AI transformation of Xinyang stores: First, turn the real world into data

Just concluded, the 2026 World Artificial Intelligence Conference saw enterprise AI become one of the most closely watched tracks. Compared to previous years, which focused on displaying model parameters and generation capabilities, this year’s booths featured many more agents, industry solutions, and enterprise deployment tools. AI companies’ products are increasingly extending to connect enterprise data, systems, and business processes. But when AI truly enters the physical world, things won’t go as smoothly as product demonstrations suggest. Take clinics as an example: which consultation room a customer enters, what procedures the doctor completes, where the equipment is moved, why a particular medicine is picked up and then returned—none of these things will automatically turn into data. Recently, Xing Jin, founder of So-Young, held a nearly two-hour exchange with Wall Street Insights·全天候科技 and other media. So-Young is promoting AI transformation in over 50 of its light medical beauty stores, hoping AI can be used for facial consultations, treatment quality inspection, customer triage, and store operations. What was more impressive about this exchange were the foundational tasks So-Young undertook to enable AI to enter the stores. For example, to track how long a consumer waited at different stages inside the store, So-Young tried tablet sign-in, wristbands, Bluetooth, Wi-Fi, and other methods; to enable headquarters to remotely monitor whether doctors were following standard treatment procedures, cameras were installed in consultation rooms, eventually revealing the need to upgrade network bandwidth at each location nationwide. So-Young began building this digital system in 2023, investing more than a hundred engineers. Three years on, Jin Xing believes the company is still mainly in the “first half”—data governance—of AI implementation. “With this accumulation of data, AI can learn according to your business,” Jin Xing said. “We are still in a stage of extensive digitization.” So-Young’s experience also highlights a hidden threshold for AI implementation in chain businesses: before AI can be put to use, companies must first turn an ever-changing real world into authentic, continuous, and machine-understandable data. Data Collection’s Hidden Costs For chain enterprises, one of AI’s most enticing values is its ability to replicate the performance of outstanding stores and employees. A single store can rely on its manager, doctors, or skilled staff to maintain operations. But when the number of stores grows from ten to dozens or even hundreds, the enterprise must know whether each operates according to the same standards. So-Young aims to become a standardized light medical beauty chain. According to Jin Xing, the idea is that a consumer with the same needs should get highly similar treatment plans from different stores and different doctors or consultants. Yet currently, So-Young still cannot achieve this. What consumers ultimately receive is still largely dependent on the personal knowledge and experience of doctors and consultants. To minimize this variability with AI, it first needs to know exactly what happens inside the stores. A complete medical beauty treatment includes at least the consumer voicing their needs, skin analysis, doctor’s diagnosis, treatment plan design, preparation of medicines and consumables, treatment completion, and post-treatment reassessment and feedback. But this information is presented in various forms, scattered across different areas. Consumers’ needs are a conversation during a consultation; skin condition comes from diagnostic equipment; what medicines and dosages are used is recorded in the supply chain system; the treatment process is a video; the outcome must be judged by before-and-after photos, reassessment reports, and reviews. For different types of data, So-Young uses different collection methods. Consultation conversations are transcribed via speech-to-text; skin diagnostic machines connect to the backend, turning what was once just viewable reports into structured data; medicines and consumables are tracked by the ERP system; doctors’ treatment procedures are recorded by video. According to Jin Xing, So-Young has compiled over 1.75 million treatments, collecting more than three million before-and-after comparison photos, and nearly 530,000 reassessment reports. But data governance is not about storing as much information as possible; the first challenge is ensuring the data is authentic. In practice, So-Young once wanted to track how much time consumers spent in each stage—consultation, skin analysis, pre-treatment prep, and treatment. Initially, tablets were placed at each stage for manual sign-in, but when stores got busy, employees easily forgot to sign in. If headquarters tracked sign-in rates, people would pre-fill or belatedly sign in. Records in the system became complete, but the time data was no longer authentic. “This data is dirty data for us, and dirty data is useless,” Jin Xing said. So-Young then tried wristbands with recognition codes, allowing consumers to scan at each area. But wristbands interfered with customer experience. Why should customers wear wristbands? Would different colors be interpreted as membership tiers? To make wristbands seem valuable, the team even considered adding functions like opening lockers. Bluetooth and Wi-Fi didn’t solve the problem either. Manual collection increases workload, while contactless collection requires trade-offs between accuracy, cost, and customer experience. Ultimately, So-Young had to combine multiple methods to best restore the real location of consumers, staff, and devices inside the store. This exposes a set of contradictions in offline chain store data governance: poor collection mechanisms may increase employee workload, even modify behavior, and produce superficially complete but actually distorted data. AI Enters Treatment Inspection Although So-Young is still working on data governance, expected to be completed by year-end, AI is already gradually entering the stores. Video inspection of treatment processes is becoming a scenario for AI implementation. For So-Young, the core issue video inspection aims to solve is monitoring whether different stores truly execute unified treatment standards. To this end, So-Young developed standard SOPs for each treatment. For a BBL project with ten steps, the SOP specifies the procedure and duration required for each step. Meanwhile, dual screens are installed in consultation rooms; one displays the procedure flow for the current treatment, allowing doctors to follow along in real time, and letting consumers see the progress. To ensure every step is accurately performed, So-Young uses manual video inspection for supervision. Headquarters has a dedicated inspection team that remotely reviews treatment videos from different stores and rooms to check if doctors operate according to SOPs. But as the number of stores and treatments increases, the volume of videos produced daily grows, and manual inspection becomes inefficient. So-Young is trying to use AI to transform this process. According to So-Young’s comments to 全天候科技, a new intelligent inspection system is planned to go online at the end of July. This new system will introduce AI visual recognition, doing frame extraction, object detection, and behavioral analysis on treatment videos, identifying who appears, what step the doctor is executing, what equipment is used, whether actions meet standards, and if any suspected abnormalities arise. According to So-Young’s plan, the system will gradually achieve 24/7 intelligent inspection, with humans continuously confirming and correcting detection results to improve model accuracy. But to roll out this system nationwide, the primary challenge isn’t model recognition, but network bandwidth. Medical beauty shops didn't originally need continuous video uploads; the existing network already handles customer service and business systems. Once intelligent inspection is launched, consultation room footage must continuously be sent to headquarters; the original bandwidth quickly becomes a bottleneck. So-Young found that each store’s bandwidth needed to expand to at least 200 Mbps, or else video uploading would affect other system operations. The company had to contact different regional and mall network providers to upgrade store networks one by one. From manual spot checks to AI inspection, what appears to be an AI visual recognition feature ultimately involves cameras, network bandwidth, video storage, SOP definition, manual annotation, and more. This also means that introducing AI into stores isn’t a quick profit-yielding upgrade. To let the model “see” and understand the real treatment process, companies first have to invest significant engineering resources to transform existing systems, hardware, and business processes. Jin Xing does not believe AI will necessarily make any single store more profitable in the short term. But for chain enterprises, the true gains from AI lie not just on one store's profit statement, but in expanding the entire organization's management capacity—transforming management once dependent on personal experience into systematic abilities that headquarters can monitor, invoke, and replicate. But before all this, the company must first see the real world. Turning an offline store into a world that machines can understand still requires a lot of slow, concrete work. Risk Warning and Disclaimer The market is risky, and investment needs caution. This article does not constitute individual investment advice, nor does it take into account the unique investment objectives, financial circumstances, or needs of any individual user. Users should consider whether any views, opinions, or conclusions in this article are suitable for their specific situation. Investment based on this article is at their own risk.