Six people now do the work of 58: how a Korean startup went from a $1.4M loss to profit in six months as an "agent organization"
Nov 11, 2026Original 티타임즈TV(송인혁 유니크굿컴퍼니 대표) · Microsoft Work Trend Index · MIT NANDA ↗
In 3 lines
- Six people do what 58 did and the company is profitable; the key was moving from personal AI to organizational AI
- Role-based agents failed; what worked was one agent called from a Slack-like chat, with no new habits to learn
- MIT blames a "learning gap" for the fact that most enterprise AI pilots show no measurable return
In January, Uniquegood Company, the Seoul startup behind the location-based game platform Realworld, had 58 employees and a loss of about 2 billion won (roughly $1.4 million). When CEO Song In-hyuk told the staff how bad things were, 80% left within two weeks. Today six people, including the two co-CEOs, run the same business. In two interviews posted by TTimes TV on Nov. 9 and 10, Song said "we didn't cut people because of AI" and walked through how the company turned profitable in six months. We read his account alongside Microsoft's Work Trend Index and an MIT report.
Losses grew as headcount neared 60: "Whose job is this?"
Uniquegood makes games in which players become the lead character of a story and solve missions in physical spaces. A single venue in Seongsu drew 100,000 visitors a year, but every new project needed more people, and each time the company neared profit it slid back into the red. Song said the company was profitable when it was small and the losses grew as it approached 60 people. Work that belonged to no team, such as getting the company known, kept piling up, and when he tried to assign it the answer was "that's not our team's job." He said this happens in any organization that grows.
80% left in two weeks, and the handover exposed a black box
On Dec. 23 last year, a partner pulled out of a large project the company had been preparing. In January Song gathered the whole staff and disclosed everything down to the cash balance; 80% left within two weeks. The entire development and design teams resigned. As the remaining 12 took over the work, he noticed a problem. Every employee had a paid AI account, yet there was no way to know who had prompted what or how an output had been produced. Individuals got faster with AI, but the know-how disappeared when they did.
Role-based agents failed: keep the habits, call one name
The team first tried a knowledge graph and role-based agents such as a "planning agent." Neither worked. People couldn't tell which agent to call, so the old "whose job is this?" problem came right back, and staff pushed back against learning yet another tool. Song drew two conclusions: don't make people change their habits, and take apart the individual black box. So in a Slack-like chat, people call a single agent named AXIT, which summons other agents as needed, splits up the work and remembers every instruction and conversation.
From a 15-language docent to a billion tokens a day
Confidence came from a small docent project for the Seoul International Garden Show. The team built it for the web, something the old development team had said couldn't be done, and it grew into a management system usable across exhibitions with voice guides in 15 languages. The company now uses about a billion tokens a day. At API prices that would be around $8,000 a day, though Song said the real cost is lower because subscriptions are mixed in. When users submit improvement requests, an agent picks them up and ships them; he said the system is upgraded about 50 times a day.
The support person now ships the fix: "nano ventures" and those who didn't cross over
Song calls this shape, one employee running one business, a "nano venture." The customer experience staffer who used to only relay complaints inward now fixes issues with AI on the spot, and employees spend most of their time with customers. Token usage is visible to everyone, and he said the heaviest users clearly performed best. Not everyone made the jump. As the gap widened day by day, some colleagues lost motivation; the company ran extra training, but 12 people became six. He said that part was painful.
"It's not an AI problem": being broadcast and being found come first
In the second interview he said the core of AI transformation has nothing to do with AI. The essence of a business is whether the loop of getting the product out and getting it discovered by customers keeps turning, yet most talk is about individual productivity. Even with 58 people, he recalled, meeting customers and taking rejection was left to the two CEOs while everyone else spent their time producing documents internally. He said he had mistaken that for productivity. About 80% of the company's customer inquiries now arrive through AI search.
Song In-hyuk, CEO of Uniquegood Company · YouTube · TTimes TV ↗Two people, 1.2 billion won a month; a hospital director cut no-shows by 80%
He gave two examples. An acquaintance who had led a 1,000-person retail company quit, started exporting containers with one friend, wired the whole process together with AI and passed 1.2 billion won (about $850,000) in monthly revenue. The director of a general hospital with roughly 300 staff transcribed three months of call-center recordings, added conditions such as weather and underlying illnesses, and built an algorithm that predicts no-shows. By changing how reminders were sent, the hospital cut no-shows by about 80%. Neither had a development background. The figures are as Song relayed them.
Song In-hyuk, CEO of Uniquegood Company · YouTube · TTimes TV ↗MIT's "learning gap": building in-house succeeds a third as often
The GenAI Divide, a report released last year by MIT's NANDA initiative, found that only about 5% of enterprise generative AI pilots achieved rapid revenue acceleration. It drew on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public deployments. The cause it named was less about model quality than a "learning gap": tools that don't learn from or adapt to workflows. That echoes Song's individual black box. The same report found that buying tools from specialized vendors succeeded about 67% of the time, while internal builds succeeded only about a third as often. Read that way, not every company should build its own AXIT.
MIT NANDA · Fortune ↗"A company that uses AI, or a company that runs on AI?"
In its 2025 Work Trend Index, Microsoft called companies where humans and agents work as one team "Frontier Firms" and predicted every employee would become an "agent boss" who builds, delegates to and manages agents. Eighty-two percent of leaders said they expect to use digital labor to expand their workforce within 12 to 18 months. Song's test for an organization is simpler. Can the organization see its members' conversations and actions? Does an agent remember that record and improve on it by itself? Those two questions, he said, separate a company that uses AI from a company that runs on AI.
Microsoft · Official Microsoft Blog ↗
