The Field Asks First: How Far Can AI Go?

The Job That Now Takes 7 Minutes in a Fragrance Warehouse

At a fragrance manufacturer in Seoul, balancing inventory used to consume half of every month. Because each fragrance has different properties and expiration dates — and some raw materials can substitute for others — the person in charge had to work through every variable by hand to calculate optimal stock levels, a process that took fifteen working days a month. After the company integrated an AI inventory management system from domestic startup Deepflow, the same task was cut to seven minutes. Demand-forecast accuracy improved from 70 percent to 80 percent.

An electronics manufacturer using the same solution reported a 49 percent reduction in stockouts, a 70 percent cut in excess inventory, and monthly inventory cost savings of approximately 1.1 billion won. The first on-the-ground fact uncovered by this reporting: the cases that produced the most dramatic numbers were, strictly speaking, not 'generative AI' but machine learning specialized in demand forecasting — not a conversational model like ChatGPT, but a predictive engine trained on lead-time variability and seasonality. The AI that has actually saved money on the shop floor and the generative AI everyone talks about these days may share the same label, but they are different things.

Why a Seongdong-gu Café Owner Stopped Calling Designers

The clearest picture of generative AI entering a small shop appears on the marketing side. Kim, a business owner in her thirties running a dessert café in Seongdong-gu, Seoul, used to outsource every promotional piece to a designer — an expense that was always a burden. Now Kim uses AI to retouch menu photos, design posters, and produce short promotional videos, all on her own. 'Marketing costs have dropped significantly,' she said. Another self-employed person running an online shopping mall said using AI to write product descriptions, draft customer inquiry responses, and analyze reviews had led to 'noticeably higher work efficiency.'

These are not grand systems. This is tangible automation — tools built on ChatGPT that draft a Naver Place review reply and post it in under ten seconds. The government is also channeling money into this space. In February 2025, the Korea Small Enterprise and Market Service (SEMAS) overhauled its 'Small Business Knowledge Hub,' adding a function that uses generative AI to summarize the key points of lectures, while also expanding programs supporting small-business owners in generative AI literacy and digital transformation.

But Half of Manufacturing Sites Are Still Standing at the Door

While the small-business sector is moving quickly, small and medium-sized manufacturers are far more cautious. A survey released by the Korea Federation of SMEs (KBIZ) on October 19, 2025 captures the mood. When 502 small and medium-sized manufacturers that had participated in the large-SME cooperative smart factory program over the past five years were asked, 47.4 percent said AI adoption in manufacturing processes was 'necessary.' Including those who answered 'neutral,' 78.5 percent expressed a positive view. The message: awareness of the need is there.

The problem comes next. An overwhelming 44.2 percent cited 'upfront cost burden' as the main reason AI adoption is difficult, followed by 'lack of specialized personnel' at 20.5 percent. What they want from the government was equally concrete: 'direct financial support' topped the list at 72.3 percent, followed by 'AI-specialized consulting' at 21.9 percent. Even among those already running smart factories, 'lack of specialized operating staff' (43.8 percent) and 'high maintenance costs' (25.9 percent) were holding operations back; and the top reason given for being unable to analyze process data was also 'insufficient dedicated staff and specialists' (50.4 percent). The gap between the seven-minute success story from the fragrance warehouse and these survey results is the true position of Korea's small and medium-sized manufacturing sector: there are models to follow, but no people or money to follow them.

A Surprising Twist: Hours Fell but Output Didn't Rise

At this point, the direction of this reporting needs to shift. Collecting adoption stories alone easily leads to the conclusion that 'AI is an unqualified gain.' But data released by the Bank of Korea carries the opposite warning. A Bank of Korea issue note analyzing roughly three years since the launch of ChatGPT found that workers who used generative AI saw their average working hours fall by 3.8 percent — equivalent to about 1.5 fewer hours per week on a forty-hour workweek.

What is striking is that those saved hours did not translate into greater output. At the individual worker level, the correlation coefficient between the rate of working-hour reduction and the rate of increase in work output was zero. The Bank of Korea calls this a 'productivity disconnect.' 'AI raised efficiency at the level of individual tasks, but it did not extend to improving workflow, changing organizational structures, or redeploying personnel' — leaving time emptied while performance stayed flat. The only groups where productivity genuinely rose were, as exceptions, the self-employed, professionals, and a small minority of very heavy AI users.

This is the critical passage. A self-employed person like Seongdong-gu café owner Kim, who controls all decisions alone, can immediately convert saved time into more marketing or higher sales. In an organization, however, those saved 1.5 hours evaporate into reports and meetings. The data thus provides a structural explanation for why generative AI works better for small-business owners than for larger organizations.

What Didn't Work — and One Line of Dissent

What failed must also be recorded honestly. Hallucination — generative AI's chronic flaw of producing false information in a confident tone — poses the same danger even in small shops. These are overseas cases, but Air Canada faced a consumer compensation ruling after its AI chatbot provided incorrect discount information, and 'MyCity,' a business-guidance chatbot for New York City backed by Microsoft, generated responses that encouraged illegal conduct, including suggesting it was acceptable to pocket employees' tips. For a small-business owner who has handed customer service entirely to AI, this is far from someone else's problem. The approach recommended in the field is therefore a hybrid: AI handles only the most frequently asked inquiries, while complex issues are passed to a human.

There is also a clear counterargument. In a survey released by Korea IDC in October 2024, 78 percent of domestic organizations reported productivity improvements from generative AI, and LG Electronics created an in-house AI that writes SQL code enabling hundreds of terabytes of data to be analyzed without any IT knowledge — a view that AI skepticism is overblown. But that 78 percent reflects 'improvements perceived by adopting organizations themselves,' while the Bank of Korea's correlation coefficient of zero is 'the result measured in actual output.' For the same technology, subjective experience and objective measurement diverge. The reason an SME factory owner hesitates in front of a consulting invoice may be precisely because they instinctively sense the gap between those two numbers. The seven minutes in the fragrance warehouse are real — but not everyone can afford to buy those seven minutes.