PoC (Proof of Concept)

## Differences from Prototypes PoC and prototype are often confused, but they serve different purposes. A PoC verifies "whether something is technically feasible," regardless of appearance or usability. A prototype verifies "whether something works as a user experience," and is often conducted after the PoC. For example, in a PoC for an AI chatbot, it is sufficient to connect to an API and measure response accuracy. A minimal command-line interface is perfectly acceptable for the UI. Screen design and user flows are only developed in detail at the prototype stage. ## How to Conduct a PoC The process generally follows these steps. First, clearly articulate the hypothesis to be validated—in a specific and measurable form, such as "Using RAG to search internal documents will reduce inquiry response time by 50%." Next, build a minimal system configuration and collect data to validate the hypothesis. The process typically takes two to four weeks. ## Common Traits of Failed PoCs There are several patterns in which PoCs fall apart. Expanding the scope of validation too broadly, having vague success criteria, and validating with sample data instead of production data—when these factors combine, the result is often a situation where "the PoC succeeded, but it couldn't be used in production." This is especially true for AI-related PoCs, where the quality and volume of training data have a significant impact on outcomes. Even if 90% accuracy is achieved with 100 sample records, it is not uncommon for accuracy to drop sharply when applied to tens of thousands of records in production. Using data that closely resembles production data from the PoC stage onward is key to preventing costly rework.