Original · Unique Research · 2026-04-01 · Shanghai
Editor's note: This historical roundtable includes both the source's six-section overview and its extended transcript; their repetition is retained. The introduction and overview place Nexu's launch late last year, while the transcript says last week and the moderator says this year. All versions remain as written; no launch date has been inferred. The source calls the earlier trust chain three-party without identifying all three parties. It pairs a Chinese tax-confirmation expression with “Revenue Recognition”; that mismatch is retained, not presented as equivalent accounting terminology. Participant names, Yiyan Technology, Yunti Technology and moderator handle Lajiao are provisional English renderings. Product functionality, local-only data handling, compliance, staffing, office counts and efficiency are speaker/source claims, not independently verified or statements of current availability. Cultural comparisons are the speaker's examples, not universal descriptions of national teams; “human quality” and the brain/circulation analogy are metaphors. Future agent decision-making and industrial-software timelines are forecasts, not completed capabilities. Ten source images still require content review.
Unique Awards
AI Is Reshaping Global Collaboration,
but What Separates “Real Empowerment” from “False Empowerment”?
A roundtable bringing together four industry practitioners explored what it takes to put AI to work in global collaboration.
At a Unique Awards · Hangzhou AI WEEK roundtable, four practitioners from different fields sat down together. Their topic was clear: has AI genuinely helped businesses collaborate globally? If so, why do some companies feel empowered while others feel it creates more trouble?
Li Jinwei / CMO, Refly.AI—an entrepreneur building open-source AI Agent products; the introduction says he launched the open-source digital-employee product Nexu late last year.
Wang Pengfei / Co-founder, Mapping Intelligence—focused on connecting industrial data silos; proposed the concept of an “AI universal plug.”
Liu Yuchen / Founder & CEO, Yiyan Technology—a frontline practitioner of global collaboration for internationally expanding companies, with business in the United States, South America, Europe and Southeast Asia.
Sean Li / Billing Product Lead, Airwallex—a product operator at a global fintech company, responsible for revenue-management and commercialization products.
Moderator: Lajiao / Partner, Unique Research
1. Three “Mismatches” Lie Behind Companies' Difficulty Using AI
Li Jinwei's observations pointed to three practical problems.
First, a capability mismatch. A company introduces professional AI software expecting employees to use it immediately, but in reality they must write their own Prompt, configure workflows or even possess low-code skills. The result, he said, is that 80% of the company cannot use it.
Second, a mismatch in granularity. The company wants a comprehensive strategic upgrade of its AI capabilities, while individual employees often need to solve only one or two specific tasks. The scale of the ambition and the need do not match.
Third, a trust mismatch. Many AI products are cloud-based SaaS. As companies use them more deeply, core data starts flowing outside, forcing them to abandon the products. Demand for private and local deployment has long been constrained.
Based on those three judgments, the narrative says Refly launched the open-source product Nexu late last year, emphasizing three things: deployment simplified to one-click installation, AI as “an intelligent assistant inside your chat software,” and all data staying local without passing through the cloud.
“
You do not need to begin with complex scenarios. First let the intelligent employee do one or two concrete things, such as collecting group-chat data each day and automatically generating weekly reports. As it gradually understands your context, its value emerges.
2. Data Silos Are Not Just a Technical Problem, but a Cultural One
Approaching the issue from industrial manufacturing, Mapping Intelligence's Wang Pengfei offered a counterintuitive observation: data silos are not just disconnected systems. They also reflect different cultural rhythms.
He gave an example: Chinese and European R&D teams work very differently. German teams favor “write the documentation before executing,” while Chinese teams lean toward “rapid iteration and immediate feedback.” In his account, those differences create friction in cross-border collaboration every day.
Mapping Intelligence first helps companies connect the “blood vessels” between systems that do not communicate, such as ALM, PLM and MES. Once data flows locally, an AI Agent can learn what style of output teams with different cultures need—formal change documentation for a German team and iterative, immediate feedback for a Chinese team.
“AI is not just a tool,” Wang said. “It gradually develops a 'human quality.' That's one of the things we find increasingly interesting in industry.”
3. Have We Cleared the Language Barrier? The Real Bottleneck for Overseas Teams
Liu Yuchen of Yiyan Technology raised a problem many companies expanding abroad hesitate to discuss publicly: language.
“Even when everyone speaks English, accents and ways of expressing things are completely different. Our Chinese engineers are technically very strong, but their English writing is weak. Frontline colleagues can communicate smoothly with customers, yet struggle to convey requirements accurately to the back-end team through a high-quality PRD.”
They had tried various solutions: buying translation-enabled glasses for traveling colleagues and using different translation plugins. But these treated symptoms rather than the root cause.
Liu offered another interesting observation: sometimes AI tools are so powerful that they increase collaborative friction. “Some colleagues finish a meeting and ask AI to generate a Summary and PPT straight away, without digesting or organizing the material themselves. The other side receives a mass of unprocessed information, and communication becomes less efficient.”
His current approach is that people must personally check critical points. AI is an assistant, not a replacement.
4. AI Efficiency Gains Are Real, but the Chain of Trust Has Grown Longer
Airwallex's Sean Li shared his experience in finance.
In his view, AI's efficiency gains are clear: from automated account-opening approvals and transaction-level risk controls to optimizing financial expenditure, the benefits are visible. But risk has genuinely increased too—essentially because the chain of trust has lengthened.
Under the traditional model, the company trusts employees, who execute approvals; the source describes this as a three-party chain of trust. With AI, it becomes five parties: company → employee → AI → customer, potentially also represented by AI → regulator. There are more links, and in this account each link's risk factor rises.
“This is also the 'robots versus robots' topic that AI researchers have been discussing,” Sean Li said. “It's quite interesting.”
He also identified a nontechnical cost: as everyone in the company eagerly tries AI tools, IT and information-security teams bear the greatest burden, assessing each tool's impact on financial-grade data security. “The process adds management costs, but it has compounding value over the long term.”
5. Workflows Are Not Obsolete. The Barriers Are Disappearing.
One thought-provoking clash of views at the roundtable concerned whether products such as OpenClaw meant the workflow era was already over.
Li Jinwei's view was that the era of people manually orchestrating workflows was passing, but “workflows” themselves would continue. “Whether it's first-generation Workflow or today's large language models, the underlying dependency is still on workflows. There is simply less orchestration that people need to do themselves.”
He prefers to define Refly's products as “digital employees” rather than “tools.” A digital employee must fit into a company's everyday operations, understand context and help execute production-grade tasks—not be a universal assistant that constantly needs training.
Wang added an industrial perspective: over the next three to five years, industrial software would not become fully “AI-native,” but would remain mostly “AI-enhanced plugins.” In ten years, however, he predicted, an industrial-grade native operating system would certainly emerge.
6. The 2026 Outlook: Vertical Specialization, Localization and Agent-First
In the closing segment, each of the four guests shared, in one sentence, the 2026 trend they were most optimistic about. Together, their four judgments formed a complete picture:
Li Jinwei: A shift in thinking “from building tools to building digital employees.” Localization and open source will become clear trends because protecting high-value private data is a real need, while the true barrier to entry is the context accumulated over time.
Wang Pengfei: Companies expanding overseas do not know how to develop those markets? Replace traditional CRM with AI and build an AI CRM solution. In his view, no standardized product yet exists in this area, but one will certainly emerge.
Liu Yuchen: A new trend in software expansion abroad in 2026 is “software localization”—not just translation, but local support channels, local development of some features to meet compliance requirements, and local data storage.
Sean Li: Product builders should extend YC's principle from “Make something people want” to “Make something agents want,” because agents will increasingly act in place of people's habitual operations.
More from the Roundtable
Lajiao: Welcome, everyone! Our roundtable theme is “New Opportunities and Challenges in Reshaping Enterprise Collaboration.” Before we begin the discussion, let's introduce ourselves and our companies, starting with Mr. Li.
Li Jinwei: Hello, everyone. I'm Li Jinwei from Refly.AI. Refly was actually our company's first product, launched last year. We began with open source and launched a cloud-deployed version last December. Our original motivation came from how we understood AI product forms: Agent products have a low barrier to entry and can be driven through natural language, but lack sufficient controllability and stability; Workflow products, meanwhile, have too high a barrier. We therefore wanted to build AI-native software driven through an Agent's natural-language interaction. Last week, we launched our second open-source product, Nexu. We hope people will follow it on GitHub and contribute. It combines our underlying workflow-orchestration capabilities with the OpenClaw product form, allowing one-click local deployment of stable AI digital employees for everyday tasks.
Lajiao: OK, Mr. Li has just introduced the AI agent project Nexu that you launched this year. Since today's topic is AI's role in global collaboration, what customer pain points do you encounter when connecting business needs and putting AI into global collaboration, and which can your product solve now?
Li Jinwei: Three pain points stand out most from what I've observed and experienced. First, people debate whether AI tools provide real or false empowerment. That they can empower businesses is beyond question—but why do some companies struggle to use them?
The first is a capability mismatch. A company wants professional, highly capable AI software, but employees often need to configure their own Prompt and workflows, or even possess low-code skills. That leaves 80% of the company unable to use it.
The second is a mismatch in granularity. Strategically, the company wants to implement an AI system that upgrades everything. Each employee, however, often just wants to solve one or two discrete tasks. That mismatch makes it difficult to put into use at the employee level.
The third is trust. Much AI software now comes as cloud SaaS. Deeper use of workflows involves private data, which companies often do not want flowing to other platforms. They need products that can be privately and locally deployed.
We designed the open-source Nexu around those three issues. First, simplify the capabilities and remove the complex UI of traditional AI software, making it an intelligent assistant in the chat software you already use, such as WeChat. If you can chat, you can use it. Second, do not begin with extremely complex scenarios. Deploy the intelligent employee on one or two specific problems, such as collecting group-chat data daily and summarizing weekly reports. Finally, we launched an open-source, locally deployed product: all data stays on the customer's premises without passing through the cloud, addressing privacy concerns.
Lajiao: Understood—a very good product. Thank you for sharing. Next, Mr. Wang Pengfei: in your own business experience implementing AI for global enterprise collaboration, what customer pain points or problems exist, and what new approaches will address them this year?
Wang Pengfei: Let me briefly introduce our company. Our team started developing SaaS for R&D management in the automotive industry in 2021, under Yunti Technology / Mapping Space. Last year, we saw a new opportunity in AI and founded a second company, Mapping Intelligence, focused on “AI plus international expansion.”
Industrial manufacturing has many data silos. R&D uses ALM systems, while production and manufacturing use PLM and MES systems, and connecting them is very difficult. Businesses trying to connect R&D through to production face three challenges: high integration costs, the difficulty of understanding the entire industrial process, and the security red line that core data must not leave its designated boundary. Around those issues, Mapping Intelligence built a product like an “AI universal plug.” Our own platform connects ALM, PLM and MES so enterprise data can flow safely on-site.
Lajiao: An “AI universal connector”—that is indeed an interesting product. Mr. Liu Yuchen, do similar problems arise in Yiyan Technology's current operations and global AI collaboration, and how does your product address them?
Liu Yuchen: We are actively expanding globally, with offices and operations in the United States, South America, Europe and Southeast Asia. We have encountered many problems, such as managing global teams and collaborating seamlessly. The biggest right now is still language. People may all speak English, but accents and varieties differ. In particular, our engineers in China are technically strong but weak in English, while frontline staff—such as presales colleagues in Singapore or an FTE in Argentina—communicate well with customers but cannot convey requirements to the back end through a strong PRD. There are many translation tools now. For example, I bought translation-enabled glasses for a South American colleague visiting China, but these treat symptoms rather than causes. Better solving language problems in collaboration for overseas teams is a key issue, I think.
Lajiao: Thank you. I think Airwallex has considerable experience to share here, because by this year it has around 40-plus Office locations worldwide, as described in the source. Sean Li, what pain points arise when employees inside and outside Airwallex collaborate globally, and how do you solve them?
Sean Li: Thank you. Let me introduce myself: I'm Sean Li. I was previously an entrepreneur and now lead Billing—revenue management and commercialization—at Airwallex. Airwallex is a global fintech company whose vision is to build a global Financial Operating System, helping customers manage income, expenditure and cash flows. We aspire to become the next era's HSBC. Our internal philosophy is “make payments invisible and growth visible.”
On implementing AI, we have more than 1500 people, and everyone is very actively experimenting with AI in the current environment. What we do is complex and involves professional barriers—we handle money and regulation—and it is global, so communication and explanation are costly. AI can help build bridges, simplify communication and make complexity simpler. Our sales and business colleagues feel this strongly, not just product and R&D.
Lajiao: Understood, thank you. Many industries now promote using AI for internal efficiency, but some companies report that investing in AI has not necessarily improved efficiency and has actually added enormous costs. Have you experienced this in your businesses, and how have you addressed it? Let's start with Sean Li at Airwallex.
Sean Li: I think it happens, but the calculation is quite clear. The swings in efficiency are considerable. From my perspective, the only cost may actually reflect something positive: people are enthusiastic about AI and keep trying the latest enterprise tools. The people who feel the most pain are our IT and information-security teams, who assess what this means for financial-enterprise data security. I think it is a good process. Even though no one has fully worked this out from an information-security perspective in this era, you can see it as an added cost—but it has compounding value over the long term.
Lajiao: Understood. Airwallex operates at scale, and pure financial activities become more complex in anti-money-laundering and fraud scenarios. Sean Li, another question: how do you view the boundary between efficiency and risk control when using AI?
Sean Li: Let me continue. From an efficiency perspective, the clearest example for a global financial system is account opening. It used to require manual approvals, whereas agents can now make better judgments. From that through identifying signals at the transaction level to design risk controls, and then optimizing financial expenditure, the efficiency gains are clear.
But it is a double-edged sword. In the traditional relationship, the company trusts employees to approve things—a three-part chain of trust, as I describe it. With agents, the chain lengthens: the company trusts employees, employees pass that trust to an agent, and a merchant opening an account may also use an agent. The chain expands from three participants to five, and the risk factor naturally rises. The key is avoiding the additional risk introduced by these two AIs. This is the “robots versus robots” topic that AI researchers have long discussed. It is quite interesting.
Lajiao: Thank you. Returning to the earlier topic, Mr. Liu Yuchen: from your business experience, what might explain what people call AI's “false empowerment”?
Liu Yuchen: I do not entirely agree with the term “false empowerment.” I feel there really is empowerment. For example, after we introduced OpenClaw for employees, someone with a humanities background wrote an automation skill themselves: transcribing meeting recordings, then automatically generating reports with a PPT template. That went far beyond my expectations.
If there is an area with some false empowerment, it may be Coding. When AI writes a lot of code, engineers must spend more time on Review to make sure large projects do not take on risks. Controlling it through scaffolding—Harness Coding—so AI does not go wrong during prolonged reasoning, and having engineers write clean, clear Agent and Prompt definitions, are challenges right now.
Lajiao: Understood. Mr. Wang, what conclusions or thoughts do you have here?
Wang Pengfei: The industrial development path is quite clear: stage one is an “automation assistant,” stage two an “intelligent analyst,” and the third, ultimate form is a “process decision-maker.” If a problem in industry causes large-scale factory returns or a production shutdown, the cost is enormous, just as in finance. So we cannot rely entirely on AI to make decisions now. We are still in a stage of human-machine collaboration. Ultimately, it is a risk-control mechanism: important work choices still go to people for judgment.
At first, that means added management costs, because both digital assistants and intelligent analysts need people to manage them. But at the third stage, it can make judgments and decisions itself. That is also the ultimate form of Germany's Industry 4.0, in my view. We are eager to see how far it evolves over the next ten years.
Lajiao: Thank you. Mr. Li Jinwei, Refly has worked in AI for a long time. How do you view AI empowering companies internally?
Li Jinwei: Both phenomena do exist, but I agree with the previous speakers that this is an upward spiral. Large models have certainly crossed the productivity threshold by now. What stands in our way is whether your team's and company's Context can be brought together and connected so a large model fully understands the company and every individual. That makes a huge difference.
Why did OpenClaw become so popular this year? When we built workflows before, some customers still could not create highly productive workflows. OpenClaw naturally became an entry point for collecting enterprise context. With that data, workflows can genuinely function within a company. Simply put a bot into enterprise chats: as it gradually understands the knowledge, it can deliver real value.
Lajiao: Do you therefore think the workflow era is about to pass, with AI bringing the next era?
Li Jinwei: I would phrase it differently: the need for people to orchestrate workflows will pass, but workflows themselves will continue to exist. Whether first-generation workflow or today's large language models, the underlying dependency remains workflows. There is just less that people need to orchestrate. Our thinking for new products is also to make orchestration progressively simpler, so it exists invisibly within the enterprise environment.
Lajiao: Understood. Mr. Wang Pengfei, the industrial sectors you cover, such as automotive and aerospace, have more complex scenarios and must encounter bottlenecks connecting data. How does Mapping Intelligence address data silos?
Wang Pengfei: That is a very precise question. The hardest part of innovation in global manufacturing is what we discussed: enterprises must pay high integration costs and deeply understand each domain's processes. Our solution does not replace existing systems. It brings their data into our platform for local deployment.
Once systems' data is connected, we can try scenario-specific AI collaboration. Chinese and European R&D teams, for example, have different rhythms: European teams emphasize documenting before executing; Chinese teams emphasize iteration and immediate feedback. Our Agent with a human touch gradually learns to provide change documents for German teams and immediate feedback for Chinese teams, accelerating cross-cultural communication. AI is not just a tool; it gradually develops a “human quality.” That is something we find increasingly interesting in industry.
Lajiao: Mr. Wang's topic brings us to a second issue: does AI-enabled global collaboration simplify processes, or create new barriers? Some people think AI tools and data interoperability have themselves become obstacles. Have you experienced that in practice? Sean Li, please begin.
Sean Li: I think processes will always exist, but the question should be: “whose” process is this? Replacing human participation in a process with AI participation must be empowering, in my view. Ultimately, a process is path optimization with different Milestone points. Given sufficient Context, an AI agent can participate in and execute that optimization, saving people unnecessary time and effort.
Lajiao: So your central point is who makes decisions and who executes.
Sean Li: Personally, I feel people still make the decisions now, but I believe most decisions, including relatively high-level ones, will gradually be taken over by agents. Over the past ten years, we built a financial operating system. The next step is to use agents not merely as tools, but as process advisers and ultimate decision-makers. That is a major direction.
Lajiao: In actual cross-border payments, then, how do you use AI to address financial decision-making?
Sean Li: Our revenue-management product Airwallex Billing was built for that pain point. Once a company expanding overseas finds product-market fit—PMF—it must commercialize and make money. Then it has to consider geopolitics, cultural differences and different tax systems: for example, pricing by seat or usage, local purchasing power, and even tax compliance. AI magnifies these revenue- and compliance-management needs, so we provide a product specifically for such scenarios.
Lajiao: Mr. Liu of Yiyan Technology, how can AI tools standardize global-team collaboration and make it more process-driven, reducing confusion?
Liu Yuchen: Mr. Wang's example was very illuminating. We do not have one definitive best practice either. People use small tools of various kinds, such as one that translates a screenshot from a foreign language into Chinese to help with meetings held entirely through English PPT presentations. Large models also provide speech-to-text transcripts during meetings.
For collaboration tools, we moved from Feishu to Lark to address overseas compliance and permission separation. After building our knowledge base, our focus is reducing friction. AI is powerful, and some colleagues' output has become virtually unlimited: they finish a meeting and ask AI for a Summary and PPT without summarizing and organizing the material themselves. Others receive too much undigested information. Cases where good tools increase friction are happening too. We now require people not to rely on AI for everything and to personally check the critical points.
Lajiao: We have repeatedly mentioned executors and decision-makers. Mr. Wang Pengfei, compared with traditional integration solutions, what new changes arise when an AI Agent executes complex industrial processes?
Wang Pengfei: There is an underlying logic: earlier systems were based on fixed “processes and rules,” whereas in the AI era they are based on “change,” with more emphasis on understanding underlying logic. So far, we have concentrated on building experience in ALM—R&D management. Our strategy is to build step by step.
Traditional manufacturers have a heart, liver, spleen, stomach and kidneys—their different systems—but these do not communicate. First, we “build blood vessels,” using our Connector platform to connect every organ. Once circulation begins, the second step is to accumulate Skills and grow a “nervous system.” Finally, the nervous and circulatory systems work together to form an “intelligent brain.” That is our development rhythm. If anyone here works in this area, please contact me so we can explore it together.
Lajiao: On software, Mr. Li: Nexu launched this year. How does it address the slow configuration of traditional workflow software? Can nontechnical teams use it?
Li Jinwei: Briefly, three points. First, greatly reduce deployment costs: we simplified it to a one-click desktop-client installation that may take 5 minutes. Second, simplify configuration and use. We built a middle-layer configuration compiler to support fallback strategies and address compatibility across chat software, so scenarios can run quickly. Third, we do not want a new AI tool that creates data silos. We want a digital employee that genuinely “lives in existing tools,” understands context and helps execute production-grade tasks.
Lajiao: Our final topic: how can companies find more innovative paths for collaboration in a global context? Are there new opportunities ahead? Please share, starting with Sean Li of Airwallex.
Sean Li: Stepping outside Airwallex's perspective, I strongly agree with Mr. Wang's direction: cultivate agents in very specific verticals so they become real professionals. Large language models tend to produce a Generalist, but future opportunities lie in more specialized Agents, such as in legal services, medicine, payments and finance. That requires accumulating a great deal of nontextual experience.
For our own product, we use AI effectively to handle tedious matters for merchants, such as differences in national regulatory environments and what the source calls tax confirmation (Revenue Recognition), so they do not have to worry about tax or financial reporting, as I put it.
Lajiao: Mr. Liu, how can AI collaboration tools meet the needs of companies of different sizes and support differentiated implementation?
Liu Yuchen: We have grown from a few people to several dozen, and collaboration on Lark works well for us. I see many new opportunities. For instance, people may think AI recorders such as Plaud are an old category, but nearly every one of our presales staff has one. Actually operating a global company lets you discover many unmet market needs, which may become the starting points for new, great companies.
Lajiao: Thank you. Mr. Wang, over the next 5 to 10 years, will AI simply be an enhancement plugin for industrial software, or reshape an entirely new native collaboration platform? What role will Mapping Intelligence play?
Wang Pengfei: AI develops less quickly in industry than in consumer markets. Fully AI-native industrial software will not emerge over the next three to five years. So our strategy runs on two tracks: in the next three to five years, we help customers quickly connect data within 30 minutes through “plug-and-play industrial software.” But over the next 10 years, an industrial-grade native operating system will certainly emerge, and we will build our own industrial-native operating system then too.
Lajiao: Thank you. Mr. Li, some believe AI can eliminate language and cultural barriers; others think shared understanding and trust cannot be replaced by AI. Which does your product believe more?
Li Jinwei: I think both exist, but I believe more strongly that human trust still has to be resolved by people. The same proposal carries very different trust depending on whether AI creates it or someone who understands the boss does. We want AI to take over transactional work and free people. Those freed-up people should spend more time building trust, communicating and aligning context. That is how global collaboration becomes more efficient.
Lajiao: Finally, in one sentence each: what trend are you most optimistic about in AI-enabled global enterprise collaboration in 2026? Mr. Li first, please.
Li Jinwei: Two points. First, a shift “from building tools to building digital employees.” Second, “localization and open source” will become clear trends because they concern protection of high-value private enterprise data. What is truly valuable is the Context you accumulate.
Wang Pengfei: If you are still looking for startup opportunities, here is one direction we recommend. Chinese companies expanding overseas hit a bottleneck when they do not know how to develop those markets. Consider replacing CRM with AI and building an AI CRM solution. There is currently no standardized product in this area, in my view, but one will certainly emerge. Airwallex could also be an integrated provider within such a solution.
Liu Yuchen: We think software expanding overseas in '26 will see a trend called “software localization.” This goes beyond translation to include local support channels, local development of some features to ensure compliance, and local data storage. It is our pain point. We have even considered productizing how we solve it to provide local implementation support for others.
Sean Li: At YC, there was always the principle “Make something people want.” I think product builders now need to extend that to “Make something agents want.” Agents will gradually act in place of people's habitual operations, so the value you deliver must be adapted to connect with agents. Our current product-experience philosophy is to prioritize Agents' ability to use and understand the product when considering innovation.
Lajiao: Thank you to all four guests for the excellent insights and new perspectives. Could everyone here give our four speakers a round of applause? Thank you!
This article was compiled from the live event transcript.