Start with AI, not with IT. Korea's realizable wants to bring AI to factories one industrial park at a time

How far has AI really made it onto the factory floor? I haven't been able to stop asking myself that since I got back from Korea.

In September, I joined Korea Nomad Gateway, a ten-day program that took us through Incheon, Busan and Jeju. Incheon, where we spent the first three days, is best known for its airport and its new skyline of towers. But it has another side: the Namdong National Industrial Complex, home to more than 8,000 companies and about 80,000 workers. It is a city of factories.

In Incheon, someone told me about a service called "AI 공장장" (gongjangjang). It translates to "AI Factory Manager." In Japanese, the word for factory manager, kojocho, is written with exactly the same characters.

A factory manager that is an AI. What would it actually do? Looking into it, I also got a clearer picture of factories back home in Japan.

According to a survey by Japan's Organization for Small & Medium Enterprises and Regional Innovation, only about one in six small and mid-sized Japanese manufacturers use AI. And even among the smaller companies that do use it, only about a third use it in manufacturing or production. Far more use it for admin and back-office work.

AI still hasn't reached the machines. But it isn't a question of motivation. Factory digitalization follows a set order. First you install a system to manage production. Then you collect data and standardize it. AI comes last. Follow that order and it takes years, with the money going out first. For a small plant, it's an order that wears you out before you ever reach AI.

One company is trying to turn that order upside down: realizable, a start-up with offices in Incheon and Seoul, and the maker of "AI Factory Manager."

We had hoped to meet while I was in Korea, but the timing didn't work out. So after I got home, we set up a long online conversation, linking Osaka, Hiroshima and Korea. We booked an hour. It ran well past that.

And so, we got to sit down with realizable's CEO, Dongmyoung Won, himself!

What realizable wants to change comes down to two kinds of order.

• AI first, not last
  Instead of waiting until every system is in place, start by using AI to make just two or three key numbers visible. Add the data and systems you need afterward.
• Park by park, not one plant at a time
  Partner with a "coordinator" such as an industrial park or a manufacturers' association, and test with 10 or 20 plants at once.

Change the order, and the plants with the fewest big systems become the easiest ones to start AI with. Let's take it one step at a time.

WHO IS

Dongmyoung Won

Founder & CEO, realizable

“I want to recreate, with data, the calls a veteran plant manager makes from years of experience.”── Dongmyoung Won (from this interview)

1. 2010–2017From plant imports to overseas businessManaged plant imports as an expatriate in the CIS region for KCC Engineering & Construction, then joined the overseas business team at Daewon Industrial
2. 2017–2021Purchasing at a French-owned auto parts makerHandled development costing and parts purchasing on the procurement team at Faurecia Korea
3. NOWFounder & CEO, realizableFounded the company in March 2022 and builds the manufacturing AI 'AI Factory Manager' from an industrial park in Incheon

Won studied Northeast Asian trade at Incheon National University, with a year as an exchange student at Moscow State University. After graduating, he worked on the front lines of manufacturing deals: overseas plants, exports and auto parts purchasing. He founded realizable in 2022.

It reads the data from the shop floor and tells you what to fix next.

The name means exactly what it says. A veteran manager who has watched a plant for years will say, "Something's a little off with this line today," or "This time of year, start with that machine." Our starting point was to recreate those calls with data.

You can. What a plant manager is watching in their head is actually numbers.

How long a machine ran without stopping. How many defects came out. How long a piece of equipment went between breakdowns. These are a factory's report card, and we call them KPIs. Veterans read them from experience.

The problem is that in most small plants, those numbers are scattered across paper, Excel and the plant manager's memory. So when that person leaves, the judgment leaves with them.

Big companies are pushing ahead on their own. But in every country, most of manufacturing is small and mid-sized plants, and they are being left behind on AI. I think that's the biggest problem for a country's manufacturing competitiveness.

The owners who run these plants can't picture how AI would change their work. There's no textbook for it.

Go onto the floor and it's even clearer. Workers wear gloves, and their hands are covered in oil. Tablets are awkward to use, so records are kept with pen and paper. Later, someone in the office types it all into a computer. In Korea you still see this, not just in small plants but even on the floors of big companies.

Asking those plants to follow the same order big companies took 20 years to walk just doesn't get very far.

Right. The world's leading manufacturers have spent more than 20 years building up their digital foundations. They installed MES for production management, ERP for company-wide operations and FEMS for energy, standardized their data, and only now are they putting AI on top.

The World Economic Forum's network of cutting-edge "Lighthouse" factories has grown to 238 sites worldwide. Many of their most successful initiatives use AI.

The question is whether you can tell a small plant standing at the starting line, "Please walk the same 20-year road, in the same order." Every step costs money, people and time. While you guide one plant at a time through every stage, the gap keeps growing. By the time you catch up, the leaders have already moved on.

So I believe we have to change the order. Instead of finishing digital transformation and then adding AI, put manufacturing AI in first, and fill in the data and systems you need working backward.

The first thing you enter is the name of one representative product.

Say you type in "auto door parts." The AI proposes the standard processes for that product, the KPIs worth watching and the typical equipment setup. The plant's staff just adjust that to fit their own plant. They don't start from a blank page.

That works because we've organized 260 industries, 50,000 processes and 8,000 KPIs into modular building blocks and trained the AI on them. It starts from a base learned from Korean manufacturing data, then adapts as each plant's own data is added.

In Korea, we already do this in training programs run by public agencies. Participants bring just a laptop or a smartphone. Getting two or three key KPIs onto a dashboard takes 30 minutes.

That's always the first question. The answer is an international standard.

We didn't invent our KPI framework. We use ISO 22400, the international standard for KPIs in manufacturing operations. It's designed to work across industries and defines 34 KPIs across production, quality, maintenance and inventory.

Availability, defect rate, MTBF, inventory turnover. The industry may differ, but the yardstick is the same.

Let me share three voices from customers in Korea. We keep the company names private.

VOICE

At 3 a.m., the AI reacted to an unexpected change in temperature and adjusted the warehouse temperature. Nobody had to come in, and our defect rate went down.

VOICE

The AI reshuffled our machine schedule. Same output, and our electricity bill fell 15%.

VOICE

The AI found the bottleneck and redesigned our work sequence. Our output per hour went up.

None of these came from replacing a company-wide system. They came from picking two or three KPIs and starting there.

At 3 a.m., in an empty plant, the AI fixed the temperature. When I read that line, I caught myself saying, "Now that's a plant manager." Noticing something in the middle of the night and acting on it is exactly what a veteran does.

We don't try to get rid of paper overnight. We go in steps.

Before electric cars, there were hybrids, right? Factories are the same.

Take inspection: checking whether a product meets its specs. It used to be that a person compared the product with the spec sheet by eye and ticked boxes on paper. In the next stage, a camera reads the handwritten sheet and enters it into the system automatically. No more typing in the office.

In the stage after that, the paper disappears. A camera with sensors and the AI look at the product directly and decide whether it meets the spec. If it passes, it moves straight on to the next process.

You take paper out of one process at a time. In Korea today, smoothing those in-between steps is part of AI's job too.

That's the most delicate part. You can tell owners everything we've just discussed. But without cooperation from the people on the floor, a project like this goes nowhere. And you can never walk onto a shop floor and say, "We're cutting people."

There is a way to move forward without a clash, though. Many plants struggle to attract young workers, and only the veterans are left. In plants like that, we present it not as "AI to replace people," but as "AI to support the places people no longer come to". We turn people's experience into a standard anyone can use.

It may sound harsh, but if they don't act now, the plant itself may not survive. We move forward while confirming that reality together with the people on the floor.

That's why we propose something we call "Multi-PoC." A PoC, or proof of concept, is a trial before full rollout.

Instead of running trials one company at a time, we work with organizations that bring plants together, and test with 10, 20 or 50 plants at once: industrial parks, manufacturers' associations, chambers of commerce, public support agencies, or a large company's network of suppliers.

I'm not against one-by-one trials. The issue is speed and reach. If you want to change 100 plants, do you really need to start from scratch 100 times?

Going plant by plant also means hearing the same worries every time, from scratch: "Is our data really safe?" When a coordinating organization is involved, it can take on that explaining and coordinating. We can focus on the trial itself. In Korea, we've even partnered with a real-estate company that manages an industrial park.

That's a natural concern, and it's also why not just anyone can run a Multi-PoC.

Running several trials at the same time and handling several companies' data in a single system are completely different things. You have to design three boundaries up front.

• A data boundary
  Separate the data that can be shared across companies from the data each company must keep to itself.
• A policy boundary
  Decide which data can feed shared features and which data must stay inside each company.
• A system boundary
  Many companies share one platform, but each company's data and AI environment stay safely isolated.

Multi-PoC isn't a sales tactic for bundling trials together. It's the work of building a manufacturing-AI structure that many plants can safely use together.

I see three benefits.

First, a single program can support many companies at once. That spreads the impact further than supporting companies one by one.

Second, you find out quickly which manufacturing AI actually works on the floor. Test at 10 plants at the same time, and you can compare them with the same yardstick.

Third, and most important, you're left with a manufacturing-AI model built for that region and that industry. With one-by-one trials, the results stay inside one company. But when data builds up from many plants at once, it becomes an asset for the whole region.

In the long run, we see shared manufacturing AI as infrastructure, and company-specific AI as a service on top. Factories don't each build their own power grid or water supply, after all.

Yes. The people who actually use the AI are the individual manufacturers. But the customers and partners we mainly have in mind are industrial-park operators and developers, manufacturing associations, and public institutions.

Here's the model we're aiming for:

• The coordinator delivers manufacturing AI as shared infrastructure
  A park or an association offers manufacturing AI to its tenant or member companies as a common digital foundation.
• Each factory's AI keeps getting better on top of it
  On that foundation, company-specific AI services are refined further to match each manufacturer's needs.

We don't want manufacturing AI to belong only to a few advanced companies. Any company inside an industrial park or a manufacturing community should be able to use it. We hope it becomes everyday, universal industrial infrastructure.

And the key to making that shared infrastructure truly useful is real-time data.

There's a lot of debate. Will physical AI, AI with a body like a robot, take over the factory? Will AGI that can do anything a human can appear, and its maker do everything? Or will a big data-analytics company lock up the whole market?

I see two clear paths.

One is that no AI can succeed without access to real-time data. However smart an AI is, it can't run anything if it can't know what's happening in the plant right now. The more plants it touches, ten thousand, a hundred thousand, a million, the better it can run. Multi-PoC and our work with industrial parks are also about bringing in as much shop-floor data as possible to make the AI better.

The other is a change in how money is made. Until now, success meant one excellent service winning the market. In the AI era, I think real revenue will come from transactions across an entire ecosystem of connected AIs.

It may sound a bit paradoxical, but I think the opportunity lies in the fact that they haven't installed big systems yet.

Large companies can't tear down what they've spent 20 years building. Japan's Ministry of Economy, Trade and Industry made this point in its 2018 DX Report: about 80% of IT budgets go to maintaining and running existing systems. Only 20% is left for anything new.

Plants without big core systems aren't tied down that way. They have Excel, handwritten records and a few sensors. That's not "falling behind." It also means they still have the freedom to design.

Of course, many plants already have MES, ERP or FEMS. It's the same for them: rather than replacing what they have, we make it usable. AI Factory Manager connects to those systems and also to Excel, handwritten documents and sensors.

In Korea, I've heard people say, "We installed an MES, but it's still not used on the floor." It's not that the system is missing. It goes unused because there's no baseline to compare against.

We're aiming to enter Japan and Southeast Asia at the same time. In Southeast Asia, we're in talks with potential partners, mainly in Singapore.

Japan is a country with a long history of making things and very high skill on the shop floor. That's exactly why I don't expect a foreign start-up to earn trust right away by showing up and saying, "Let's add AI."

So rather than going in alone, we want to spread that trust across several partners: companies with AI capabilities like ours, companies that know how to talk to Japanese factories, and companies that manage industrial parks. People in Japan understand Japan far better than we do.

In Japan, too, we don't want to stop at selling solutions to factories one at a time. We want to work with industrial-park operators, manufacturing associations and public institutions to bring manufacturing AI in and spread it widely.

In fact, Japan already has a precedent. Japan's White Paper on Small and Medium Enterprises describes how eight small companies, including Hiroshima Metal Work in Hiroshima, jointly developed production-management software. Development too big for one company, done together by several. That's the very idea behind Multi-PoC. This time, there's no need to build new software from scratch. AI Factory Manager already provides a shared foundation.

In Korean, there's a word, 철부지 (cheolbuji). Written in Chinese characters, it means "someone who doesn't know the seasons." It's said to come from farming days, when knowing when to sow and when to harvest was the wisdom you needed to survive.

I believe the season for manufacturing AI is spring right now. Not every technology is finished. That's exactly why now is the time to sow, not the time to judge the harvest.

And the seeds of manufacturing AI aren't company-wide systems. They're the two or three KPIs a single plant chooses for itself.

Even in farming days, not everyone studied the calendar alone. Experienced people told those around them, "It's about time to sow." In today's manufacturing, I think industrial parks, manufacturers' associations and public institutions can play that role. They see the market more broadly than any single plant.

If any parks, associations or local governments are interested, I'd love to talk. Choose one plant, and we'll show you on the spot, from typing in a product name to a working KPI dashboard. Seeing it on screen is faster than any slide deck.

When you've just moved into a new apartment, you can place the furniture however you like. Try that in a place you've lived in for ten years, and you have to start by moving the big bookcase. Factory systems are a bit like that.

Japan's DX Report noted back in 2018 that about 80% of Japanese companies' IT budgets went to keeping existing systems running, and that for more than 40% of companies, the share was over 90%. These are eight-year-old figures, but the pattern, where bigger systems make it harder to move, matches what Won meant by "can't tear down."

Installing a system isn't the whole answer, either. Among Korea's small and mid-sized manufacturers with smart factories, 75.5% are still at a basic level, and 99.8% have connected only part of their processes. The systems are in, but people still make the decisions. "We have an MES, but we don't use it" is the flip side of those numbers.

So Japan's 16.4% isn't simply a lag. Plants without heavy systems are also in a position to start with the newest way of working. What matters is less what you install than what you look at first.

A health checkup is useful because blood pressure and blood sugar are measured the same way at every hospital. That's what lets you compare yourself with last year, or with the average for your age.

Factory KPIs work the same way. A plant that measures availability as "time the machine was switched on" and one that measures it as "time it was actually producing" can show the same number and still not be comparable. ISO 22400 exists to align those definitions. The Korean government is building its 240 manufacturing process reference models on the standard's KPIs for the same reason: so plants can be compared.

I also checked the Japanese precedent Won mentioned in the white paper. Eight small companies, including Hiroshima Metal Work, met through a course for business successors and, in 2003, jointly developed production-management software with help from Shizuoka University. After rolling it out across all processes in 2017, sales per employee rose 8.6% and working hours fell 15.9% by 2021. The defect rate was reported to have dropped by 97%.

More than 20 years ago, those eight companies had to start by writing the software themselves. Today, the shared yardstick and the AI foundation are already in hand. As I see it, the real value of Multi-PoC isn't just testing together. It's being able to compare, on the same yardstick, which AI actually worked.

Japan already has plenty of organizations that bring plants together. According to the Japan Chamber of Commerce and Industry, there are 516 chambers of commerce nationwide. Towns and villages have their own commerce and industry associations, 1,587 of them by the national federation's count. Add industrial parks across the country and industry associations for each sector. The seats for Won's "people who tell others when it's time to sow" are already there.

Korea is moving toward that "bring plants together" model as a country. In June 2026, it announced a Manufacturing AI 2030 Strategy to invest 20 trillion won from public and private sources by 2030. In August, it reported that across 42 sites that adopted AI, productivity rose 30.1% on average and defect rates fell 15.5%, and set targets of 500 AI factories and 25 AX demonstration industrial parks by 2030. The government explicitly named the loss of skilled workers' tacit knowledge as a national problem behind those targets.

"Park by park" isn't just one start-up's sales idea. It's also a unit the Korean government has started setting targets in. In Japan, too, more and more plants are short of workers. Grow manufacturing AI at the level of parks and regions, and make it as easy to use as electricity or water. The future Won describes may be closer than we think.

What to watch from here depends on where you sit.

• Local governments, chambers of commerce and industry associations
  Instead of handing out subsidies one company at a time, could you design a program that tests 10 plants at once? If results can be compared on the same yardstick, they also justify next year's support.
• Plant owners
  Before weighing a big system, pick the two or three KPIs you want to see every day. That alone makes it clear what to ask AI to do.
• Engineers and data professionals
  Aligning definitions with ISO 22400, safely separating many companies' data and connecting shop-floor data in real time: manufacturing AI has become a contest of careful, unglamorous design.
• Students
  Manufacturing is one of the places where AI still has the most room to grow. In both Japan and Korea, as fewer people work in factories, there's demand for the people who build these systems.

Factory AI is shifting from a tool one company buys to infrastructure a region grows together. That's the view I most wanted to bring back from this interview.

Factory digitalization has always followed a set order: systems, data, then AI. Follow it step by step, and small plants run out of energy before they ever reach AI.

What realizable does is swap that order and turn a plant-by-plant journey into a park-by-park one. Enter one product name. Make two or three KPIs visible. Test at several plants at once and compare them on the same yardstick. Bit by bit, a plant manager's experience becomes data, and eventually an AI that every plant in the region can use.

By the end of the interview, the conversation had already moved on to what comes after this article. How to rework the explanation materials for Japanese plants. Which parks and associations in Japan could take the first step with us. Won was clear that he wants to rebuild his materials to fit the way things are communicated in Japan.

Starting with this interview, DATA WORLD will work with Won to connect realizable with factories in Japan. Beginning with Hiroshima, a city with a strong manufacturing base, we want to create places where industrial parks and associations across Japan can try a Multi-PoC.

If you work at an industrial park, a manufacturers' association, a chamber of commerce or a local government, or if your company would like to try it at its own plant, we'd love to hear from you. Please get in touch with DATA WORLD. As Won says, choose one plant, and we can show you on screen, from typing in a product name to a finished KPI dashboard.