1. Current Pain Points
When selecting skincare or health products, many individuals often find themselves in one of two extremes: the first group makes impulsive purchases based solely on advertisements claiming effectiveness, resulting in a collection of unused bottles at home; the second group invests considerable time researching ingredients, comparing prices, and reading reviews, only to become paralyzed by choice overload, leading to a lack of consistent skincare habits even after three months.
From a business perspective, brands face similar challenges. In traditional retail models, Customer Lifetime Value (LTV) is extremely low due to highly fragmented and unpredictable consumer purchasing behavior. Brands invest heavily in advertising to acquire new customers but struggle to establish effective repurchase mechanisms, causing Customer Acquisition Costs (CAC) to rise continuously. A more critical issue is that existing recommendation systems primarily operate at the “collaborative filtering” level, failing to dynamically adjust based on individual physiological cycles, seasonal changes, or even stress levels.
The missing element is not the product itself, but rather a continuously operational personalized formula engine and automated fulfillment system. When the architectural design remains focused on one-time transactions, it will perpetually burn cash in the red ocean of traffic acquisition.
2. Underlying Logic Breakdown
To transform skincare and health formulas into scalable services, it is essential to understand three foundational modules: State Collection, Formula Computation, and Fulfillment Automation.
The first layer is state collection. Traditional questionnaires can only capture static labels (age, gender, skin type), while human physiological and psychological states are dynamic. A lightweight journal-style input interface needs to be designed, allowing users to report their current sleep quality, stress index, and skin condition at minimal cost (e.g., 30 seconds of simple answers or sliding scales daily). This time-series data is the true valuable training sample.
The second layer is formula computation. This is not a simple if-else rules engine; rather, it requires the establishment of a multi-objective optimization model. The input includes the user’s real-time state, historical feedback, inventory costs, and seasonal factors; the output is a prioritized list of products with dosage recommendations. The model can initially be built using decision trees or gradient boosting methods, and once sufficient data is accumulated, deep learning can be introduced for fine-tuning. The key is to ensure that this computational logic is explainable; users need to understand why vitamin D is added this month or why the recommendation is to pause acid-based skincare this week.
The third layer is fulfillment automation. Once the formula is confirmed, the backend must integrate with subscription billing systems (such as Stripe Billing or local payment APIs), inventory management modules, and logistics scheduling engines. The design focus here is on flexible fulfillment: allowing users to adjust the next shipment date at any time, pause specific items, or trigger formula adjustments based on real-time feedback. The entire process must achieve zero human intervention to minimize marginal costs.
3. AI Automation Solutions
In practical implementation, the following technology stack can be adopted. The frontend can utilize LINE Bot or a Web App as the primary interaction interface, integrating DialogFlow or a custom NLU module to handle natural language input, enabling users to quickly update their status through conversation. Each dialogue record is directly written into a time-series database (such as InfluxDB or TimescaleDB), serving as a data source for subsequent model training and real-time inference.
The formula computation layer can be established using Python + scikit-learn to create the initial recommendation engine, with training data sourced from users’ historical feedback and product effect annotations. As the data volume reaches a certain scale, AutoML tools (such as H2O.ai or Google Vertex AI) can be introduced for automatic parameter tuning and model selection, reducing maintenance costs. Inference results are output via API, with the frontend rendering a personalized monthly formula list.
The fulfillment side connects to a subscription management SaaS (such as Chargebee or ReCharge) to handle periodic billing and accounting, while the inventory system receives orders in real-time via Webhook, triggering picking and shipping processes. An advanced approach includes adding a predictive restocking algorithm that automatically sends procurement requests to suppliers based on the subscription users’ formula combinations and fulfillment cycles, preventing stockouts or excessive inventory.
Finally, in terms of data feedback loops, each time a user reports product efficacy (e.g., “this serum absorbs faster” or “this probiotic has no noticeable effect”), this feedback is immediately written back to the training dataset, allowing the model to continuously improve. This exemplifies the flywheel effect: the more active the users, the more accurate the recommendations, leading to higher retention rates and better data quality.
4. Revenue Expectations
From a financial model perspective, the core advantage of this architecture lies in the dual leverage of increasing LTV and decreasing CAC. Assuming a single transaction unit price of 1500, the traditional model may yield a repurchase rate of only 20%. However, with the introduction of personalized subscriptions, as long as the formula recommendation accuracy is sufficiently high and the fulfillment experience is smooth, the repurchase rate could rise to over 60%, tripling the LTV.
Examining the cost side, automated fulfillment can reduce the manual processing time per order from 15 minutes to nearly zero, and customer service demand significantly decreases due to transparent formula explanations and flexible adjustment mechanisms. Assuming maintenance of 1000 subscription users monthly, the traditional model might require three full-time customer service representatives and two warehouse staff, but with automation, only 0.5 personnel would be needed to handle exceptions, resulting in a 70% reduction in operational costs.
More importantly, the accumulation of data assets is significant. Each user’s state logs, formula history, and effect feedback represent highly valuable structured data. This data can not only optimize the in-house recommendation model but can also be licensed to upstream brands for product development references or packaged as anonymized industry insight reports for external sales. Assuming the derivative value of each data point is 10% of the monthly subscription fee, 1000 users could generate an additional data revenue of 150,000 monthly.
Overall, once the system is launched and operates stably, it is reasonable to expect that gross margins could increase from 35% in traditional retail to 55% in subscription models, while the CAC recovery period could shorten from 12 months to under 4 months. This transformation is not achieved through sales tactics or marketing strategies, but rather through architectural design fundamentally altering the leverage efficiency of the business model.