Original · Unique Research · 2025-10-29 · Shanghai
When it comes to AI, I increasingly believe one thing: not everyone should build models or create technology, but no one can avoid one question—how exactly do you plan to use AI? A while ago, at the China-ASEAN Expo in Nanning, I spoke with Gao Jiahui, founder of Luoji Technology. That conversation gave me a much more concrete sense of what it means to put AI to work in industry-specific scenarios.
To be honest, conversations about AI often begin with high-level concepts such as models, technology, and parameters. Yet AI businesses that actually work usually begin somewhere very practical. Gao did not come to Nanning to watch the launch of a large model or to explain how disruptive AI could be. He brought an intelligent customer service system that had already proven itself in China’s hotel industry, hoping to see whether Nanning—this gateway city—could help take it into Southeast Asia.
That starting point may not sound especially thrilling. But companies that cultivate deep industry expertise are precisely the ones most likely to turn AI into a real business.
Gao mentioned a large hotel group they serve that employs roughly several hundred customer service agents. This is a typical to B scenario in an industry where costs are plainly visible. The greatest value of AI-powered customer service is not making people notice that you are “using AI.” It is directly reducing labor requirements, improving efficiency, and avoiding errors.
But that raises a question. AI customer service sounds like something anyone can build, and creating a chatbot is not difficult. So why should a customer choose you?
Gao’s answer was refreshingly practical: being good at AI is not enough. You must first be an industry expert and only then an AI expert. The idea sounded familiar; it seems to be repeated by everyone working in AI+ vertical industries. But Luoji Technology’s approach is genuinely different. The company progressed from smart hardware and intelligent systems to deep learning, with every step tied directly to the hotel industry. That is why customers trust the company, are willing to provide data, and train models together with it.
Isn’t this a classic case of exchanging trust for real-world deployment?
Back to Southeast Asia. Luoji Technology has not yet truly established operations there, but at this China-ASEAN Expo it demonstrated versions of its product in more than a dozen languages. Why?
Gao offered a very direct example. A Chinese traveler staying at a hotel in Malaysia may want to order water, request cleaning, or check hotel services, but cannot communicate in the local language. If the traveler can speak Chinese through a mini program or smart terminal, the system can automatically convert the request into the local language and dispatch it to a hotel employee, who receives the task directly. Wouldn’t that become a “cross-language intelligent work-order system”?
At first glance, this may sound like “voice assistant plus translation.” But Gao stressed a crucial point: it is not enough to provide an answer. The answer must be correct, and the subsequent action must be executed accurately. AI customer service is not a chatbot that keeps you company; it is a business entry point that must solve a problem. Errors may be tolerable in some C-end scenarios, but in B-end settings—especially guest-facing hotel operations—even one mistake can become an incident.
Luoji Technology set two core metrics for its system:
The first is response speed, with a target of responding within 5 seconds. To achieve it, the company uses a routing mechanism built around a small model + large model + agent. Simple questions go to the small model for rapid resolution, while complex questions proceed to the large model for deeper reasoning. It may not look especially cool, but it is an extremely practical architecture.
The second is accuracy, which has now reached more than 95%. That figure was not invented. Hotel customers supplied tens of thousands of real use cases for testing, and the system was repeatedly iterated and optimized. This is an AI system that can genuinely go live and run business operations, rather than a capability that exists only in PPT form.
When many people discuss AI, they say things such as “how many parameters our model has” or “how advanced our embedding is.” Customers simply do not care. They look at two things: can it be used, and what happens when something goes wrong?
Gao emphasized that their goal in building this AI customer service system was not to replace every customer service agent. It was to assign highly repetitive, standardized work to AI, allowing human agents to devote more energy to serving guests and improving the experience.
I strongly agree. Discussions about AI often fall into a binary either-or mindset, as though people become useless once AI arrives. Reality is usually different. People plus AI can free humans to do the work they should have been doing but never had time to do.
If hotel customer service agents spend every day answering “What time is checkout?” and “How do I connect to Wi-Fi?”, when will they have time to discuss sightseeing recommendations, weather reminders, and other small details that can improve a guest’s experience?
After hearing Gao’s account, I think anyone who wants to take AI global—whether by building an application or a platform—should first think clearly about three things:
Do you genuinely understand the industry? Do not simply say, “AI can empower the xx industry.” Explain which process, which step, what method, and what value it can deliver. If you do not even understand the customer’s business, how can you solve the customer’s problems?
Have you genuinely deployed a product? A demo is not enough. You need evidence from real customers, real metrics, and real repeat-purchase data. That gives you leverage in customer negotiations and provides the foundation for trust with partners.
Do you have the capability for “localized operations + system delivery”? Going global is not accomplished simply by translating a system into English. You must understand which App Southeast Asian users rely on, what experiences they prefer, which brands they trust, and how you will work with local companies or governments. That is what it means to “go global,” not merely to “change languages.”
At the end of the interview, Gao made a particularly powerful statement: AI is the greatest opportunity that those of us born in the 80s will encounter.
I have heard people ask more than once whether AI is a bubble and whether only large companies can build large models. But Gao’s actions offer another answer: not everyone needs to build a large model, but everyone can use AI to do something—especially within an industry they already know well.
This may be the AI era that truly belongs to entrepreneurs. The question is not “whether to do AI,” but “how to use AI and turn it into a business that can be implemented, sustained, and replicated.”
What about you? Are you genuinely using AI in the work you are doing now? In your industry, where do you think AI should be deployed first?
That might just be the starting point for the next Luoji Technology.
Q1: What concrete plans or progress does Luoji Technology currently have in the Southeast Asian market?
Gao Jiahui: We have not yet truly established operations in the Southeast Asian market. We came to the China-ASEAN Expo for two main reasons. First, we hope to use this channel to open the market for AI products that have already matured and been successfully deployed in China. Second, we want to gain a deeper understanding of Southeast Asia’s specific demand for AI. At the expo, we have already spoken with representatives from Vietnam, Thailand, Malaysia, and other countries.
Q2: What preparations have you made for the Southeast Asian market, and what problems does your product mainly address?
Gao Jiahui: The first problem we need to solve is language. The product demonstrated at our booth already supports more than a dozen languages. For example, a Chinese guest staying at a hotel in Malaysia may speak neither Malay nor English. Through our mini program or WhatsApp, the guest can make a request in Chinese—asking for a bottle of water, for instance. Our system automatically dispatches the task to a hotel employee through the internal work-order system, and the employee receives it in the local language.
Q3: How is your hotel AI customer service system different from other chatbots on the market?
Gao Jiahui: It is more than simple language conversion. We believe AI customer service must not only be able to “produce an answer” and “be usable.” More importantly, it must “give the correct answer,” and the subsequent “execution must be accurate.” This is a guest-facing C-end function, but it requires powerful B-end execution behind the scenes. It is an enterprise product, and we believe this is where the experience we have accumulated in the Chinese market gives us an advantage.
Q4: You mentioned “giving the correct answer.” Enterprise customers (To B) demand much higher accuracy than ordinary users (To C). How do you address accuracy?
Gao Jiahui: Yes. If an answer is wrong in a To C setting, a user may simply think the system is “stupid.” But in a To B setting, a single incorrect answer can become an incident. We currently focus on two metrics:
Response speed: We can currently respond within 5 seconds.
Accuracy: Our current accuracy in online gray testing is approximately 95% or higher.
Q5: How do you achieve a 5-second response time? Do some questions not require deep reasoning by a large model?
Gao Jiahui: Correct. Before the large model, we added a small model + Agent as a routing layer. If a guest simply asks for a bottle of water, the small model identifies the simple request directly and turns it into a work-order flow, which is very fast. Only complex questions enter the large model for deeper reasoning. We started at 10 seconds, then 8 seconds, and kept optimizing at the millisecond level until we found the present balance of under 5 seconds.
Q6: How did you calculate accuracy of more than 95%, and what do you do about the remaining 5% that is inaccurate?
Gao Jiahui: Our strategic partners, including Tencent Cloud and Chinese hotel groups, gave us tens of thousands of test cases drawn from real past events. We continually train and test the model with this data.
For the remaining 5%, we have two very important mechanisms:
Human fallback: When the AI recognizes a problem that it may not be able to solve, the system immediately transfers the case to a human customer service agent.
Sensitive-content handling: Especially when serving C-end guests, we have mechanisms to deal immediately with unhealthy or inappropriate content. At this stage, therefore, the solution must combine AI + people.
Q7: What is the core value of your AI solution for hotel groups? How do they measure ROI (return on investment)?
Gao Jiahui: We cannot “use AI merely for the sake of AI.” In actual deployments, the first requirement from our hotel-group customers is to reduce customer service costs. Their groups employ roughly several hundred customer service agents, and our goal is to help them cut that number in half.
Q8: Where do the half of customer service employees replaced by AI go?
Gao Jiahui: AI primarily handles repetitive work. The labor capacity freed in this way can be redirected toward more guest-facing, human-centered service. I think this is an important role for AI.
Q9: Do customers—hotels—need to possess AI capabilities themselves? How do you help them develop “AI Native” employees?
Gao Jiahui: We are a technology company, and we provide the technology platform and Agent technology. But subsequent maintenance and improvement—especially adding “corpus material” for specific scenarios—still need to be handled by customers themselves. Large hotel groups generally have their own IT departments or a CIO, and this responsibility falls within their role. We work with them, of course, although this places somewhat higher demands on their employees than on traditional general staff.
Q10: What is the greatest challenge when promoting AI to B-end customers?
Gao Jiahui: The greatest challenge is the mismatch between “expectations for application scenarios” and “the current level of technology.” Take the 5-second response time we just discussed. Customers want 2 seconds, but existing technologies of every kind—whether small or large models—cannot currently deliver that. Or, if they can achieve 2 seconds, they must sacrifice other performance and cannot reach an overall balance.
Another core point is that you must first be an industry expert and only then an AI expert. You cannot wait for customers to say, “I want AI.” You must be able to tell them precisely: “Given your situation, which problem should you solve first, and in which scenario should you deploy AI?”