The AI transformation underway at Musinsa

For many years, Musinsa, South Korea’s premier online fashion platform and marketplace for Korean designer brands, streetwear, and beauty products, has described itself as a company dedicated to AI. They bring in AI-native talent and give employees room to experiment, yet meticulously manage token usage and work performance.

And instead of handing over AI completed by engineers to the field, they have the field restructure their own work using AI. So if external solutions are deemed lacking, they analyze the cause and create their own solutions.

Gil Gi-yong, the company’s director of core AI and CX engineering, began actively utilizing AI shortly after joining Musinsa last year when he saw AI evangelists, from the CTO to new employees, directly create results. He says that Musinsa changes faster than organizations he’s experienced before, so he’s diligent about anything new that comes along to augment its developing internal AI platform.

Director of operations Kim Dae-ho is on the same page, too. Kim, who joined in September 2022, currently leads the operations strategy team that’s responsible for planning and executing tasks to improve customer and partner experiences, and improve operational efficiency across Musinsa’s commerce ecosystem. Recently, most of his work has been closely related to AI, so he’s collaborating with the development team to lead AX projects.

Musinsa is investing so much in AI is because trends change rapidly and fashion consumers can be impulsive. So platforms have to be fast, anticipatory, and intuitive to meet demand. In addition, they should support areas such as sales management for partners, and AI is the means to meet these needs and boost business competitiveness. That’s how Musinsa has started solving unique fashion industry problems.

A prime example is how outfit photos are shot. To showcase products in the past, models were hired, they dress themselves, and it’s all shot in a studio. Now, AI is used as a coordinator to generate images and the goal, according to Gil, is to elevate visual AI tech to a top standard by H1 2027.

Plus, AI is used in the product registration process of partner companies, so when a product is registered, AI automatically infers categories, colors, materials, and more. In the past, these tasks were manually entered. Holistic AI has also been integrated into customer service centers. The AI chat agent operated by 29CM, a South Korean fashion and lifestyle platform operated by Musinsa, currently handles about 25% of all customer inquiries.

From business operations to ROI

Companies typically approach AX in two ways. One is for the engineering team to get the requirements of the field, develop the system, and then say try it, and pass it on. If it’s hard to solve internally, the other way is to leave it to an external vendor. But Gil sees that AX often fails in these two approaches.

“The reason is that engineers do AX, but they don’t fully understand the processes of the field,” he says. “That’s why we focus on creating an environment where field experts can directly improve the actual work processes.”

In fact, Musinsa’s AX-related ideas are directly proposed by each business team. For example, the planning and engineering departments review how to improve on repetitive tasks with LLMs by setting quantitative goals such as reducing tasks from 270 man hours per week to three.

The core of the approach Musinsa choses is next. Engineers don’t write scripts for them. Instead, the core AI team suggests guides until target figures are achieved through repeated testing. Gil says this process ultimately helps the business operator measure how to write prompts and what improvements they use to improve them.

This approach has also changed the process by which AI spreads. Instead of copying systems created by developers, colleagues directly use AI to improve actual work. As small success stories accumulate and are shared, overall AI utilization capabilities also increase.

Gil adds that this approach must align with the direction of leadership. Simply sending tasks down by saying try using AI and costs will be covered isn’t enough to drive adoption. At Musinsa, many leaders, including the C-suite, first experiment with AI and freely share results. This culture encourages voluntary experimentation and accelerates the spread of AI. 29CM’s AI chat agent was reputedly developed using this method.

In fact, this tech was developed in collaboration with outsourced companies, and produced a certain level of output, but the strategy was revised to in-house development. Those results were actually better, according to Gil, and they succeeded in launching quickly while reducing costs to about a third of the outsourcing method. The development period was also about a quarter of similar projects at other companies.

However, this speed isn’t simply due to AI. ROI verification is thorough and Musinsa first fixes timelines and budgets, then prioritizes requirements to implement within that framework.

Even before starting, cost issues are carefully considered. For example, they calculate a five-year TCO and first check when investment costs can be recovered as savings. When using external services, they don’t sign long-term contracts based on simple discount benefits. Kim says that starting a project based solely on rough cost estimates may result in significant investment but doesn’t achieve the expected efficiency.

For AI chat agents, they quantified the estimates generated by the technical team, and simulated how much customer satisfaction would change and inquiry rates decrease. Investment costs were also calculated by year and month.

After implementation, monthly tracking is conducted to see how costs differ from expectations, how the AI processes inquiries, and how close performance is to the original target. If differences from expectations occur, improvements are made, including which variables to add, if the scope of inquiries handled by AI should expand, and if agent workflows should change.

Generous support, but objective evaluation

For this approach to continue, a supportive organizational culture is necessary. First, Musinsa provides a substantial amount of tokens so its members can fully utilize AI. But token usage is actively managed, so each leader regularly checks it using separate monitoring tools. If too many are used for tasks that could be handled with fewer, it’s looked into. Conversely, tasks that could be completed much faster with AI are still subject to inspection even if they’re handled using traditional methods.

Kim says that ultimately, what matters isn’t whether AI was used, but what results were achieved and how many resources were invested in the process. If AI can improve the quality of deliverables and reduce time in the work, it should be actively utilized. But the costs invested and the results achieved should be considered together.

Naturally there are times when using AI when multiple rounds of testing are gone through before desired results are reached. This means more tokens and time are required, and costs increase. But Musinsa focuses more on what trial and error was experienced, what was learned in the process, and what can be done differently next time. Failure is also seen as data that increases the chances of success next time.

AI leader standards for talent

What kind of talent do AI-led companies need? Not so much proficiency in AI itself but the ability to properly define problems. Kim defines planning as a journey of problem-solving, meaning a series of processes to identify them, develop solutions, and execute. It’s become even more important for humans to define what to ask of AI. If the problem definition is ambiguous, no matter how plausible the AI is, it’s difficult to apply it to actual work.

Of course, acquiring these skills isn’t easy. Kim says that training to constantly ask why is helpful. He doesn’t just accept initial causes as they are, but searches until a real solution emerges.

“At first, it’s confusing,” he says, “But through intense discussions and debates, capabilities develop. If the opponent isn’t a human but an LLM, you can repeat the same process much faster.”

Gil’s standards to evaluate development talent are similar. He believes people who can think differently from traditional methods are now needed, and relying solely on experience and technical expertise is no longer a sufficient nor competitive advantage.

In fact, this is reflected in Musinsa’s AI native developer recruitment. The focus isn’t on simple algorithm proficiency, but on how well the applicant defines real-world problems, grasps the essence of the problem, and uses certain indicators to determine whether to solve the problem, even when freely using AI.

The principles of leadership aren’t much different. Kim defines his style as autonomy-based leadership, while Gil prefers to clarify responsibilities and authority. The common goal, however, is to have an organization that doesn’t rely on any single individual.

In terms of next steps in AI transformation. Gil believes that competitiveness depends on who’ll first leverage new technologies, and how quickly that experience will transform into organizational capabilities. But AI led by tech organizations has its limits. Models that worked in commerce don’t work offline as is, and a single model can’t transplant across all areas of the company. So each organization must handle AI based on its own domain of expertise.

This is where Musinsa is headed. It’s not a company that relies on a few AI experts, but a place where anyone can wield AI in their work. “I want to establish a company where every employee solves their own problems with AI, and accumulates that experience as an organizational asset to show we use it well,” says Gil.