【China】So-Young Founder Jin Xing: To Introduce AI into Physical Stores, We Must First Overcome the ‘Data Governance’ Barrier

At the recently concluded 2026 World Artificial Intelligence Conference, enterprise AI emerged as a major focus. Unlike previous years, where the emphasis was on flaunting model parameters and generative capabilities, this year’s exhibition booths featured more AI agents, industry solutions, and enterprise deployment tools. AI companies are now reaching into the deep domain of connecting enterprise data, systems, and business processes. However, when AI enters the physical world—especially in settings heavily reliant on offline, face-to-face interactions like chain clinics—things do not proceed as smoothly as product demonstrations suggest.

For chain enterprises, one of the most compelling values of AI is its ability to replicate the capabilities of top-performing stores and employees. While a single store can rely on its manager, doctors, and experienced staff, as the number of stores grows from ten to dozens or even hundreds, the company must ensure all stores operate to the same standard. So-Young aims to build a standardized chain of light medical aesthetic clinics. In Jin Xing’s vision, the same consumer with the same needs visiting different stores and consulting with different doctors or counselors should receive very similar treatment plans. However, in reality, the final plan a consumer receives still heavily depends on the individual knowledge and experience of the doctor or counselor.

For AI to reduce this disparity, it must first understand what is actually happening in the stores. A complete medical aesthetic procedure involves multiple stages: the consumer communicating their needs, a skin examination, the doctor diagnosing the problem, designing a treatment plan, preparing medications and consumables, performing the procedure, and post-procedure follow-up and feedback. However, this information is scattered across different media. Needs are in consultation conversations, skin condition comes from examination equipment, medications and dosages are recorded in the supply chain system, the procedure is on video, and the final outcome is judged by before-and-after photos, follow-up reports, and reviews.

So-Young adopted different collection methods for different data types. Consultation conversations are transcribed via voice recognition. Skin examination equipment interfaces are integrated with the backend to decompose previously view-only reports into structured indicators. Medications and consumables are tracked through the ERP system, and the doctor’s procedure is recorded on video. According to Jin Xing, So-Young has accumulated over 3 million before-and-after comparison photos and nearly 530,000 follow-up reports from approximately 1.75 million procedures.

However, the first major challenge in data governance is not information storage, but ensuring data authenticity. So-Young once tried to statistically measure how much time consumers spent in each stage—consultation, skin exam, pre-procedure preparation, and the procedure itself. Initially, they placed tablets at each stage for employees to manually check in. However, when stores were busy, employees would forget to check in. When headquarters made check-in rates a performance metric, some employees began checking in all at once, either before or after the fact. While the system records appeared complete, the time accuracy was lost.

Later, the company tried wristbands with identification codes, which consumers would scan upon arriving at different areas, but the wristbands compromised the consumer experience. Why wear a wristband? Would different colors be perceived as distinguishing membership tiers? To give the wristbands meaning, the team even discussed adding a function to open lockers. Methods like Bluetooth and Wi-Fi also failed to solve the problem alone. Manual collection increased workload, while non-sensing collection forced trade-offs between accuracy, cost, and customer experience. Ultimately, So-Young chose to combine multiple methods to reconstruct, as accurately as possible, the actual positions of consumers, employees, and equipment within the store.

This highlights a core contradiction in data governance for offline chain stores: if the collection mechanism is poorly designed, it can increase employee burden, alter employee behavior, and produce data that appears complete but is actually distorted. So-Young began building this digital system in 2023, deploying over 100 engineers. Three years later, Jin Xing still believes the company is primarily in the “first half” of AI deployment—the data governance phase.

While data governance is not expected to be complete until at least the end of this year, AI is already being gradually introduced into stores. Video auditing of procedure processes is a key application. For So-Young, the core problem video auditing aims to solve is monitoring whether each store is actually executing unified procedure standards. To this end, So-Young has developed standard SOPs for each procedure type. For example, a BBL (Broadband Light) treatment with 10 steps has an SOP that clearly defines the operations to be completed at each step and the standard time required.

At the same time, consultation rooms are equipped with dual displays. One screen shows the operation steps for the current procedure, which the doctor can reference during the procedure, and the consumer can also see the progress. To ensure each step is executed correctly, So-Young has been supervising the procedure process through manual video auditing. A dedicated auditing team at headquarters remotely reviews procedure videos from each store and consultation room. However, as the number of stores and procedures grows, the daily volume of generated videos increases, and the efficiency of manual auditing declines.

So-Young is now attempting to transform this process with AI. According to information provided by So-Young to All-Weather Tech, the company plans to launch a new smart auditing system by the end of July. The new system will incorporate AI image recognition capabilities for frame extraction, object detection, and action analysis of procedure videos. It will identify who is in the video, which step the doctor is performing, what equipment is being used, whether the actions conform to standards, and whether any anomalies are suspected. The plan is for the system to gradually achieve 24/7 smart auditing, with model accuracy improving through continuous human verification and correction of identification results.

However, when deploying this system across stores nationwide, the first challenge So-Young faced was not model recognition accuracy, but network bandwidth. Medical aesthetic clinics typically do not need to continuously upload large volumes of video; their existing networks are already strained supporting customer service and other business systems. After introducing smart auditing, consultation room footage must be continuously transmitted to headquarters, quickly making the existing bandwidth a bottleneck. Through practical operation, So-Young determined that each store’s bandwidth must be expanded to at least 200 Mbps. Otherwise, video uploads could affect the operation of other systems. The company had to contact network providers in different regions and commercial properties to upgrade the network for each store.

The transition from manual spot checks to AI auditing—which seems like a simple function of AI image recognition—ultimately involves cameras, network bandwidth, video storage, SOP definitions, and manual annotation, all working together. This also means that deploying AI in stores is not a quick path to profitability. To enable the model to see and understand actual procedure processes, the company must first invest significant engineering resources to modify existing systems, hardware, and business processes.

Jin Xing does not believe AI will quickly improve the profitability of a single store in the short term. However, for chain enterprises, the true return AI can bring is not just what appears on a single store’s profit and loss statement. It lies in expanding the organization’s overall manageable scope and gradually transforming management capabilities that previously relied on individual experience into system capabilities that headquarters can invoke, monitor, and replicate.

So-Young’s practice provides a highly representative sample of AI deployment across the chain industry. It highlights a reality that is often overlooked: in an era of rapidly evolving AI large model capabilities, the hardest part of deploying AI into the real physical world is often not the algorithm itself, but how to convert the ever-changing real world into truthful, continuous, machine-understandable data. From tablet check-ins to wristbands and then to multi-technology consumer flow tracking; from manual auditing to smart video analysis requiring bandwidth upgrades for each store—each seemingly minor engineering detail holds the key to whether AI can truly take root in offline settings.

But before that, enterprises must first make the real world “visible.” Transforming a single offline store into a world that machines can understand still requires a lot of slow, concrete work. As Jin Xing says, only after this data accumulation is complete can AI truly learn for specific business tasks. For chain enterprises considering AI deployment, So-Young’s story may offer a sobering reference point: the release of AI’s value begins with patient and honest recording of the real world.