing to pay for early-stage investment, and have more mature subscription habits.
"Same team, same tech, different exam room—from 'having to explain everywhere' to 'speaking through results.'"
At the roundtable's end, the host asked: if an enterprise starts an AI project tomorrow, what must be confirmed before initiation?
Combined, their answers are basically a health checklist:
Are you truly committed to this, or do you just feel "I want AI"? Is there an internal champion with influence and execution who can coordinate departments? (They've seen Agents already deployed, then at the cross-department testing step, relevant people simply don't show up.) Can every step's math be clear—not just cost savings, but organizational investment and process reform?
Further down are two slower things: are there rules and culture as backstops, telling employees what's mandatory and what the company encourages? Have you cultivated internally a cohort of AI-Native people, like Linghe's "AI Refiners," who understand the business, are willing to tinker, and can continuously teach changes to Agents? Finally, what AI handles and what humans handles must be drawn first, so employees feel safe.
Huang Kecheng's criterion is the most direct: look at the liaison the client sends. Execution, awareness, and voice—all three are required. If the other side sends someone who spends 90% of time on other work, only 10% cooperating, and keeps asking "what's the use of this thing," the project probably won't move.
"Technical problems can always be solved—models get stronger, and many problems you don't even need to solve yourself. But organizational problems don't disappear on their own. POC tests whether the model is smart; production tests whether the enterprise is clear-headed."
---
Speakers
Huang Kecheng (黄可铖), CEO, Shenyong Intelligent / Emerging AI (深涌智能)
Wang Fang (王仿), Co-founder & CTO, NoDesk AI
Zhang Qiming (张栖铭), Co-founder, Ouraca
Wang Ting (王婷), Chief Customer Officer, Linghe Shuzhi / Linkcross (灵核数智)
Host
Wu Wei (吴畏), Founder & CEO, Unique Research
Wu Wei: You've all helped enterprises deploy AI; you've encountered many blockers and pitfalls. Today I hope you share these experiences as honestly as possible, so other enterprises can avoid detours. First question: from POC to actually entering production, blockers can come from organizational structure, budget, or an enterprise's own IT/digitalization foundation being incomplete. If you could only pick one, what do you think is the most critical blocker?
Huang Kecheng: We are Shenyong Intelligent, English name Emerging AI. We help enterprises in China and overseas with one-stop AI deployment, currently focused on finance—overseas brokers, private equity, family offices, and other financial institutions. We also do some consumer products; the core is bringing Autonomous Finance Intelligence from top-tier institutions and quant funds down to non-top institutions and ordinary people. That's our vision.
I think POC and actually entering Runtime/production are two completely different things. In the POC phase, the client can engage many vendors simultaneously; we can do POCs for many clients simultaneously. Both sides' resource investment is limited—essentially a lower-level resource matching. But once truly in production, the requirements on both sides are much higher. How much information the enterprise is willing to open, how many team resources they can invest, whether both sides have enough awareness and trust, whether teams can effectively collaborate—all become blockers. If I had to pick one, I think the most critical is enterprise resources, including data, compute, and how far the enterprise is willing to open on security and trust.
Wu Wei: That may relate to serving finance—finance demands extremely high trustworthiness and precision; many steps can't have any error. Wang Fang, what's your view?
Wang Fang: I come from an algorithm background, previously doing search/recommendation and NLP at Alibaba and Sogou. I left Alibaba in 2023 and joined Zhipu; I was among the early group in China working on enterprise large-model deployment. Now NoDesk AI's team is mostly alumni from Zhipu and Alibaba, focused on domestic To-B enterprise AI deployment.
From our practice, the biggest blocker from POC to production is: business value was never clearly defined before project initiation. Even if the POC effect is great, when it comes to actually continuing investment and going to production, people still feel it's not that valuable, and the project stops. Especially in 2023-2024, many enterprises picked admin assistants, HR Q&A, or internal knowledge assistants. You can tune the effect well, but the boss eventually finds people use it a few times a year, and won't keep investing. Now AI is entering more core scenarios like customer service; bosses usually see value, but internal people may not agree. Business departments keep challenging: POC looks okay, but what if lots of problems emerge after launch? What if the model errs? Who bears the risk? So before initiation, you must argue not just value but also the accountability mechanism. Otherwise the project can't truly enter production.
Wu Wei: So you can't just try because "everyone is doing AI." Before initiation you must be clear: what business problem does it solve, what value does it create, and who ultimately bears accountability. Zhang Qiming, what's your view?
Zhang Qiming: I'm from Ouraca, short for "Our Academy." The company initially wanted to build a university for the future. After large models appeared, we felt many things kids used to learn may no longer need to be learned the old way, because models already do them better. Our final conclusion: in the future, people should learn how to use Agents well.
Our first product is called Library. We believed building a university should start with a library. We turned many books into podcasts, then delivered explanations based on user personality. After launch, we interviewed many users. Many told us: "I don't actually like studying or reading. I read because I need to solve problems in life and work." This hit me hard.
Later we tried extracting knowledge from growth books into Skills. Users hand their documents to the Agent, which uses these Skills to tell them: which features should be built, which shouldn't, which customer segments to focus on. He doesn't need to finish the book; he can directly apply its methods to solve problems. So we gradually shifted from "helping people learn knowledge" to "teaching people how to use Agents well."
In this process, we interviewed many enterprises, including internet companies like Alibaba and ByteDance, as well as manufacturing and supply chain companies in the Jiangsu-Zhejiang-Shanghai region. Personally, I think enterprise AI deployment has two important blockers. The first is whether the top leader's awareness is in place. The enterprises we contact are all very FOMO, but different bosses' understanding of AI varies widely. Some bosses think it's enough to "distill" existing employees' experience; but some manufacturing bosses already realize AI isn't simple efficiency improvement—it's a redesign of technology roadmaps and production methods. For example, some 3D printing companies feel old mechanical design methods no longer apply; AI needs to redo the entire mechanical design process. That's not optimization on old knowledge; it's something disruptive. Internet companies currently care most about how to truly use AI Coding. But using AI Coding well involves not just giving programmers a tool—the entire organizational method must adjust. Going-global e-commerce companies focus on cost reduction and efficiency improvement—how to acquire more users, do SEO and growth through AI. So different industries' core needs are completely different; you can't cover all enterprises with one uniform solution.
Wang Ting: I'm from Linghe Shuzhi, product called Linkcross, mainly helping manufacturing enterprises with AI digital transformation. Our founder comes from a manufacturing family with 40 years of history and once led the entire enterprise's digital transformation. He found Chinese manufacturing has huge room for improvement. Many manufacturers still rely on Excel—the "spreadsheet jockeys" everyone mentions; data is offline; some approvals are still on paper. We believe AI is an important opportunity to transform manufacturing. Our slogan is "Help Chinese manufacturing remain great." It may sound grand, but for us it's a long-term starting point. I'm Chief Customer Officer, mainly responsible for customer-side AI deployment, so I feel these issues sharply. The previous speakers mentioned resources, business value, and top-leader awareness. I'll add a commonly overlooked issue: talent transformation. FDE—frontier deployment engineers—is popular now, but deploying AI into an enterprise is only the first step; what truly matters is whether it can be continuously applied after deployment. Whether an enterprise can sustainably use Agents hinges on whether employees can collaborate with Agents.
Wu Wei: What capabilities do existing employees need?
Wang Ting: First, they need to learn to converse with the Agent. Think of the Agent as a new employee entering the enterprise. After being deployed to a role, it still needs continuous learning and growth. I just saw a client yesterday. They had an Agent already deployed and running for a while, then the client's spreadsheet changed. The old workflow ran through RPA; once the spreadsheet changed, RPA broke. With AI, it still needs to know what business changes occurred. At this point, employees need to tell the Agent: "This is a certain old client; here's what changed; you need to adjust your capabilities to adapt." Sounds like a simple sentence, but many employees don't know how to describe it. They can't even say where the change happened or what they need the Agent to do.
We solve this in two ways. First, we help enterprises establish the "AI Refiner" role. This isn't about us dispatching someone; during the building process, we work with the enterprise to find and cultivate the right internal people. I previously worked at ByteDance; in 2021 I founded the "Efficiency Pioneer" program. In my current role, I increasingly find that one of the most important future capabilities is the ability to converse with AI and to distill experience. So we call these people AI Refiners. An AI Refiner must first learn to decompose their own work and articulate the workflow. In fact, as long as an employee can articulate the process, AI can directly generate the SOP. Our system combines SOPs and Skills to let the Agent start executing. The core capability employees need is: when business changes, they can discover the change, describe it, and tell the Agent how to adjust. Without training, many employees indeed can't describe their work clearly. But after this kind of training, people's thinking also changes. We're running an internal "Refiner Competition." Now almost every role, every person at our company has at least three Agents.
Take my work: starting from client research, I only need to tell the Agent: who is the client, what industry, who am I meeting today, what does this role own. The Agent generates interview questions. After the interview, meeting notes go to the Agent; we've already written the downstream SOP. It outputs client requirements, solution, and technical breakdown: which capabilities the platform already has, which are new technical blockers needing R&D. The whole process is basically AI-completable.
Wu Wei: Which parts of manufacturing have you deployed?
Wang Ting: From sales order entry, order analysis, order forecasting, to production planning, production scheduling, then procurement, quoting, and shipping—we cover all of it. Our approach isn't having one Agent replace one isolated role; it's using Agents to help enterprises reshape the entire business process. It's not simple "business plus AI"; it should be "AI plus business."
Wu Wei: In one case you mentioned, your Agent helped a client save three people, corresponding to about 1.5 million RMB in labor cost. This kind of directly quantifiable result is very important for enterprises judging project value.
Wu Wei: Wang Ting just mentioned employee capabilities and organizational transformation. Zhang Qiming has long worked in education—how does your internal talent and organization adapt to AI?
Zhang Qiming: Let me use the internet industry as an example. Internet companies may be more extreme in AI transformation than manufacturing, so the characteristics are more obvious. Our own team initially used the traditional R&D model: upper layer sets strategy, product managers write PRDs, designers produce drafts, engineers write code, testers test, layer by layer downward. When AI Coding first became popular, we thought product managers could use AI to write docs, designers to use AI for design, programmers to use AI to write code. Result: every role looked a bit faster, but overall development progress didn't improve much.
Later we broke the whole process apart and found the longest time sink wasn't execution—it was alignment. I need to explain my thinking completely to you; you understand, then pass to the next layer. Every layer pursues certainty: wait until upstream writes the doc clearly enough, then I start the next step. So each layer's execution speed improved, but the communication and alignment time in the middle didn't decrease; organizational friction remained very high.
Later, when I had an idea, I'd first use AI to quickly generate a page or demo, then take the concrete thing to communicate with the team. That was indeed faster. But new problems appeared. During the process, frequent modifications were needed—a little change here, a little there. If you keep letting others follow your modifications, they'll "lie flat," feeling you must fully think through the requirements before handing them over.
Later I hired several programmers with startup experience. I found these people have stronger business understanding. Once they understand the business, they can independently complete many things and form a closed loop. So we began deliberately cultivating everyone on the team: get used to generating something from one sentence, understand the business yourself, make your own judgments. We also wrote many Skills to assist this process. For example, when you don't know what to do next, run a Skill. We also formed some SOPs, including first doing competitor research, finding benchmarks, then adding our own unique understanding.
But in this process, managers also need to change. If managers micromanage—today require changes here, tomorrow there—employees will lie flat again. So managers must let go of their desire for control and lower their ego. Second, every colleague must truly understand the business. If they can't understand yet, use Skills and moderate communication to help.
Wu Wei: With current Coding Agent and model capabilities, if one person has entrepreneurial initiative, what's the maximum scale of project they can self-complete?
Zhang Qiming: Personally, as long as it doesn't involve particularly deep underlying technology or algorithms, most work can be done by one person. In manufacturing, besides needing specialized algorithms for mechanical design, most other business systems can basically be done through AI now. We currently have ten full-time colleagues. Excluding operations and external-facing roles, each core R&D colleague completes tens of thousands of lines of code per week. Compared to last year, model capabilities are stronger now. Writing code used to often produce weird issues; now those are fewer, and basically after one or two "card draws," results are usable.
Wu Wei: NoDesk AI currently focuses on e-commerce retail and manufacturing. Can you discuss a concrete case of how an enterprise truly deploys an Agent into the business?
Wang Fang: We recently built an import/export customs declaration Agent. Customs declaration sounds standardized, but actual execution is very complex. Frontline employees use cameras or phones to photograph various export goods, sending lots of images, Word, and Excel files to the backend. Backend staff must extract dozens of fields from these complex materials, query related information in internal systems, standardize the fields—including entity alias alignment, entity conversion for different export countries, weight verification, etc. After verification, these fields must be organized into standard customs declarations, packing lists, and contracts. Previously, manually completing one customs declaration usually took 20 to 30+ minutes, and was very error-prone. Now with the Agent, dozens of tasks at once can be processed in about ten-plus minutes. The frontend is perception and understanding layer: AI automatically understands image content, extracts data from Excel and other documents, then calls different systems for verification and document generation.
But production is completely different from POC. In POC, everyone provides good-quality photos. Once live, photo quality rapidly degrades—at least 10% are so blurry you can't even make out a person. At this point, the Agent must know: "I can't see clearly; this needs a human." Rather than forcing a parse. Because in customs, even one wrong character can cause a serious incident.
Wu Wei: So the process itself is standard, but the execution is highly non-standard.
Wang Fang: Right. To take an Agent to expert level—or 90% of expert level—the people building the Agent must themselves become experts in the domain. You need deep understanding of the Agent's goals, boundaries, exception handling, which systems and Skills to call. Only when these are defined clearly enough can the Agent run on the production line. An auxiliary tool at 80% accuracy already impresses people. But once on the production line, if accuracy can't reach 90%+, it usually can't go live, because one problem can equal one incident. What about the remaining 10%? Ultimately humans and Agents back each other up. Our model is called "AI First": high-confidence tasks execute automatically, but the whole process must be controllable and observable, with every node visible to humans. When the Agent's confidence is low and it feels uncertain, it proactively hands the task to a person. The final backstop is always business staff.
Wu Wei: Are these projects standardized products, or do they require deep customization?
Wang Fang: Partly standardized, partly customized. Knowledge base products can be standardized, SaaS-like pricing, monthly or yearly subscription. But knowledge governance, internal system integration, and connecting these systems to Agents usually require implementation fees. Currently, one project typically costs around 300,000 to 500,000 RMB.
Wu Wei: How do enterprises measure this 300,000 to 500,000 RMB investment? What's the expected output?
Wang Fang: Some scenarios are easy to calculate. For example, e-commerce video generation used to cost 500 RMB manually per video; with AI it may drop to 300—the cost change is direct. Some labor-intensive roles like customer service and various moderation roles may have dozens or hundreds of people. After Agent launch, enterprises calculate how many people are needed, how much labor can be reduced, or whether additional hiring can be avoided. E-commerce during peak seasons like 618 and Double 11 usually requires lots of temporary contractors. With AI, enterprises may not need to add as many contractors during peaks. Currently, most customers still use reduced labor cost as the primary way to measure AI investment ROI.
Wu Wei: Finance has extremely high requirements for precision, security, and compliance rules. Why did you ultimately choose this direction, and why did you first see results in overseas markets?
Huang Kecheng: We didn't start in finance on day one. We started in Infra—compute, models, data, efficiency optimization, all hard parts. In 2025, after DeepSeek appeared, an obvious change was that every industry started discussing Enterprise AI and considering deploying large models internally. All-in-one machines were also popular; essentially all to solve local deployment and business application problems. At that time we touched many industries, including e-commerce and manufacturing—especially manufacturing. Our founding team is relatively young, with MIT, Peking University math, Tsinghua backgrounds, plus some big-company experience. The team is campus-like; many have overseas experience. This team has two characteristics: first, it pivots quickly without strong industry lock-in; second, everyone has relatively strong math foundations.
When we did manufacturing, we felt deeply: manufacturing is truly a great thing, but once inside, you find lots of dirty work unavoidable. In POC, enterprises don't invest much; project acceptance is relatively easy. Some overseas clients have more mature commercial environments and are willing to pay part of POC costs. But once truly in production, enterprises find they need employee training, capability building, plus massive OCR, spreadsheet digitization, knowledge base construction, historical knowledge extraction. The whole process sounds very troublesome. If the boss needs to pay 500,000 or 1 million RMB, he will definitely calculate how many people are saved and how much value is generated. But spending is always painful; such projects push slowly.
Later we found financial trading scenarios are different. Our team has strong math foundations, so we took the Infra capabilities we'd been selling externally and brought them back to do our own Training and trading experiments. The first impression finance gave us is high digital maturity. Second, it's very easy to do Evaluation. "In letters there's no best, but in combat there is." In financial trading, if returns are good enough, they basically override other arguments. Now we batch-communicate with private equity funds in Hong Kong, Singapore, etc.; the process is fairly fixed. We don't need to spend lots of time explaining technology upfront; we put returns on the table first: is the result impressive enough? If the result holds, other questions are easy to discuss.
Wu Wei: So, what you initially built was something like an AI quantitative trading system.
Huang Kecheng: That's the starting point of all our business. Finance itself is a very end-to-end system. We didn't initially intend to become Finance Bros, but after building the main chain, you find lots of capabilities can spill outward. Including underlying Infra, middle-layer data, factors—each part may directly correspond to a profit center for private equity, private banks, family offices, or brokerage departments. If an institution lacks a certain link and we happen to have that capability, communication becomes much easier. Customers may first see our trading results, find them attractive, then ask: "Besides trading, we have digital employee problems—can you do that?" At this point the whole business order flips, and communication becomes much smoother.
Wu Wei: Does this also determine you'll enter overseas markets more?
Huang Kecheng: Yes, and it relates to team DNA. The US market, especially Enterprise Level projects, is harder than before. We currently focus on Asia-Pacific and the Middle East, with Singapore, Japan, and Hong Kong as relatively stronger entry points. These market clients are more willing to pay for early-stage investment and have more mature subscription habits. If the business model shifts from one-time payment to commission, projects are easier to negotiate. Of course, Chinese teams going overseas also have their own issues. Customers care about the relationship between overseas business and the China-based team/tech architecture. Especially in finance, this involves data compliance and financial compliance. Even if you don't apply for your own license, you must comply with different markets' regulations.
Wu Wei: Have you considered becoming a licensed financial institution yourself in the future?
Huang Kecheng: People ask this often. Running a traditional financial institution isn't our strength. We can put some money in, or let existing shareholders put some in, but if you ask us to truly become a traditional financial institution, doing sales and growth like Finance Bros, we're not good at that. We'd rather keep moving along the trend of AI advancing financial democratization.
Wu Wei: Last question. If an enterprise says it's starting an AI project tomorrow, what's the most important thing to confirm before initiation?
Wang Ting: One thing may not be enough. Based on our enterprise delivery experience, at the organizational level at least three questions must be confirmed. First, is the enterprise truly committed? Many enterprises just feel "I want AI" but don't know what problem to solve, and it's unclear who decides whether to start. Second, who is the internal guide and driver? This person needs enough influence, execution, and some decision authority. The boss is usually too busy for daily driving, but someone inside must be able to coordinate departments. We've encountered Agents already built and deployed, but at the cross-department testing step, relevant people simply didn't show up. The project got stuck. So clarify: who decides, who drives. Meanwhile, the enterprise needs to cultivate a cohort of AI-Native talent—select people inside with 3-5 years of work experience, some business experience, willingness to accept new things and continuously improve. That's why we built the AI Refiner system. Third, the enterprise needs rules and cultural safeguards. People need clear incentives and constraints. Rules tell employees what must be done; culture tells them what the company advocates, encourages, and needs. When decision-maker, driver, talent, rules, and culture are all in place, enterprise AI deployment has a relatively good starting point.
Zhang Qiming: I'll add one: ROI. Enterprises must make every step's ROI clear. Can this task be quantified? Can the final result be accepted? Can all costs from AI investment, human backup, and organizational reform be fully accounted for? ROI here isn't narrow financial ROI but a broad ROI. Beyond direct cost and revenue, you must calculate organizational involvement, process reform, and organizational gains. If the math is clear, the boss's decisions and organizational driving become much easier.
Wang Fang: I'll add from the organizational relationship angle. AI deployment has a big impact on the organization. The boss usually wants it, but frontline employees and middle managers are conflicted. Many things frontline employees used to do, Agents can indeed do now. Employees' value and scarcity may be diluted by AI. Even if the enterprise calculates the financial math clearly, when actually pushing down, you still hit strong resistance. So the enterprise must answer upfront: what does AI handle, what does human handle? Where will human scarcity manifest in the future? Only when employees know their relationship with AI and have basic security during deployment can the project truly move forward.
Huang Kecheng: Technical problems can ultimately be solved. As AI gets stronger, many problems don't even need us to solve ourselves; foundation models can help. My deepest insight is: if I'm the vendor, I hope the client has a counterpart inside who can collaborate with me on equal footing; if I'm the client and truly want to drive enterprise AI transformation, I can't place all hope on the vendor. There must be a person or a small group inside with enough force to push this—think of it as an AI transformation team. We judge whether a project should be temporarily abandoned based a lot on the liaison the client sends. For example, if they assign someone who doesn't truly understand the business to lead AI transformation, who spends 90% of time on other work, only 10% cooperating, and keeps asking "what value does this bring"—such projects usually can't move. The liaison must simultaneously have three things: execution, awareness, and voice—all three required.
Wu Wei: From POC to true production, technology is only one part. Enterprises also need clear business value and ROI, establish appropriate decision and driving mechanisms, find leads with execution/awareness/voice, handle the human-AI relationship, and keep Agents running through rules, culture, and talent cultivation. If these problems aren't solved in advance, no matter how good the POC effect is, it's hard to truly enter production.