AI Transformation of So-Young Clinics: First, Turn the Physical World into Data

At the recently concluded 2026 World Artificial Intelligence Conference, enterprise-level AI emerged as one of the most watched sectors.

Compared to previous years, which focused on showcasing model parameters and generative capabilities, this year’s exhibition floors featured more agents, industry-specific solutions, and enterprise deployment tools. AI companies’ products are extending further to connect corporate data, systems, and business processes.

However, when AI truly enters the physical world, things do not go as smoothly as they do in product demonstrations.

Take clinics as an example: which consultation room a customer enters, what procedures a doctor performs, where equipment is moved, and why a medication is dispensed and then returned—these events do not automatically become data.

Recently, Jin Xing, founder of So-Young, engaged in a nearly two-hour discussion with media outlets including Wall Street News · All-Weather Tech.

So-Young is advancing AI transformation across its more than 50 light medical aesthetics clinics, aiming to apply AI to consultations, treatment quality control, patient triage, and clinic operations.

What left a deeper impression during this exchange was the extensive foundational work So-Young undertook to enable AI integration into its clinics.

For instance, to track the waiting time of consumers at different stages within the clinic, So-Young experimented with tablet check-ins, wristbands, Bluetooth, and Wi-Fi. To allow headquarters to remotely monitor whether doctors adhere to standard treatment protocols, the company installed cameras in consultation rooms, only to discover that network bandwidth at clinics nationwide needed to be upgraded individually.

So-Young began building this digital infrastructure in 2023, investing over a hundred engineers. Three years later, Jin Xing believes the company is still primarily in the “first half” of AI implementation—data governance.

“Only with this accumulation of data can AI learn specifically for your business,” Jin said. “We are still largely in the phase of digitization.”

So-Young’s practice also reveals an invisible threshold for AI adoption in chain businesses: before AI can be implemented, enterprises must first transform the ever-changing physical world into authentic, continuous, and machine-understandable data.

The Hidden Cost of Data Collection

For chain enterprises, one of the most appealing values of AI is the ability to replicate the capabilities of top-performing clinics and staff.

A single clinic can rely on its manager, doctors, or skilled employees to operate. However, as the number of clinics grows from ten to dozens or even hundreds, the enterprise must determine whether each location operates according to the same standards.

So-Young aims to become a standardized chain of light medical aesthetics clinics.

In Jin Xing’s vision, if the same consumer with identical concerns visits different clinics and undergoes consultations with different doctors and consultants, the resulting treatment plans should be highly consistent.

However, So-Young has not yet achieved this.

The final treatment plan a consumer receives still largely depends on the individual knowledge and experience of the doctor and consultant.

To narrow this gap, AI first needs to know exactly what is happening inside the clinics.

A complete medical aesthetics treatment involves at least the following stages: the consumer stating their concerns, skin analysis, the doctor diagnosing the issue, designing a treatment plan, preparing medications and consumables, performing the treatment, and post-treatment re-evaluation and feedback.

However, this information is presented in various formats and scattered across different locations.

Consumer concerns are captured in consultation dialogues; skin conditions come from testing instruments; medications and dosages used are recorded in supply chain systems; the treatment process is captured on video; and final outcomes are assessed through before-and-after photos, re-evaluation reports, and reviews.

So-Young employs different collection methods for different types of data.

Consultation dialogues are recorded via speech-to-text; skin analyzers are integrated with backend systems to break down previously view-only reports into structured metrics; medications and consumables are tracked through ERP systems; and doctors’ treatment processes are recorded on video.

According to Jin Xing, So-Young has accumulated over 3 million before-and-after treatment photos and nearly 530,000 re-evaluation reports based on approximately 1.75 million treatments.

However, data governance is not just about storing as much information as possible; the first challenge is ensuring data authenticity.

In practice, So-Young initially aimed to measure how much time consumers spent at various stages, such as consultations, skin analysis, pre-treatment preparation, and the treatment itself.

So-Young initially placed tablets at different stations for employees to manually check in. However, during busy periods, employees often missed check-ins. Once headquarters began assessing check-in rates, some staff members started backdating or retroactively entering data.

The records in the system became complete, but the time data was no longer accurate.

“This data is considered ‘dirty data’ for us, and dirty data is useless,” Jin said.

So-Young subsequently tried using wristbands with identification codes, requiring consumers to scan them upon entering different areas. However, wristbands interfered with the consumer experience. Why would customers need to wear wristbands? Would different colors be interpreted as distinguishing membership tiers?

To make the wristbands seem valuable, the team even discussed adding features such as unlocking lockers.

Bluetooth and Wi-Fi solutions also failed to solve the problem independently.

Manual collection increases workload, while seamless collection requires trade-offs between accuracy, cost, and customer experience.

Ultimately, So-Young had to combine multiple methods to reconstruct the real-time locations of consumers, employees, and equipment within the clinics as accurately as possible.

This exposes a contradiction in data governance for offline chain stores: poorly designed collection mechanisms can increase employee burden, alter employee behavior, and create a dataset that appears complete on the surface but is actually distorted.

AI Enters Treatment Inspection

Although So-Young is still undergoing data governance, expected to be completed by the end of this year, AI is gradually being introduced into its clinics.

Video inspection of treatment processes has become one of the scenarios for AI implementation.

For So-Young, the core issue video inspection aims to address is monitoring whether different clinics are genuinely adhering to unified treatment standards.

To this end, So-Young has established standard SOPs for different treatment items. For example, for a BBL procedure involving 10 steps, the SOP specifies the required actions and standard duration for each step.

Meanwhile, dual screens have been installed in consultation rooms. One screen displays the operational workflow for the current treatment, allowing doctors to follow along during the procedure and enabling consumers to view the treatment progress.

To ensure each step is executed accurately, So-Young initially supervised the treatment process through manual video inspections. Headquarters set up a dedicated inspection team, with inspectors remotely viewing treatment conditions across different clinics and consultation rooms via video to monitor whether doctors followed the SOPs.

However, as the number of clinics and treatments increased, the volume of daily video footage grew, making manual inspection inefficient.

So-Young is attempting to change this process with AI. The company informed All-Weather Tech that a new intelligent inspection system is scheduled to launch by the end of July.

The new system will introduce AI visual recognition capabilities to perform frame extraction, object detection, and behavior analysis on treatment videos. It will identify who appears in the video, which step the doctor is performing, what equipment is being used, and whether the actions comply with standards or show signs of anomalies.

According to So-Young’s plan, this system will gradually achieve 24/7 intelligent inspection, improving model accuracy through continuous human verification and correction of recognition results.

However, the primary challenge So-Young faces in deploying this system across all nationwide clinics is not model recognition, but network bandwidth.

Medical aesthetics clinics originally did not need to continuously upload large volumes of video, as existing networks had to support customer service and other business systems simultaneously. With the launch of intelligent inspection, consultation room footage needs to be continuously transmitted back to headquarters, quickly making the existing bandwidth a bottleneck.

After actual implementation, So-Young found that the bandwidth for each clinic needed to be expanded to at least 200 Mbps; otherwise, video uploads could disrupt the operation of other systems. The company had to contact network providers in different regions and shopping malls to upgrade the network for each clinic individually.

Transitioning from manual spot checks to AI inspection means that a feature seemingly belonging to AI visual recognition ultimately involves cameras, network bandwidth, video storage, SOP definitions, and manual labeling.

This also implies that bringing AI into clinics is not an upgrade that yields immediate profits. To enable models to see and understand real treatment processes, enterprises must first invest substantial engineering resources to transform existing systems, hardware, and business processes.

Jin Xing does not believe that AI will necessarily make individual clinics more profitable in the short term.

However, for chain enterprises, the true return on AI may not be reflected solely in the income statement of a single clinic, but rather in expanding the management radius of the entire organization. It gradually transforms management capabilities, which previously relied on individual experience, into systemic capabilities that can be invoked, monitored, and replicated by headquarters.

But before this can happen, enterprises must first see the real world. Transforming an offline clinic into a world that machines can understand still requires a great deal of slow, concrete work.