Companies That Use AI vs Companies That Build AI

The starting point was connecting the fit experience confirmed offline to online purchases. The confidence gained from touching and trying on clothes in stores was easily lost online, and this gap led to size failures and returns.

Initially, they used multimodal LLMs like Nano Banana, Claude, and GPT to quickly validate the flow from offline photography to digital model generation, app integration, virtual try-on, and purchase conversion. Thanks to the speed and tangible results, they could confirm "is this possible as a product," but at this stage, Musinsa was still closer to a company that uses AI well.

As the project took shape, the core question changed. Rather than one-off result images, they needed a structure where each individual's body-shape criteria would be saved and accumulated, and the offline experience needed to remain as a digital asset. Ultimately, the goal became not a feature wrapped around an LLM, but the underlying technology to turn experience into an asset.

The reasons for the strategic shift were also clear.

• Cost: As usage grows, the cost burden of multimodal LLMs increases.
• Prompt dependency: Results are hard to explain and control, and improvements tend to get stuck in prompt tuning.
• Limits to differentiation: If competitors use the same API, they can quickly build similar features.

So they decided to position the LLM not as the core of the product but as a development tool, while letting the core asset accumulate inside Musinsa itself. Currently, the O4O team is continuing its own AI project—including the Fit My Box (working title) prototype—under the 20% Project system, connecting body-shape understanding that starts offline all the way to online.

In the end, the choice became clear. Rather than AI use that ends as a quick experiment, the direction is to build Musinsa's own AI asset that accumulates body-shape data and experience.