EP 117
Gangbuk IT Is Changing: ERP and Knowledge Work in the AI Era (Jongok Ko, COSMAX PI Head)
AI-Native ERP with COSMAX’s Head of PI Innovation 0:00
Chester Roh We’re recording on Saturday morning, October 3, 2026. Today I’m at COSMAX in Pangyo. I also run B Factory, and serve as a digital advisor to COSMAX, where I’m also working on its enterprise AI transformation to turn COSMAX into an AI-driven company. Today I’m joined by someone I work with: Jongok Ko, head of COSMAX’s PI Innovation Division.
Jongok Ko Hello. I’m Jongok Ko, head of COSMAX’s PI Innovation Division. Today I’ll share some interesting stories about AI-native ERP with you.
2–3% Paid Users and the Very Early Days of AI Adoption 0:40
Chester Roh We’ve overcome a lot of hurdles, but AI is moving so fast right now. Some say it’s moving too fast and we need to slow it down, or even stop altogether, but I think that’s absurd. Everyone around us uses AI, and it seems like everyone uses Codex and Claude Code, but if you look at adoption overall, only about 2–3% of the US population uses a paid plan. There are reportedly about 1.5 billion knowledge workers worldwide. I heard that figure on an a16z podcast. There are about 30 million coders. Those roughly 30 million coders account for most agent traffic on tools like Codex and Claude Code. Nearly all of it is coding.
Beyond coding, there are many uses, like summarizing documents, but companies and most people are still far behind, so we’re at a very early stage. There’s an analogy for this. A lot of people have been drawing parallels with history lately. Companies often say to their employees: We’ve given everyone $200 plans and even unlimited API plans, but while something seems to have changed, company productivity hasn’t risen. Some people have certainly found their work easier. They finish their tasks faster and go home. They may be doing less work for the same pay, but CEOs looking at the overall results wonder why productivity across the company isn’t improving. So people use this analogy: it’s like the shift from steam engines to electric motors. When steam engines arrived, work people had done by hand at home was brought into factories and made more efficient. Early on, one large steam engine would run, turning a main shaft through a system of pipes. Numerous belts were connected to that shaft to distribute power to the factory’s upper and lower floors. But when electric motors first arrived, factories kept that setup and simply replaced the steam engine with one large electric motor. Factory owners asked whether electricity would make things run faster and get the work done sooner. It turned out it didn’t.
So when did factory productivity really increase? When electric motors became small enough to be installed in different factory machines, each for its own purpose, and factories began redesigning their layouts completely. That’s when productivity rose. Likewise, we’re still using systems for knowledge work that we’ve built over nearly a century. Unless the entire workflow changes, the transformation won’t happen. That’s the argument people are making.
Gangbuk IT That Pangyo Doesn’t Know 4:05
Chester Roh Come to think of it, many of you listening to this podcast probably have backgrounds in IT or software. Every company I’ve worked for was a software-based tech company, and every company I’ve run was a software company. In software companies, we’re used to building the workflows we need from the backbone up. Each company has its own workflows, and often the service itself defines the company, so they’re structured in all sorts of ways. But someone told me there’s “Pangyo IT,” and there’s “Gangbuk IT,” which people in Pangyo know little about.
By Gangbuk IT, they meant the kind of work done by familiar names like IBM and Oracle and large systems integration firms. The names you hear a lot in that world are SAP, Salesforce, Oracle, and Microsoft. To return to COSMAX for a moment, that Gangbuk-style IT is firmly established here. It uses a great many solutions we’d recognize. The story goes back to last fall. Was it around the fall of 2025? At COSMAX, I was mainly involved in AI and new business initiatives, and related areas, but we needed to upgrade our ERP system. We struggled with that for a long time, until Jongok Ko joined us as division head, and we made great progress. Jongok Ko knows the whole story, so I think that would be a good place to start.
COSMAX’s Need for an ERP Upgrade and the Project’s Launch 5:20
Chester Roh SAP is an ERP system, right? ERP is meant to be the backbone of a company’s entire workflow, bringing every department’s resources into one system and letting work flow through it. And most companies use ERP, right?
Jongok Ko They all use it in one form or another.
Chester Roh But SAP usually sells its software through licenses. So, like buying Office 2023, you buy it once and keep the license for life. You sign maintenance contracts and keep the system running that way. SAP is now moving that model to the cloud. It’s changing the old licensing model to a private cloud offering called PCE, with software upgrades along the way. It’s a model we’re familiar with from Google Cloud and AWS. SAP has been changing its model to make that transition. That’s where this all started. We needed to switch to cloud pricing and make the transition, but the cost is substantial. These are huge projects requiring tens of billions of won. Companies usually call them next-generation ERP projects and undertake one roughly every ten years.
A Next-Generation ERP Project Worth Tens of Billions of Won 7:31
Chester Roh But moving to the cloud is complicated, and entering the AI era at the same time makes it even more so. Around then, COSMAX’s top executives gave me a mission. It was a daunting one: investigate how we should decide on the SAP project and propose the right direction for us in the AI era. Since I’m not an enterprise guy at all, I watched a huge number of SAP videos on YouTube last fall. I needed to learn something before I could make a judgment. I watched a ten-hour course and all sorts of other lessons, including ones on components, programs, and ABAP. And I’m being completely honest here: even after ten hours, I had no idea what they were talking about. I realized they had their own ecosystem, using a language we weren’t familiar with, and that the ecosystem was enormous.
I’m not saying that’s wrong. The enterprise ecosystem has been around for a long time. But from my perspective as someone running a business, what matters is whether this is really the system our company needs, and whether it can support everything our company wants to do as well as it should. If it can, then we just need to consider how good the results are for the money we spend. But as I looked at how things worked in practice, I found so many systems I couldn’t understand. Division head, please correct me if I’ve misunderstood something. ERP, or Enterprise Resource Planning, didn’t seem to serve as the backbone of the company’s operations. A company ultimately exists to make money, so its foundation includes finance, HR, and general affairs. And you need to run the numbers to see whether you’re making or losing money, so finance often becomes the underlying backbone. In the end, it felt as though all the other teams had to fit a finance workflow for calculating costs and issuing tax invoices.
Jongok Ko That’s right. ERP is basically designed to support FI, with everyone producing data and working to build the accounting ledger. Then, as business processes get built into it too, ERP becomes far too sprawling. And the biggest problem is that ERP projects usually happen every ten years, so changes to a company’s many business processes can’t be reflected quickly. Once every ten years, when you overhaul the whole system, you also have to update all the processes from the previous decade.
Ten-Year Reconstruction Cycles and SI Consulting Practices 11:00
Jongok Ko So it isn’t just a matter of licensing costs. Rebuilding the ERP system I was surprised that a setup requiring enormous SI consulting and development costs is simply taken for granted in this industry. Anyone who’s worked in software knows it’s hard to charge a B2C customer even $10 these days. And for those selling enterprise software, even making tens of thousands of dollars in revenue is difficult. Yet here, projects worth tens of billions of won are running all the time.
It makes me wonder what I’ve been doing all these years in software. In a way, it feels like a lucrative opportunity was here all along. I’ve been thinking about that a lot.
Chester Roh But SAP is very powerful software. For someone in B2B software engineering who considers themselves a top-tier engineer, there were plenty of things they could do. They could build a great B2C app or consumer service, or build B2B applications, what we call SaaS. They could sell those products to various customers and had plenty of ways to start their own businesses. So the ten years before the AI era began really were the era of B2B SaaS. People turned even small processes into SaaS products and sold them to enterprise customers. As recurring revenue built up and they managed customer churn, a formula emerged for valuing a company once it reached a certain scale. That formula dominated the past decade, but now people are questioning it too.
Jongok Ko These days, companies are building their own versions of Slack and Notion. Increasingly, they seem to be building just about any SaaS tool they need in-house.
A 10-Year-Old Backbone with Adult Limbs Attached to a Baby’s Torso 13:24
Chester Roh We’re not saying SAP is useless. Strictly speaking, Cosmax is also preparing a project to move to the next generation of SAP. Preparations are underway, and the project will formally begin. But before that, from our company’s perspective, we felt we needed to overhaul these systems. So we’ve brought in AI and are doing a great deal of work with it. To return to SAP, I think it would help to start from this point.
Let’s talk about what the existing systems looked like and why they’ve become a problem. First, the system is SAP. We bought it in 2016 and built the ERP system around our workflow at the time. We built modules to handle incoming customer orders, send them to the lab, and produce samples. Once the PO was confirmed, purchase orders would go out. Those orders would be linked to packaging materials, raw materials, and so on, forming what we call a BOM. Then, when it moved into production, it had to follow the production process. Once it went to the customer, we’d issue a tax invoice and handle payment. If a problem arose, the process for tracing everything forward and backward hadn’t changed in ten years, but a company doesn’t stand still for ten years. The business has expanded far beyond its workflow back then, and customer demands have grown and become more complex. Yet the backbone stayed the same for ten years. So naturally, you undertake system upgrade projects.
Jongok Ko That’s right. A very heavy ERP system tends to be a poor fit for a rapidly growing industry. As I said earlier, companies typically undertake a next-generation ERP project every ten years. It’s like having a baby’s body with an adult’s hands and feet. The hands and feet grow up enough to handle the existing work, but the body is still a baby, causing many system problems. To solve those problems, the body needed to grow too, but the ERP system has to stay running at all times. As long as the company operates, the factory can’t stop for even a day. So the challenge is to keep improving the system while it’s still running.
Chester Roh Right. I can tell you’re choosing your words carefully and trying hard to put this delicately. When I looked at it, the original system was still there, with only the bare minimum needed to generate financial statements. A separate gateway had been created for new projects, with a great many additional projects attached to it.
Jongok Ko To begin with, there are too many projects. There are too many solutions as well. But if you change the legacy system, all those solutions have to change with it, With so many vendors involved, integrating everything at once is no small task.
Chester Roh For those who work with enterprises, and for company owners, let me explain the reality. A company’s top decision-makers usually aren’t software or IT experts, so from the perspective of someone running a software company, they often don’t know things that seem obvious. For the companies brought in to do these projects, what’s already piled up is politely called “legacy,” or, to be blunt, “crap.”
Jongok Ko We call it technical debt.
Chester Roh They leave all that accumulated technical debt untouched. They set it aside and build a new data pipeline just to get their own solutions and new tools running. That’s how the systems keep evolving.
Jongok Ko That lets them finish a project in three to six months, but eventually, you have to deal with the whole thing. Yet deciding to tackle the whole thing isn’t easy, and it’s a huge burden for any one executive to take on.
Chester Roh Right, right. At a small startup, you could just strip out the existing legacy systems and start over. But this system already handles hundreds of billions, even trillions, of won. Improve that system? Nothing can go wrong while it’s running, either.
Jongok Ko That’s the biggest challenge.
Chester Roh You have to improve it and, at some point, make a clean transition. That’s extremely difficult for anyone. So when I said last fall that I’d taken on this ERP project and sought advice from ERP experts outside the company, they told me a next-generation ERP project normally takes three or four failures. That’s how hard it is.
That’s the reality, though something has changed in the meantime: AI can now do a lot of coding and similar work with just a few clicks. But there’s a gap in expectations here. Owners ask, “Can’t we just have AI do all that?” Of course it can’t do everything.
The Gap Between Management and the Front Line on AI Adoption 19:23
Chester Roh As enterprises adopt AI, they need to change too, but in what direction? It’s taken us quite a while to get to today’s topic: What are the right steps? That’s why I’ll occasionally talk with Division Head Jongok Ko about enterprise issues on AI Frontier. But both Jongok Ko and I are closely involved with sensitive matters inside the company, so even as we speak,
Jongok Ko We’re constantly self-censoring.
Chester Roh It’s always on our minds: “I can say this. I can’t say that.” Those thoughts keep running through our heads, but still, let’s talk about improvements as factually as we can.
Jongok Ko First, here’s where AI can help most directly: To undertake a next-generation ERP project, you first need what’s called an AS-IS analysis: all the company’s existing work processes need to be documented. But the faster a company grows, the less likely those processes are to be documented. Work gets passed from senior employees to junior employees, while resourceful employees solve individual problems themselves with Excel. Others leave problems unresolved or just muddle through. These practices are scattered everywhere, and from a software perspective, you can’t develop software until you’ve sorted them out.
Yet next-generation projects always start before that work is done. It’s the same worldwide. That’s why IT consultants come in. They say, “I’ll get it all organized for you,” and that’s how these projects get started. But when it’s your own company’s work, its people are the domain experts. Of course, a consultant may also be a domain expert. But beyond processes specific to the industry, each company has its own effective ways of working. It’s important for the company to document those itself. Working through that with AI is very effective.
Chester Roh This may be a sensitive question. We’ve dealt with it in our own way, but that unique system you mentioned, the one specific to each company— no one built it for them. It’s a system the company developed as it evolved and emerged as a market leader. And in today’s AI terms, this system is the company’s tacit knowledge. That knowledge builds up across the organization in different forms, and it’s highly fragmented. ERP carries essential data, such as BOMs and financial data, but it doesn’t reflect how people actually work. So, as you said earlier, much of that work lives in Excel files or documents on the most capable employees’ PCs, or in PowerPoint files. And the organization runs on that system. It runs on those Excel files, sending only the final results to ERP for processing. “Adaptation” is the right word. That’s what it is. That’s how an evolving organization has adapted to a system that hasn’t evolved with it.
The Company’s Tacit Knowledge Accumulated in the Smartest Employee’s Excel Files 22:08
In-House Claude Code Training and 100 Front-Line Employees Committing to GitHub 23:28
Chester Roh We held AI training inside the company at first, remember? Strictly speaking, it was AI coding training. Last fall was just when Claude Code was starting to take off. That fall, we invited employees to volunteer for Claude Code training. Codex wasn’t very popular then. So we held the Claude Code training. Naturally, the people most interested were often forward-thinking employees responsible for core operations. They signed up in droves. Then we trained them. One exceptionally talented person came in and taught the whole course, and their enthusiasm inspired many of the participants. What did those participants build after the training? Naturally,
Jongok Ko they began building and using the systems their own departments needed. Previously, when a team needed software for a business process, it would commission a systems integration project or buy a solution. Now, across departments, people build what they need themselves, apart from very complex systems. Among the people already doing that work, whom we call domain experts, some now commit code to GitHub at a developer’s level. There are already more than 100 of them in our organization alone. It was a remarkable start.
Chester Roh Changing a culture like this takes someone between the business domain and technology, someone with almost a CEO’s level of passion. I’ve always thought that kind of person is crucial. To me, that’s entrepreneurship. If there is even one person with ownership and a strong entrepreneurial spirit, that gene—what I think of as a cultural gene— gets copied throughout the organization.
I thought there were 30 to 40 people, but it’s already reached 100.
Jongok Ko It’s close to 100 people now.
Chester Roh So, strictly speaking, most of those 100 people come from non-engineering backgrounds—
Jongok Ko people who tried development for the first time this year, or for the first time last year. That’s roughly their level of experience.
Chester Roh Yes. Of course, that brings its own challenges, like retraining them and figuring out how to bring skilled engineers together with domain experts. Those are challenges we’re working through. But to return to the original point, those people naturally automate the company’s core work— each department’s core workflows—with AI. Then they build a system that feeds the results into ERP. Right? So what naturally happens to the many legacy subsystems we already had? What happened to those systems?
Jongok Ko We can’t get rid of them all. There’s something important here, and I like to put it this way: a company is like a living organism. If you remove an organ, something else must take over its function. Whether that takes a cyborg form or an organic form, or even a steampunk form, something has to be there. So the legacy systems haven’t changed dramatically yet.
Excel used to sit between the work and the legacy systems, and software built by the business teams themselves began to replace it. The huge advantage is that while Excel can certainly save files, much of the intermediate calculation data in those files doesn’t make it into ERP. Some of that data could be very important. As the work moved into software, the steps in the business process began to be captured as internal company data.
Here’s what’s interesting: ideally, you’d complete DX first, then use that data to move toward AX. But given what manufacturing has today, As individual innovators use AX in many areas where DX is still incomplete, the necessary DX gets completed along the way.
Manufacturing Moving from DX to AX 27:47
Chester Roh That’s a very important point. In the past, with digital transformation, people in a given domain couldn’t turn their work into a system themselves, so they had to explain everything to someone who could or outsource the work.
Jongok Ko That was the role of consultants—IT consultants.
They acted as intermediaries.
Chester Roh But the work those IT consultants used to do has been taken over by the domain experts within the company.
Jongok Ko “Domain experts” sounds impressive, but it just means existing employees now do that work themselves. It does look impressive—wow, they’re developing on GitHub and doing all these things. But the more important point is this: they can fully digitize their own work and record all of it in the ERP system. And whenever their work changes a little, they can update the system themselves as they go. I think that’s becoming the foundation for ERP to go beyond adapting and actually evolve.
Chester Roh In their work, employees—the domain experts— provide a lot of context to accomplish a task. Even if the data in that context exists only as raw text, the information gets pulled out when we feed it to AI. So naturally, the data formats for the company’s entire workflow, task definitions—what goes in and what must come out— and tacit knowledge about exceptions have been converted into digital data.
Jongok Ko As each person simply focuses on their own work,
these things are documented automatically and even mapped as graphs. Results that once came from consulting periodically—no, once every ten years— are now being produced every moment. It really is a remarkable change.
Chester Roh So, in the end—and this was possible because of AI— we ended up with a thick white paper on how the entire company operates. We’d be in big trouble if someone took it out of the company.
Jongok Ko The summary alone came to about 1,200 pages. The full version seems to be nearly 7,000 pages.
Chester Roh Of course, some parts of it are more central than others. A company is built as that knowledge passes from one department head to the next, and then to the next. So, we’ve reached that point. The individual work systems were running, so we organized them and built systems that feed into ERP. To return to SAP, how did it come to hold its current position? As I explored the enterprise ecosystem, I thought SAP was an incredibly strong company, with a great business model and a self-sustaining ecosystem. But I also thought AI must be giving SAP a serious headache right now. In enterprises, especially in finance, there’s a strong need for this: if the data didn’t come from SAP, it can’t be trusted. Part of that is trust in the brand. But what exactly does the system do that makes people say data from SAP can be trusted?
Why Accounting Firms Trust Only SAP Data 31:02
Jongok Ko From the perspective of accounting firms and auditors, it has well-established safeguards throughout. For example, if we build our own ERP, access the database, and use INSERT, anyone can change that data with a SQL statement. That could make it hard to tell whether the data was fabricated or to trust it at all, couldn’t it? With SAP, though, for example, across FI, management accounting in CO, and the SD sales module, data created through those business processes is entered according to rules and can only be changed according to rules. From an accounting standpoint, those rules seem strong enough to make the data trustworthy. For accounting firms, that applies to SAP data and Douzone data alike. They trust Douzone’s data, too. If they don’t trust other data, you have to prove it’s trustworthy. In practice, companies have little choice but to use a well-known ERP system.
Chester Roh It’s software that has served as an industry standard for some 40 to 50 years, and you can’t discount the strength of its brand. From a software perspective, an accounting ledger uses double-entry bookkeeping: entries in different places ultimately balance, allowing the records to validate themselves. That’s a defining feature of double-entry bookkeeping. SAP has what you might call checksum checks built into each module,
Jongok Ko across many areas. For example, after the month-end close, you can’t suddenly change a cost. If you do need to change one, you can’t do it directly. You reverse the original entries through a specific process. It’s like a database transaction: if someone opens several apps to withdraw money, you have to prevent simultaneous withdrawals by grouping the actions into a transaction. Many business processes are linked as transactions in the same way. Because the system has been around so long, it has earned a great deal of trust in accounting.
Chester Roh So if a ledger passes those audit checks, unless someone has forced the numbers to fit, an error would be caught by a check along the way, making errors unlikely in the final result. That’s the standing the system has.
But from what I’ve seen of SAP on YouTube and in its manuals, FI/CO, the financial and management accounting modules, is the most important system for a company. But some companies manufacture, others distribute, and others do something else. There are countless business processes. To calculate the figures in that ledger, those modules have to feed it the right data.
Jongok Ko That’s right.
Chester Roh From a software perspective, you might say you could just set up a PostgreSQL database and build all of that yourself. But as we said earlier, Gangbuk IT lacks the engineers who can do that.
Jongok Ko There aren’t any, and they’re far too expensive. You’d need a top-tier developer on hand all the time.
Chester Roh So SAP packages it into modules: “What industry are you in? For that industry, buy this module and this one, configure them like this, and you’re set.” For anything those modules can’t handle, you add a custom-built system using ABAP—
Jongok Ko People usually call it ABAP.
Chester Roh You build a system in that language and add it through customization, correct?
Jongok Ko Basically, with SAP, you have to write ABAP code. And using that ABAP code, people used to build all kinds of business processes. So to change a business process, you need a developer who knows ABAP. I studied computer science and have worked in software my whole life, but I’ve never written ABAP code. That’s how rare ABAP developers are. And ABAPers, meaning ABAP developers, often haven’t worked with other languages. As a result, they seem to occupy a very unusual position in the market.
Chester Roh It isn’t a conventional software engineering role. Only people who’ve built careers as engineers or consultants within the SAP ecosystem could do it. But I have another question. FI/CO is certainly a powerful system, but there are all those individual modules that run on top of it. Even at Cosmax, we have sales and marketing, then R&D, and production is a huge part of the business.
The Cost and Limitations of SAP Module Customization 36:56
Jongok Ko Production is the biggest part.
Chester Roh Can you adopt those systems and make them work? They can’t fit every company’s work perfectly, but can enough custom development make them work?
Jongok Ko With ABAP? Yes, you can. But every time a process changes, you have to maintain and modify them.
Chester Roh Maintaining those projects is very slow, too.
Jongok Ko It takes a long time and costs a lot.
Chester Roh That’s also how SAP built its ecosystem.
Jongok Ko I doubt SAP intended all of that, but the developers seem to have very different ways of working, too.
Normally, if developer A is working on something and the frontend developer is out today, someone else can readily step in. They can check GitHub and pick up the work. Oddly, though, in my experience with ABAP, if someone is responsible for FI and that person isn’t there, the work stops. And if someone is responsible for CO, the CO developer doesn’t know FI, and the FI developer doesn’t know CO. And CO developers know nothing about production-related PP, either. There are strong boundaries between them, too. They don’t cross into each other’s areas or even try to. Even job postings ask for someone who knows ABAP, or someone to handle the FI module.
Chester Roh So the ecosystem evolved to work that way as a matter of course.
Jongok Ko Yes. It’s structured so people only have to handle their own work.
Chester Roh Do the modules built into each area still serve their purpose?
Jongok Ko As I mentioned, in a fast-growing company, work keeps being added and changed, and the legacy system can’t keep up. SI work acts like glue to fill the gaps, so things don’t run smoothly, but they somehow keep running.
Chester Roh I see. We were talking about the SAP ecosystem earlier. It’s worth studying from a business perspective, too.
Jongok Ko It seems like a fascinating business case.
Chester Roh I’ve met many people who work with SAP in Korea, and people from SAP headquarters at conferences in the US. Talking with them, I get a clear sense that the ecosystems are entirely separate. Even the terms they use and the terms we use reflect very different basic philosophies. They’re more focused on business than software engineering— much more on the humanities and business side. When we talk, it’s about 20% humanities and business and 80% engineering. With them, it’s reversed: 80% humanities and business, 20% engineering. It’s quite different. I thought it would be interesting to talk about what makes the SAP ecosystem distinctive.
The SAP Consultant Ecosystem and the Engineer Shortage 40:32
Chester Roh As I said earlier, there’s Gangbuk IT and Pangyo IT. But before AI started coding and completely changed the fate of software engineers, software engineers were always in short supply. Skilled software engineers tended to go to software startups, big tech, or Korea’s leading internet and platform companies. In Korea, they went to those places. So for traditional companies— whether Cosmax or another large manufacturer— if we could have hired highly skilled software engineers and product managers to handle our work, we would have built the systems ourselves. But we couldn’t find those people. I think that was the biggest problem. We simply couldn’t find them.
And large B2B companies filled that gap. Companies like Salesforce, SAP, and IBM came in and filled that entire gap. Then companies like SAP and Salesforce built their own vast software ecosystems and established a group of engineers who worked only with their proprietary code, strengthening roles like ABAPer and SAP consultant. This ecosystem is enormous.
Jongok Ko Yes, yes. It is.
Chester Roh After that, only those people could do this work, and once companies implemented SAP, they had to hire people to run the system internally. But those people also came from the SAP ecosystem: existing consultants or people who had managed SAP at other companies. I also found it interesting how closely this community’s talent pool was connected to the people working inside companies. Why does it work that way?
Jongok Ko It’s like engineering students being closer to each other and having more to talk about. Humanities students tend to stick together too. In practice, if you look at the market, SAP developers and ABAPers have their own communities. So they often don’t talk much with conventional software developers. Their work is different, too. And without much contact, they have few chances to become close. But ABAPers and SAP staff at different companies often seem closer to one another. I don’t think there’s any particular purpose behind it. It may be because they do similar work, and that ultimately affects their career paths and their ability to change jobs.
Chester Roh SAP didn’t build this ecosystem overnight. It built it over several decades. SAP positioned itself as the provider of the solution, and it never sells directly to individual companies.
It always sells through a third-party consulting firm. It’s a structure that guarantees
those firms very high profit margins,
and those third-party firms have plenty of engineers, including ABAP specialists. SAP issues certifications, you know. Those firms hire people who hold those certifications. That let SAP maintain a strong brand and set the rules for the ecosystem: “We’re going to charge a lot for this, you have to do at least this much, even competition has to work this way, and we’ll train those people.” That made it an extremely powerful business with no real alternative and a strong moat. It’s operated that way for 40 to 50 years.
AI Shaking SAP’s Moat and Bringing Manufacturing Talent In 44:48
Chester Roh What’s changing this in the AI era isn’t me. It’s the arrival of AI. Companies have frozen software engineer hiring across the board. That’s simply a fact. They aren’t hiring. Instead, they’re seeing how much more they can get out of the people they already have. New hiring has stalled, and compared with the past, even senior software engineers are now entering the job market. So people working in IT in Gangbuk can now hire people they previously couldn’t recruit.
Jongok Ko We also have an AX innovation team under our PI innovation division, and we’re preparing to hire more than ten people. Until now, software developers with strong résumés generally haven’t applied to manufacturers. They usually preferred software companies. They used to go to search engine companies, and for a while, to well-funded software startups. But as we reviewed nearly 100 résumés, I felt the applicants’ qualifications had really improved. People we previously couldn’t attract to manufacturing are now applying in large numbers. So I’ve been thinking that AI may ultimately strengthen manufacturers’ software capabilities in an unexpected way.
Chester Roh As we’ve been talking about enterprise IT, we’ve also done a great deal ourselves. Systems we couldn’t build before are now up and running. And we’ve seen how much damage a vibe coder without sufficient software engineering knowledge can cause. In our experience, that includes wiping out a database.
Jongok Ko That does happen sometimes.
Chester Roh Even so, the company as a whole is doing things it couldn’t do before. So we’re designing safeguards against those risks while continuing to introduce systems that let people do more of the things they want to try. We’ve talked a bit about the SAP ecosystem, but even so, we’re also going to move to the next version of SAP.
Jongok Ko We have many legal entities, including some small ones. Above a certain revenue threshold, they use SAP; below it, Douzone. In some cases, they use another ERP system. We have three different ERP systems within our company alone. So ultimately, I think we’ll have to keep using multiple ERP systems. But most importantly, assuming the company keeps growing rapidly, when it comes to business processes, factory production, and sales, can that heavyweight ERP keep covering all of it? I think that’s something we need to consider.
Chester Roh To take that discussion a little further, we keep asking a fundamental question: what is the ideal shape of an organization that operates with AI? And we’re trying various experiments to find out. The foundation of those experiments is people who know their domain well, what we’d call domain experts: people who were already outstanding employees at our company. The starting point is the sharp rise in productivity when those people combine their expertise with AI.
From there, we talked earlier about Excel and things like that. And with the workflows those people have created, as systems gradually shift to AI, let’s call the original team the human team. Think of the existing workflow using my earlier analogy: a system built around a steam engine and one fitted with an electric motor. Sometimes those two systems coexist; sometimes they complement each other, and sometimes they clash. Or one may be judged better or worse than the other. Sometimes that drives the choice, and from a broader perspective, senior management may end up considering both systems. Situations like that arise.
Jongok Ko All the time.
Chester Roh Could you share some interesting stories about specific cases?
7 Billion Won in Unused Solutions and a 70% Reduction in Maintenance Costs 49:50
Jongok Ko Management’s main question is: We’ve adopted AI, so what exactly has become easier? That’s all qualitative. So it’s hard to answer. But in quantitative terms, after seven or eight months of hard work, we calculated which solutions we’d otherwise need to adopt and found we could avoid buying about 7 billion won worth. That’s software spending that would have counted as new investment. We’ve already avoided roughly 7 billion won in development investment over the next two or three years. We also have about 2 billion won in annual maintenance costs, which we’ll cut by over 70% starting next year. We’ve already cut maintenance costs by 1 billion won this year alone. So the financial results are already showing.
Chester Roh As we discuss this, it’s good for the company, but greater efficiency means less of that money keeps circulating. Across the economy, AI clearly has a deflationary effect.
Jongok Ko That’s certainly one aspect, but I’d look at it this way. In the past, large companies bought many software solutions, and that spending helped software companies grow. Now, software professionals are joining those large companies directly. Less money flows out to suppliers, and ultimately each company may become more competitive at its core. At the national level, that could actually make us more competitive globally. Since our company focuses on K-beauty, we’re moving toward a model where AI-driven software capabilities directly strengthen K-beauty’s competitiveness.
Chester Roh That’s a nice way to put it. Fundamentally, the people who use AI well first
Jongok Ko are the ones who benefit.
Chester Roh Either way, it starts with greater efficiency. Going further, when someone figures out how to tackle the fundamentals, productivity itself rises. That means less money flows to surrounding businesses. Across the economy, less money circulates, but stronger companies bring other gains. Of course, that raises questions about concentration and distribution, among other things. Aside from handling work internally that we used to outsource, are there parts of the existing workflow that AI has fundamentally changed or greatly improved?
Why the SSOT Exists in Excel Outside the ERP 52:53
Jongok Ko An ERP system like SAP or Douzone is supposed to be a company’s SSOT, its single source of truth. How many units do we have in stock right now? What do we need to make a particular product? That information should be in the ERP. But often, it’s outside the ERP.
Chester Roh Why is that?
Jongok Ko I didn’t understand it at first either. I thought, “Then what’s the ERP for?” But that’s how it is. For one thing, the company has grown so rapidly that creating a single source of truth requires joining and combining an enormous number of tables to produce the information.
Our material master data alone has hundreds of thousands, even millions, of records, all linked to one another. Calculating which materials go into each product, and how much, is difficult to do in real time.
Chester Roh It can’t be done. So that information also lives in each department’s Excel files?
Jongok Ko Say some materials arrive today, and today’s production data already exists in the respective tables as a source of truth. But it hasn’t reached the master data yet; sometimes it’s updated later, at regular intervals. That’s one example.
Chester Roh Why the delay? Because people are busy?
Jongok Ko There are two reasons. First, someone has to automate tasks like entering production results. Also, running that process every time would put too much load on the ERP. That’s another reason. The volume of data is so large.
Chester Roh So when you enter it, it just
takes several minutes to process, sometimes an hour. There’s data like that?
Jongok Ko Not one hour. It runs for seven or eight hours.
Chester Roh From our perspective, that means the program was poorly written.
Jongok Ko I think that’s part of it too. But I think the basic issue is this: Not every ERP system is designed to fit every industry. Each ERP system has to be designed for its industry. At Cosmax, though, we have tens of thousands of finished products. There isn’t an ERP system built for that. In terms of ERP scale, for example, a major automaker’s system is much larger. It has far more programs, or T-codes, than ours. But how many types of finished products does it handle? A dozen or so. Models get discontinued, and there are far more parts. At Cosmax, it’s the finished products that come in so many varieties. So it’s more likely the ERP wasn’t designed for our needs. There’s a lot of room for improvement there too.
Chester Roh It sounds obvious, but if you try to cobble that together with ABAP…
Jongok Ko You really can’t build it that way.
That structure is fundamentally tied to SAP’s architecture.
Chester Roh Hence the saying: rather than suffer, do it yourself. People have their own databases that are much easier to use, and for those who don’t work in software or IT, that’s naturally Excel.
Jongok Ko That’s unavoidable. Whether they confirm by phone or some other way that something has arrived, they find out a few more units have come in, so production can start right away, and they can tell the customer, “We can produce it today and deliver it in two or three days.” They have to have these exchanges constantly, even now. That’s how important real-time information is in this industry. With electronics or cars, for example, the number of products is fairly limited. If there are fewer than 100 types, there are only 100 types of inventory to manage. If you expect to ship 1,000 units this month, you can prepare 1,500 without having to stock huge quantities across countless products. But if we have 10,000 or 100,000 types of finished products, we can’t prepare them all in advance. So most of it is handled just in time. That makes real-time information even more important, but most large ERP systems may struggle to reflect changes in real time. That’s another aspect of it. This is just one example.
Chester Roh So the work that needs doing ends up in the business teams’ Excel files. They talk to each other by phone or messenger to reconcile their respective databases, then enter the results into the ERP afterward to keep a record. So that’s how the system works.
Jongok Ko That’s right. Only then can they deliver the goods, issue a tax invoice, and recognize revenue, allowing the company to get paid.
AI Eliminating Interdepartmental Communication Costs 58:03
Chester Roh Then, naturally, as AI enters systems that used to run on Excel, what changes are taking place?
Jongok Ko A great many changes are already taking place. For example, someone in a department might get their daily work through Teams messages—as many as 300 a day. They have to keep track of every one. But work like that is disappearing.
Chester Roh Gone entirely?
Jongok Ko Yes, some tasks have been eliminated altogether. I haven’t said which department, because several departments are eliminating this kind of work.
Chester Roh All of that is the cost of communication. People spent all day on the phone doing their jobs. You mentioned SSOT earlier. One synchronized database that everyone can see at once would solve it. They should be able to check it and settle the matter, but an enormous number of things didn’t work that way.
Jongok Ko To people who work only in software, this sounds obvious. But when I looked across the industry, I found that most can’t do it. It turns out ERP normally works alongside Excel. That came as quite a shock to me.
Chester Roh Most people act rationally. And wherever there’s an opportunity, someone will offer a solution. Because of AI, the existing enterprise B2B ecosystem is bound to undergo enormous upheaval.
Open-Source ERPNext and the Gap of Korean Tax Law 59:43
Jongok Ko That’s already happening. Here’s a hint for viewers: there’s an open-source project called ERPNext. It’s an open-source project in the FI and CO space, and it’s developing very quickly. It’s open source, and everyone now has AI as a tool. So it advances very quickly. We tested it. We wanted to see how reliably it could handle the safeguards across the many business processes we analyze when assessing the current state. It wasn’t perfect, but about 70 to 80% of those safeguards worked. That really surprised me.
In the end, there are just two reasons ERPNext can’t be used in Korea: it hasn’t been adapted to Korean tax law, and it lacks Korean firm banking integration. An ERP normally connects to banks and automates even incoming payments. Those two features just haven’t been implemented in Korea yet. I’m sure some startup will spot the opportunity and use ERPNext. If it adapts it to Korean tax law and adds firm banking, it could quickly offer a lightweight ERP as a SaaS product.
And as you know, software is advancing incredibly fast right now, so I think it could eventually grow into a project that seriously competes with SAP and Douzone.
Chester Roh I can see that happening. In a way, ERP used to say, “We’ll take care of everything” within a certain scope of work. Because nobody else would do it, traditional ERP projects did more than provide a basic framework. They also pieced together business process applications in a makeshift way. If most companies’ actual work still runs on each team’s Excel sheets, AI-equipped domain experts will move in from above, and existing ERP providers will retreat to their core functions unless they develop a decisive advantage there. I think a new balance will emerge with existing ERP providers pushed much further back.
AI really is making life hard for a lot of people.
Jongok Ko It certainly feels like a time of change. Every field is advancing and changing so quickly that I think it’s important to keep up with the trend.
Chester Roh Enterprise AX is a huge market, and many small FDE firms are springing up these days. But software engineers working as FDEs can’t walk in and handle this work right away.
Jongok Ko It’s a field that’s been hidden away.
Chester Roh There’s so much tacit domain knowledge and so much that applies only within enterprises. But domain experts are learning to use AI, and experienced software engineers will come after that domain knowledge too. Just as leading AI researchers in Silicon Valley now read molecular biology books at night and tweet about them, I think people will naturally enter this field if there’s a market.
Jongok Ko Startups have tried many fresh, impressive ideas, but I think there’s a clear reason they haven’t tackled ERP. Startups don’t use ERP themselves. They use Notion and things like Google Storage, and that’s all they need.
They can pay an accounting firm a few hundred thousand won a month to handle all their accounting. And in special cases in Korea, they use software like Douzone, with an accountant simply recording ledger entries. So I think people working in IT have had no chance to encounter this area at all.
If they’ve never worked in it, they can’t identify the demand in that market. But there was a huge, hidden IT market that had remained protected. And because of AI, I think it’s going to face a serious challenge.
The Owner’s Understanding and What Remains Human in the AI Era 64:25
Chester Roh On the other hand, because of the work I do, I meet many company owners. I think they’re now beginning to open their minds to this too. Before, they might have said, “I don’t understand it; just hand it to a vendor.” But an owner needs to understand exactly what this involves, what talent is needed for the transition, and what risks they’ll have to bear. Only with that clear understanding can the project get started and become viable.
Many people start without really understanding these things, then face strong pushback from existing staff. Or if the results don’t seem to meet their expectations, they simply give up. They go back to the old system or look for a new vendor. That was the pattern. One thing I’m grateful to Cosmax’s top management for is They had strong faith in the direction Chester Roh envisioned, and in the direction our division head was taking, which made it possible for us to push ahead. So those things matter in a way entirely different from AI.
And the deeper we get into the AI era, the more we ask: What, exactly, is left for humans? Discernment, taste, the will to act, and trust between people. “I’ll go with this person, or die trying.” Those are the forces that change the world. Trying to be clever and just clicking away with AI won’t make something big happen. That’s what I wanted to say.
The Future of Corporate Organizations Reshaped by AI 66:16
Chester Roh Now, about enterprise AI transformation, as division head, you know our company is very much a manufacturer, and yet we’ve moved quickly with pioneering experiments and changed a great deal. You’ve been leading that work on the front line. With AI transformation, what do you think the company’s future will look like? Right now, we have separate departments, each led by a department head. Staff do the hands-on work, calling and emailing one another and earning revenue. That revenue pays salaries and bonuses. That system keeps running, and the nature of the business won’t change overnight. But the organizations serving that market will change a great deal because of AI. What do you think they’ll look like? This may be a harsh reality, or perhaps just speculation.
Jongok Ko It’s hard to imagine what every area will look like, but if we think about it from a sales perspective, Cosmax has a very large sales team. It’s a large team, and the sales staff in domestic marketing have an enormous amount to do. They respond to customers and check on their requests. But they also connect people with the right teams. For a quality issue, they talk to the quality department. For a production issue, they talk to production. Then production might say it can’t proceed because of a particular material, so they contact the person responsible for it. If all the ingredients are there but the cap has a problem, they contact the packaging materials team. That supplier is an outside company, too. They contact it, adjust the schedule, and arrange delivery. Even if the lotion inside no longer feels as soft as it did at first, they have to talk to the research lab. There’s so much to do.
Chester Roh So domestic marketing connects all the back-end teams—
One person connects all of them.
Jongok Ko The systems handle some of it, of course, but when it’s urgent and the system hasn’t been updated, they handle it all over the phone. It’s very difficult. Revenue is growing rapidly, and so is the number of customers, but they have to manage it with limited staff growth. So the complexity keeps increasing. In the long run, I think as much of this operational work as possible will be replaced by AI.
Then domestic marketing staff can focus more on customers and finding new ones. AI can’t do that. It’s not a matter of AI sending cold emails. Ultimately, it’s about customers and their needs. At Cosmax, our top priority is finding ways to solve our customers’ problems. If a customer wants to launch a new product, we can work through the challenges together. Staff can spend more time on that and improving customer satisfaction. All those routine tasks can be checked once the data and processes are defined. Those tasks will eventually be automated, allowing staff to focus more on customers. I think that’s where we may be headed. That’s one possible future I see for domestic marketing. Each department will probably evolve in its own way.
Chester Roh We can break these into modules: this one changes one way, and that one changes another. We’ll look at those examples one by one next time. You put it somewhat indirectly, but the same point comes up here. Now, the so-called grunt work people do We should hand the work we were talking about over to AI as quickly as possible.
Jongok Ko Repetitive, time-consuming work that isn’t core to the job. Though some of that may be core work. But that will all be automated, too.
Chester Roh The higher-level work people need to do includes caring for customers, and I tell our team the same thing: “I’ll take away the busywork, so bring me insights.” We need to move toward higher-value work that produces insights. It’s an incredibly simple but clear conclusion.
How long will that change take? We’ve been leading AI transformation, mainly around Cosmax’s SAP system, for nearly a year now, and you’ve been here as division head for about eight or nine months. Yet even in that short time, we’ve changed a great deal. And the pace of change is accelerating because AI is getting better, rather than because of anything we’re doing.
Jongok Ko That’s right.
Chester Roh It isn’t that we’re so good at this. AI just keeps getting better at the work.
Jongok Ko Exactly. It’s getting cheaper, too.
Chester Roh So we really need to understand this. As a result, the pace of our own change keeps accelerating. When do you think it will be complete? How much longer will it take for AI to change the whole organization?
Jongok Ko Even with a very conservative estimate, I think it will change completely within two years.
And in some areas, over the next six months, I think we’ll see major parts change dramatically every six months. An organization transformed by AI won’t turn its people into robots all at once. It’ll be like cyberpunk: first the hands are enhanced, then the eyes, and then the legs. Sales will suddenly change, then the factory, then the research lab, until eventually everything has changed.
AX Talent That Eliminates Their Own Work First 72:36
Chester Roh All right, let’s close with a lesson from this. If another department comes in and transforms my work with AI, the future looks bleak. If they transform our organization with AX, a lot of what used to be part of my job could disappear.
Jongok Ko If you limit your role to knowing your routine tasks well and handling them well, your position could be at serious risk.
Instead, you need to use the domain knowledge you’ve gained from that work to keep asking how to improve things, and how several people could accomplish more than one person can in roughly the same working hours. That’s how you need to grow. People who think AI cut an eight-hour task to five hours, so they can take the other three hours off, are very likely to be replaced by someone else.
Chester Roh That has happened at our company, too. In fact, the people at Cosmax who moved from business teams into AX were the ones who said they would eliminate their own tasks first and then do more.
Jongok Ko This goes back to what we discussed earlier, but I keep hearing that established software solution and SI firms are rapidly facing serious difficulties, and the reason is similar. They used to assign large numbers of software engineers to SI projects and solution development. Everyone knows that AI now lets them do that work much more efficiently. If they had charged correspondingly less, I think everyone would have kept using them. But they tried to keep all the efficiency gains from AI as higher profit margins, so the cost gap compared with building it in-house grew too large.
Suppose we used to charge 15 million won for an SI project, but now we can finish in one month what used to take six. If we said we’d charge only one-sixth as much, I think everyone would have continued working with us. But prices haven’t come down. Customers know these firms can build things much more easily, and they know building them in-house no longer takes as long as it used to. So I think more companies are bringing that work in-house.
Chester Roh And the companies commissioning the work apparently say, “You’re going to have AI do this anyway, so give us a bigger discount.”
Jongok Ko These days, that’s only fair.
Chester Roh These days, delivery terms often require all the source code, and customers know that once they have it, they can make changes in-house with just a few clicks. When a new tool changes what’s possible, the world finds a new optimum.
It seems we need to adapt quickly.
We’ve talked about AX at the company level, but for knowledge workers, the lesson is to quickly use AI to disrupt my own work and realize, “I shouldn’t be doing this anymore. Then what should I do?” It comes down to who asks that deeply human question sooner, pursues it more diligently, and answers it better.
Jongok Ko Exactly. Doubling the work you used to do in roughly the same hours to increase output per person is still a fairly modest form of progress. And as you just said, when you disrupt your own work, you become capable of disrupting other work, too. That becomes your role.
Chester Roh And when you look at the world through that lens, you find many more things you can do. Then, instead of being a cog in the organization, you define yourself as someone who sees the world in a new way. I think we’ll keep hearing stories like that. So let’s close with this: it’s a question of who realizes it first. I tried something new today. With support from Jongok Ko and the leadership at COSMAX, I’m learning a great deal from the work I’m doing. Once I have a more balanced perspective, I’ll get into specific tasks next time: what happened and how best to approach them. I’ll talk about those. These are examples of best practices that will soon be widely known even if we don’t talk about them. We believe enterprise AX in Korea needs to improve much faster, so I think it would be good to gradually share what we learned through our own trial and error. All right, let’s wrap up today’s October 3 recording here. Thank you. Thank you.