Part 3: The Framework for a Solution
Most smaller retail locations use different POS systems, which makes it difficult to collect data from each register. We discovered this while working with both large chain retailers and individual stores. In nearly every case, the POS system lacked integration with third-party platforms. Files had to be manually exported and uploaded—making real-time store-level data collection nearly impossible.
If the end-goal is to predict whether a product or product line will sell at a given location, an AI system must be able to ingest diverse data types and workflows. Capturing sales data is only half the challenge; the other half lies in understanding how each organization’s merchandising and inventory strategies influence SKU-level sales. This would not be possible using today’s BI tools, but may be accomplished using large scale ML models that absorbs dozens of data points for each SKU, and applies those modeling to thousands of SKUs simultaneously.
The key challenges any collaboration framework must address include:
- The ability to obtain meaningful data from each location to build a profile location-by-location. (We believe that unless we can get per location data, our prediction models would not be optimized.)
- Converting available data into a usable format for analysis and AI model development.
- Understanding merchandise-mix decisions and modeling that process into a repeatable workflow.
- Modeling decision-making behavior, whether driven by local store managers or a centralized HQ purchasing team.
- Capturing product characteristics from the brand to map against characteristics of existing products at each location.
- Associating product-location matches with a predictive score for each SKU at each store.
Trying to tackle all these steps at once is overwhelming. Instead, we advise our brand customers to start by building a registry of retail locations that carry their products. As each location is identified and added, CRSTBL profiles the store to create a unique personality from a product success prediction (PSP) perspective. Essentially, we’re building mini-AI models for each store that generate predictive success scores for every SKU. We then implement tools to allow brands to further build the relationship with each location so they can enhance each location’s PSP accuracy.
Over time, as each location’s PSP accuracy expands to cover more products and product types, brands can use the model for other useful purposes, such as running new product concepts, marketing programs, discounts and incentives, etc. through the model to see how likely their designs will “work” at each location. At scale, brands can virtually test-market a product across thousands of stores using CRSTA AI, predicting where the product will succeed and at what price. They can then launch new products in the stores and channels that exhibit the best chance for success.