Teamsparta × LilysAI: Measuring AX and Building for a Global Market

This Bloom event featured Jaekyung Hwang from Teamsparta and Yein Kim from LilysAI.

Teamsparta is a Korean tech education company whose AX (AI transformation) team trains enterprise clients across many industries. LilysAI has been building AI summarization since the early days and is now building Readray for English-speaking markets. I learned a lot from what both teams shared, and I'm grateful to them.

Two stories stuck with me. The first was Teamsparta's AX training team putting on safety gear and walking into a steel mill, right up to the blast furnace, to pull raw data off machines with a USB stick so they could help the company analyze it and bring AI into its work. The second was Yein's. After a year of trial and error trying to take LilysAI global, they decided to build a new product designed from day one to work in the US and other English-speaking markets (so much so that they won't even open a pilot to Korean companies), and they've kept pushing in that direction. Both came from people doing the work on the ground, which is why they hit harder.

Thanks again to both of them. Here's a detailed account of the evening.

On Tuesday evening, September 15, we gathered at Teamsparta's office on the third floor of Anam Tower on Teheran-ro (Seoul's main startup and tech strip) to ask where our companies' AX really stands. The starting point was simple: everyone says they're good at AI, but nobody has a standard for what "good" means. Every desk had a copy of a report Teamsparta built by surveying 330 people who make training decisions at their companies. On stage we had the co-founder of LilysAI, which has 1.3 million users, and the consultant who wrote that report. The last hour was a roundtable, six people per table, where everyone put on the table the place where their own company is stuck.

Nobody had a standard for being good at AI

Junsu opened the evening. AI is hot right now and everyone says they're good at it, he said, but nobody seems to have an actual standard for what that means. The theme of the night was to build that standard together. When he asked who was in charge of rolling out AX at their company, quite a few hands went up. When he asked who was in HR and had already run training only to watch the business teams fail to use it, almost no one raised a hand (Junsu said he'd assume they were just shy). Close to half the room was at Bloom for the first time. We followed our usual format: an opening fireside, a main fireside, then the roundtable.

A great knife came out, so now people are cutting hair with it

I hosted the opening fireside. I'd been thinking about who could make AX feel concrete, and LilysAI came to mind because it went viral from the moment it launched. I'd wanted to meet Yein for a while, so I invited them. When I asked who in the room knew LilysAI, a lot of hands went up.

LilysAI helps people who read a lot of material understand it faster. You throw in YouTube videos, PDF papers, web articles, whatever, and it summarizes them and answers your questions. It's been out for three years. I mentioned the 1 million users I'd seen in a recent article, and Yein corrected me: 1.3 million. When I asked how it felt to first see that number, they laughed before answering.

I was happy. But a number like 1.3 million users is mostly something you use when you announce things, internally or externally. Our internal target metrics are about retention and paid users, so I just felt good about it and moved on.

The health metrics they watch internally are retention first, then paid conversion and paid retention. Signups are a number for the outside world, so there was no big celebration. Yein couldn't remember retention ever crashing. Instead they talked about the new product they just launched, Readray, aimed at North America. The share of users who come back a week after trying it is 70%, much higher than LilysAI.

Yein's explanation was usability, not intelligence. A lot of products today are so dazzled by how smart AI is that they overlook accessibility and usability.

A really, really good knife just came out, and because it's so good, people are using it to cut their hair and cut fruit. But honestly, cutting your hair is a much better experience with scissors.

Yein thinks AI's penetration into everyday life is still about 1%. Eventually there will be AI in your chair and your desk, and we're only at that 1% now. Narrow it down to the problem of reading, and the best experience is information augmented right in front of your eyes, like Iron Man's HUD or the glasses in Kingsman. So Readray follows you as you read and layers on research without you asking for anything. The bet was that if everything you need for the paper you have to read right now is laid out in front of you, that's a temptation you can't resist. The 70% retention is the result.

Yein uses Claude as their main tool, with Gemini on their phone and in Chrome, for exactly one reason: you can summon it instantly. Pasting material into Claude or ChatGPT and saying do this, do that, do it again is cutting hair with a knife. At least for reading, augmentation right where you are beats chat. Accessibility and proactive suggestions are still overlooked because everyone is distracted by how smart the models are, and Yein sees a lot of the future there.

Danggeun does 270 billion won. Grammarly does 1 trillion. Canva does 4 trillion.

Only a few people in the audience knew Readray. I started by asking why they'd build a new product when the existing one was doing well. If LilysAI is a wide workbench where you come into the service, upload material, summarize it, transform it, even make quizzes from it, Readray holds onto just one narrow flow: cramming the material in front of you right now into your head, one way or another. When I asked why they narrowed it down, the answer was about a dream, not the company.

My co-founder Hyunsu and I dream of building a product that's loved globally. We started the company because we want to make software that's truly used around the world, like Notion or Canva or Grammarly.

Here's the comparison Yein gave. Danggeun (Karrot, a Korean neighborhood marketplace app) is called Korea's most successful startup, and its annual revenue is 270 billion won. Grammarly makes 1 trillion won from grammar correction alone, and Canva makes more than 4 trillion. It took them a year to realize they'd have to get far more focused to build a product like that. They fought hard to take LilysAI global. There was a little traction in Japan, a little in Northeast Asian markets like Taiwan, but nothing at all in English-speaking markets. In Korea, everything from launch to growth went fairly smoothly. Globally, the same actions played out differently at every step, so it was like learning to walk again. Yein called that year "digging holes for nothing," and it ended with the conclusion that they needed to build Readray from scratch.

In Korea, a button without a label is scary

I wanted to know whether failing to break into the US was a product problem or a marketing problem. The answer was a mix, but because GTM and product are connected in the end, they decided they had to change the product. The differences Yein described were fascinating. North American UI is far more extreme in its simplicity. In Korea, people only press a button if it has a label. In North America, if every button has a label, people complain there's too much text. Korean users, on the other hand, won't press an unlabeled button because they're scared of what might happen. "Put your material in and you can do all kinds of things" doesn't sell there. You have to be sharp about it: "I'll augment the paper you're reading right now, right where you are, like a HUD."

I heard this as the difference between super apps and niche apps. Asia wants many features in one app, and the US wants something narrow and deep. That night I learned that the so-called Asian sensibility is about UX, not visual design. That's why they concluded it was better to simply build a new product for the US market than to keep agonizing over how to do GTM better.

When I read the phrase "AI lens" in an article, I pictured hardware, like a contact lens. It's actually a Chrome extension, with a desktop app coming soon. You open a paper, it detects it, and a button pops up asking whether you want it analyzed. It has launched, but as a closed beta, and they block Korean users and only accept North American users. I asked whether that many people really read papers.

From a Korean point of view it looks like a small market of PhDs and research labs. Yein said that Jenni AI, which targets only researchers, makes over 20 to 30 billion won a year as far as they know, and that just as Grammarly built a trillion-won business on grammar correction, this is a big enough market. Readray's customers also aren't only researchers. There are three groups: researchers, investors, and company leaders. All three have to read dozens of papers, reports, and videos a day and make good decisions. Geographically they don't see only the US. Canada, Australia, the UK, and Europe make up one English-speaking set.

YouTube added a summary button, and retention didn't move

I was curious whether each new model release, like GPT 5 or Opus 4.5, was a threat or a lever. Yein said the service hasn't had a crisis moment yet. The last year was spent focusing entirely on landing globally, and growth continued the whole time.

People worry about us a lot. "YouTube can summarize now, there's a summary button, what are you going to do?" But when that kind of thing happened, there hasn't been a single moment when our retention or churn changed.

The reason came down to who the customer is. LilysAI's customers are people who watch a huge amount of investing YouTube to make investment decisions. A five-line summary will never be enough for them to decide anything. What they need is to absorb the material completely, just faster, and Yein feels there still aren't many products that solve that. LilysAI's summaries are built in several layers to give users confidence that nothing important is missed while cutting the time to a tenth. Products that go "ta-da, look what this can do" get early signups, but the users all leave. Yein believes people eventually recognize a product that actually solves the real problem all the way through.

In Yein's view, one general-purpose AI can't handle details like that. They added that even ChatGPT and Claude hear "it's over for them" every time a competitor ships a new feature, so in a market changing this insanely fast, keeping up your product development speed is what matters. So the thing they wrestled with longest wasn't the model. It was global marketing.

On Reddit, the first response is "who are you?"

I wanted to hear the whole story of bringing over what worked in Korea and watching it fail. Here's how they first grew in Korea. They joined a bunch of open KakaoTalk chat rooms (KakaoTalk is Korea's dominant messaging app) and Discord servers like Bloom, made really good notes with LilysAI summarizing material each community would love, and shared them around. Because the material was good, people wondered how it was made and came in on their own, and influencers inside those communities started recommending it voluntarily. It grew by itself.

We tried to reproduce that in North America, but on Reddit the filtering is extremely strict, and whatever you do, people ask, "Isn't this an ad?" Korea is small, so people help each other out and look at things with an open mind. In North America the "wait, who are you?" reaction is really strong, so it didn't work well.

American communities are much more closed off than they expected. They suspect ads or phishing first and never click unfamiliar links. When I asked what they'd do differently if they could go back, Yein said there was nothing they really regretted. What's working now is social marketing with micro-influencers and DMs on X. It's working so well that they're starting to wonder if this is actually the answer.

Short-form video is the biggest piece. They don't make it themselves. They recruit influencers who are a good fit, like influencers doing PhDs or lifestyle influencers, to make it, and it runs on people who are happy with the product spreading it to their friends. The product's wow moment has to show up in the short-form video, so this is where product and marketing connect. Readray is "Read" plus "Ray," as in a beam of light, and there's a moment where a beam sweeps across the PDF or YouTube screen and scans it. They put that moment in the videos to get a "wait, what is that?" reaction. Listening to this, I thought English speakers would get "ray" right away, but non-native speakers might not remember it. Their point that they built it for North America from the start came through even in the name.

Not raising money made them good at managing costs

AI costs drop and then climb again, so I asked whether that was a burden. The answer started with funding.

We haven't raised any investment so far. We started the company thinking we'd raise if there was a reason to. We passed breakeven within a year and have kept making a net profit, so we didn't raise. Because of that, I think we managed costs really well from the start.

Three years in, they're still growing and profitable. Since every expense comes out of their own pockets, they grew up checking every month how much was going out and whether anything was leaking on useless things. Model costs were much higher at first, but they strongly believed prices would fall fast, so they assumed that and added features aggressively. In reality, prices fell even faster than they expected.

The other piece is willingness to pay. AI goes beyond assisting people. It gets to the point of replacing the work people used to do. So we've entered an era where someone who paid 10,000 won for a SaaS product now pays 100,000, 200,000, or more. As long as the value to the user is there, Yein doesn't think the cost structure is something to worry about. Recently they connected the company's spending records and API usage to AI and had it review every unnecessary expense. Within a few days, they cut about 30 million won a month. Yein's view is that no AI product fails to survive because of cost. If it didn't work, it didn't deliver enough value.

All of this is done by five people. Staying at five people while turning a profit is probably the dream of every founder in that room.

People make decisions, AI does all the execution

So how five people build a product was the part I most wanted to hear that night. Each person has a lot of authority and handles everything from A to Z. The frontend engineer doubles as the product planner and thinks about everything from product planning to GTM. Everyone uses Claude Code Max, marketers included, and each engineer runs three of them. Yein said using it at full power like that matters to them. In the US, teams of 20 are now generating tens or hundreds of billions of won in revenue, all because they maximized productivity with AI. The team's rule is that AI drafts everything first, and the only thing people do is make decisions.

Of course, it didn't run like this from the beginning.

In the AI era you can run so many things at once that one person could have a hundred things going in parallel. But at first we could only run one or two at a time. So we looked at why, and the problem was finding the work.

When they dug into why they couldn't find work, it turned out they didn't know what problems the product and team had. So they gave everyone homework: every team member finds 20 new statistics a week and shares them in Slack. If you learn 20 new facts, you get 20 new things to do. Looking at the numbers is AI's job too. You tell AI to bring you certain numbers, it brings them, and the person looks and says, "This is a problem, let's try this." Deciding work based on statistics also eliminated pointless work. Now the guideline is that everyone's Claude Code always has 20 or more threads running, and five of them have to be statistical analysis threads.

I thought this part could work at a company of any size. With enough tokens you can run a hundred things, but most teams aren't efficient because they don't know what to hand off. AI runs so fast that people have to run with it, and that homework came out of the team thinking hard about what more to give it.

Everything I touch in the browser, I've tried handing off

I asked how they personally pushed to full power. The answer was simple: hand off everything you have to do, whether it works or not. Saying they were wary of sounding like they were promoting another product, Yein gave the example of the AI browser Aside. Switching from Chrome felt like a hassle, but they just switched. It was inconvenient at first, but once they adapted, they started having Aside do everything they touch in the browser. It writes all their email drafts. They even hand over managing influencers on Discord and tell it to figure it out.

I once heard that people who are good with AI use it the way they'd use an assistant. I hand off all my email too, but I hadn't been handing off scheduling or Google Calendar. I guess that's why people say those who've had a personal assistant are the best at running AI. They've handed off everything, and the gap shows between people who've pushed through to "do I really need to hand this off too, it's easier to do it myself" and people who haven't.

I flipped it and asked whether handing something to AI ever created more work. Yein said there are only two things left that they don't hand off, which means they hand off everything except those two. The first is creating content and writing. No matter how many times they tried, it didn't work. Writing something people will actually want to read still only comes out average. So they brainstorm with AI, but they write anything that goes out to users themselves. With marketing copy and assets, AI often lands close to average and isn't good at copy with a real feel to it.

The second is communicating closely with users. We're a team that cares a lot about talking with users. We do a ton of interviews, and we still handle all customer support ourselves. AI could do it, but I think it's important for me to feel that and keep that sense.

AI can organize the results of those conversations, but understanding customers deeply is so important that it's where they spend the most time. Leaving only the work that keeps you close to customers, to people, and handing everything else to AI sounded a little bittersweet. But I had no argument against it being the right way to use it.

Get past "ta-da, cool" and do a lot of the unglamorous stuff

Finally, I asked for advice for people trying to build AI into their products and people who want their employees to use AI more. Yein said they weren't in a position to give advice, but offered two things from personal experience.

There are honestly so many products that come out going, "Ah, isn't this cool, so awesome." They can go viral, but I think in the end they become products nobody loves.

If you build products, you have to keep asking why users behave the way they do and what they really want, and do a lot of the truly unglamorous, fiddly things. That's how you end up with something worth using. On social media only the "ta-da, cool" stuff looks meaningful, but real customers feel the difference of one small usability detail and are moved by it. If you use AI, right now is the time to go for quantity over quality. Increase your volume tens of times over first, then think about quality. That's how it becomes second nature. What Yein wants to prove over the next year is that their English-speaking customer base becomes larger than their Korean one.

There were two questions from the audience. One person asked how they picked their domain, since domains are expensive. It's about 200,000 won a year, so it's cheap. The second asked how they got past resistance to AI among highly educated users, like in investing and research, and whether it was lowering that resistance or simply the product being good. Yein said there was resistance early on, because many people find it extremely hard to try an unproven product. But instead of that angle, they focused on finding people in pain right in front of them.

Someone who's like, "I'm a first-year PhD, I don't understand this research at all, and I have to read ten papers," doesn't have time to think about whether they're uncomfortable with AI.

Investors are the same. Watching YouTube is usually fun, but for someone with tens of millions of won on the line, it's a desperate problem. And for leaders, studying other companies' AX cases isn't optional. As Junsu put it, it was a very fundamental answer. Find the people who feel the pain point.

Turning gut feel into numbers

Junsu hosted the main fireside. The report on every desk was the 2026 Enterprise AX Benchmark Report from Teamsparta's corporate training team. Junsu praised it for offering a numerical standard for judging whether you're doing AX well, and pulled out the two numbers he found most interesting: 97% said AI training is necessary, but 82% of people who got training said they couldn't really use it in their actual work.

It felt a bit like studying English to me. Everyone knows they need to study English, but if you give them lessons and then say "okay, go speak English now," they say, "I can't."

LinkedIn is flooded with reports, but usually you save them to read later and that's the end of it. So that night we decided to read the report with the person who made it: Jaekyung Hwang, who leads the consulting part of the AX training team in Teamsparta's AX business division. Jaekyung has been in B2B sales for ten years. They started at LG Electronics B2B selling air conditioners to construction companies, did a short stint in overseas sales, sold on Apple Korea's business team, and joined Teamsparta two years and ten months ago.

Junsu's first question was whether they made the report as a lead magnet or because they wanted to capture the difficulties they see in the field.

When you run a lot of training, you often end up telling clients things based on gut feel. How are other companies doing it, what are other companies struggling with, what are they doing well. With each client we met, we were often going on instinct, and we thought it would be good to have quantitative data for that.

What AX leads always want to know about most is other companies. Jaekyung hoped the report would serve as background when designing training or planning company-wide AX, and as something to use in reports to leadership. They added that, of course, it was also made as material to bring in clients and start conversations. That honesty made everything after it more believable.

The number they spent the most time on was also the 82%. Personally, it made them wonder what all their past training had been for. But what they focused on more was the number next to it: when companies did a skills assessment and designed the curriculum well, training succeeded 32% of the time. In the education industry, one in three coming out with high satisfaction is a high number, and that's what told them how much assessment and design matter. For reference, the report's sample is 100 people from large enterprises, 106 from mid-sized companies, and 124 from small companies, 330 in total. By role, 108 HRD managers, 106 executives, 74 C-level, and 42 in training planning. Manufacturing is the largest industry at 45%. Overall, 56% are actually implementing AI, and 22% have reached the stage of spreading it across the organization or beyond.

Everyone within 10 meters is using AI

Junsu pointed out that half the companies said they were on par with or ahead of competitors, yet 29% of those hadn't even started implementing AI. Jaekyung didn't think they were lying. AI just looks different from the field than from the desk at headquarters. People compare themselves to their friends, not Big Tech or IT companies. If you ask a friend whether their company has adopted AI and they say not yet, you figure your company is ahead because yours has. People also don't distinguish between using AI and adopting it. If the person next to you writes reports with ChatGPT, that counts as being ahead too. Whether the company has adopted it and applied it to real work is, of course, a different story.

For people at a desk doing planning or AX, everyone within about ten meters of them uses AI. So when they look around, people are using it. But whether the company is actually using it well probably came out a bit lower.

Then I asked what a company should look at to measure where it really stands, something a practitioner could check tomorrow. Jaekyung gave three: Can you list five things you produced with AI in the past week? Can you name someone outside the team you lead who's good at AI? And is there any work that has actually disappeared because of AI? Junsu loved how concrete that was. At roundtables, he said, the most common worry is how to measure AI. He once told an HR manager he uses AI to organize meeting notes, and they asked whether that was a waste of tokens. Our company has unlimited tokens, so we use it freely, and that story showed me how differently people look at this.

There's a test for whether you're good at AI too

Teamsparta built an assessment that measures whether people are good with AI. Junsu wondered if it was something like "Which of the following is not a Claude model?" with Astra and Codex as options. Jaekyung explained that HRD managers clearly wanted to assess skills, but there was no quantitative data on whether someone is good with AI. There are tests on the market, but they're things like "pick the correct description of prompt engineering," which is closer to B2C. For B2B, they decided you have to measure the ability to use it well, not knowledge of the tools.

So the test covers four areas: prompt engineering, data analysis, ethics and security, and workflow. There are 20 multiple-choice questions and two open-ended ones. The open-ended ones give you a persona, like "You're in HR, and in this situation you need to pull this data. How would you write the prompt?" It isn't a literacy quiz. It checks whether this person and this organization have the ability to use AI well, and whether they have it after training. No company uses it for new-hire recruiting yet, but some have rolled it out company-wide, and one financial company adopted it to add an AI item to its promotion scoring.

You write the report with AI, but the template stays the same

I asked what the 82% who can't use AI after training actually looks like on the ground.

I think the biggest weakness is that people can't bring what they produced in training back into the company.

Whether it's a skill they built in training or the MD file documenting it, it stays in their personal folder. Distributing it to the team and getting the organization to use it is where things fall short. What blocks it is templates and approval chains. You learned AI, but the report template and the approval chain are the same, so nothing changes. Even if you write the report with AI, you have to retype it by hand, and then it becomes, "Eh, if it's going to be like this, I'll just skip AI and do it the old way." Junsu said that "eh, if it's going to be like this" was the line he related to most. That's why people end up doing it themselves again.

Work piles up on the people who quietly use AI well

The report says what blocks AX isn't technology. It's people. Regardless of size or industry, the number one barrier was the gap in AI skill between employees, cited by 55%. When I asked what problems that gap actually creates inside an organization, the answer was meetings.

Put bluntly, when AI ability varies, meetings inside the company fall apart quite a bit. That's something I thought a lot during consulting.

Even prep time diverges. One person prepares in 30 minutes with AI, and another spends time from Monday putting it together the usual way. The quality of the opinions and the time spent thinking differ, so the meeting itself changes. Then work starts piling up asymmetrically on the people who quietly use AI well. "You're good with AI, so go do it well with AI," and work that should be split evenly goes to one person. The same happens with executives. Because of the skill gap, they either don't support it properly, or once they do, they think of it as a magic wand: "Everything worked in the YouTube video I watched, and all you bring me is a translated business email?"

Jaekyung said they've watched, many times with clients, the company's most important meetings fall apart first, followed by the reporting structure and the deliverables. Junsu added something he'd heard at an early roundtable. A company gave an executive tokens and told them to edit a video with AI. The executive had no idea how to use it, so they edited it by hand and submitted it as if AI had done it. It's not a joke. It actually happened.

Set process to the top, training to the bottom

So when some people are good at it and some aren't, where should a company set the bar? The average? Junsu asked. Jaekyung said there's no single right answer, but split it into the two angles they use in consulting: how to change processes and how to run training.

For process, it's much easier to set it to the top tier. Pick about 15% as champions or ambassadors and encourage and push them to build processes that really fit your company. That's the most effective and the fastest. Training has to be set to the bottom tier. If you set it to the top, well, there are actually more people at the bottom. They'll give up on the training within an hour.

In one line: set process to the top, training to the bottom. I also wanted to know why some people refuse to use AI no matter what. Is it inertia, or are they in a field that truly doesn't need AI? The answer was neither.

What I feel is that they don't use it because there's no loss in not using it. It's not that they don't have time. It's not really that hard, either. But not using it doesn't change much for them. No extra benefit, and no penalty either. If anything, there's more risk.

Using AI seems to create more work. Say you used AI and expectations go up. Doing it the way you always have feels like it'll finish faster. Weigh it rationally, and there's no loss in not using it, so there's no reason to bother. That line stayed with me long after the event.

Size differences are money; industry differences are data

AI implementation rates are 76% for large enterprises, 50% for mid-sized companies, and 46% for small ones. By industry, IT is at 78% and manufacturing at 44%, a 34-point gap. I asked whether that gap comes down to the scale needed to pay for tokens. For company size, Jaekyung said, yes, it's money. 41% of large enterprises spend over 50 million won a month on AI, versus 9% of small companies. For over 300 million won a year, it's 28% versus 3%. The amount of tokens you can pay for and the number of AI accounts are the biggest gaps between large and small companies. The industry gap is about data, not money. IT already works with digital data organized in Excel and CSV, so when they go train fintech or IT companies, the data is well organized. In manufacturing, even when the data is digital, often nobody has ever opened it.

I went into a steel mill once. They said we needed to pull data from the machines, and they told us to put on helmets and follow in a line. Then they're brushing off the dust like this. You plug in a USB stick here and the data comes out.

When they asked how anyone had looked at the data until then, the answer was that processed data, massaged to produce good-looking results, was being sent up to headquarters. Jaekyung's summary was that the difference between IT and manufacturing isn't money or ability to pay. It's the difference in data sources and how much that data gets used. Junsu said he once asked to see the data in a meeting and someone brought a pile of USB sticks, "this is 2025, this is 2026," and that's when he understood why AI in manufacturing is hard.

The report shows each industry stuck in a different place. In IT, manufacturing, and healthcare, the skill gap between employees is the top barrier. In finance and retail, the top barrier is the lack of a tailored curriculum, meaning training content that fits their industry is rare on the market. In retail, 88% said it's hard to apply training to real work. In manufacturing, 68% feel a sense of crisis, but executive involvement is only 12%.

Ask what you'll build, not what you'll learn

Even among the most advanced, large enterprises, only 37% have reached the spreading stage or beyond, which means nobody is done yet. So I asked what the companies furthest ahead are doing differently.

They seem to think less about what they'll learn and more about what they can produce. It's not "teach us about Claude Code." It's "we want training where we use Claude Code to build a dashboard from this data."

When requests like that come in, Jaekyung feels the company is ahead. Next, those companies see training as accumulation rather than a program. The first cohort's materials become the second cohort's teaching aids, the second cohort's output helps the third, and so on. One company is already on its 16th cohort and keeps improving. Jaekyung also said they've been noticing cost efficiency a lot lately. Leading companies think about how to use tokens economically. They handed out AI and said "try it," costs went too high, so now they think ahead about using it efficiently and run training by level. Whether large or small, the needs of the companies out in front looked different like that.

The most dangerous companies have both urgency and budget

Junsu said the stories about companies that fail are more fun, and moved to that. Success rates for applying training to real work vary by how training is designed: the same curriculum for the whole company gets 11%, design by job role gets 18%, and level-based design after a skills assessment gets 32%. If the best is 32%, that's one in three. Why didn't the other two work?

Jaekyung said that was exactly right. When you assess skills, design well, and move into training, one in three gets results. For a training company, assessment, design, and training are the entrance. Their job is to pave a good path at the entrance and show people that going this way is easier. Making it all the way down that path and out the exit happens one time in three. One of the problems the other two run into is executive risk. Executives either don't support it, or once they do, they make unreasonable demands like, "I gave you this, so bring me that."

The ones we see as most dangerous are companies that have both a sense of crisis and a budget. If you have both, you really want a magic wand. "I'm pouring in budget and our sense of crisis is this big, so bring me more, better, faster."

Those companies can't stand the path of assessing skills, designing, and then training. The other kind is a company that can't picture what's waiting at the exit. You need a picture of what it looks like when you reach the finish line, but quite a few HR teams say, "Let's just run the training and see what we want." Then it's hard for the training to be good. Jaekyung's summary: Teamsparta can design up to the entrance, but getting to the exit has to be driven by the company's environment and systems too.

The curriculum comes last

Teamsparta trains several of Korea's top ten companies, so I asked what they look at most when designing training. The answer was unexpected.

We look at the curriculum last. The first thing we look at is the company's environment.

It might be the IT environment or something specific to the industry. Some companies use an LLM developed in-house on premises, some use external APIs. They look at that first. Next they look at the distribution of skill levels. The assessment shows whether there are more level 1s or level 2s, and they decide what distribution to watch to bring those people up to level 3 or champion. Only then the curriculum.

There are also companies where they think no design will work, however good. Jaekyung said it shows less the moment they walk in and more the longer they talk. It's the place with no understanding of its environment, doing it because someone above said so, with no concrete blueprint and no clear AX theme: "We have to train fast, the budget came down, we need results fast, we don't know what the deliverable is, so just do it with us." As a case where they'd judged wrong, they mentioned executive training at a large holding company. Executive training is usually insight lectures or mindset sessions. This company wanted laptops set up for everyone, starting from logging in and prompting, and ending with trying vibe coding. This was shortly after Claude Code had come in, and they wanted to do that.

I tried to stop them. "This absolutely won't work. The moment the executives see that black screen, the CLI window, it's just not going to work." But the feedback was that they actually loved that process.

The companies that do well share their environment honestly. They tell you this works, this doesn't, and this is what we need. In training for practitioners at a manufacturing company, every process had different data, and one person vibe coded an agent that collected and unified eight kinds of data from eight processes. When Jaekyung sees something like that, they think: this company knows where its problems are and how to prepare its data, and it's designing its ontology well. Junsu added that he'd heard executives are actually good with AI. Giving tasks to an assistant is a habit for them, so when they give tasks to AI the same way, they get results to match.

The problem isn't the tools; it's knowing where to use them

Plenty of companies probably hand employees AI tools and assume they'll figure it out. I wanted to know why that doesn't work. According to the report, adoption of ChatGPT or Gemini is over 85%, so the tools are already there.

I think these people don't so much not know the tools as not know what to use them for in their own work.

If you're in sales, your workflow runs from A to Z, and someone has to show you that C or D could be replaced with AI. Without that sense, you don't know where, in what context, or how to bring AI in. From the point of view of executives or whoever buys the tools, "we gave it to them, they'll figure it out." But figuring it out is hard. Analyzing your own work and breaking it down systematically is hard, and that's why they insisted on including workflow as one of the assessment's areas.

Juniors came up too. Seniors have done the job for five or ten years, so they know what to feed AI and in what order to get the result they want. Even if they can't use the tools, their domain knowledge is deep, and with a little guidance they get good results. Juniors don't know the work itself, so they don't know what to put in. For them, Teamsparta uses guided projects. For example: you're in marketing, there's no budget, and you have to post an ad banner launching tomorrow to the team messenger today. What prompt would you use to produce it? The design gets people with shallow domain knowledge up to a similar level of output.

Report "4 hours down to 30 minutes" and the follow-up questions come

A lot of people that night have to report AI results up the chain. Some companies used to say they'd measure by token count, which looks absurd now. Everyone uses a different standard, whether time saved or work efficiency, so Junsu asked for Jaekyung's view. Time saved is great for the first report, Jaekyung said. The problem is the second. If you lead with time saved again, executives all ask the same follow-up: what did you do with the time you saved, and what more could you do with it?

Instead of just "we cut something that took four hours down to 30 minutes," say "three reports that each took four hours now take five minutes, and we automated ten of them." That kind of report can shut down the question before it's asked.

The advice was to build the slides for the second and third reports around who used AI, what work they used it on, and how that work was converted into something else.

"Name five things you did with AI last week"

When asked what to do if you could do only one thing at work tomorrow, the answer was practical too, because practical questions show the reality of AI use most clearly.

Just ask two or three people on your team, "Tell me five things you did with AI this week." That alone gives you a rough picture. Honestly, five isn't easy. Doing things with AI at all.

If you have time, ask one more: what work have you eliminated with AI? Not reduced, eliminated entirely. I was also curious what people should ask each other at the roundtable if they could ask only one thing. Jaekyung said asking what doesn't work gets people talking more than asking what does. Ask whether there are rules or situations at your company that block AI use. Small talk starts with things like "we're only allowed to use the internal LLM" or "write ten prompts and it wears out," and then people end up asking each other what's blocking them at their company right now and why it's hard for their company to take off.

There were three questions from the audience. The first person brought their own method. Like the advice to write out your work process, they'd been repeatedly asking AI, "How do you work?" to build and update a canonical document, and they asked if there were more tips like that. Jaekyung laughed and said that if they knew, they'd be using AI far better than they do now. But if organizing your whole workflow right away is hard, focusing on one task and asking about it over and over is a good approach. Then they added one thing.

If your data isn't well organized, AI is really just a very basic model that tells you about the "King Sejong MacBook incident." So the question is how you can teach it the data you have.

(The "King Sejong MacBook incident" is a well-known Korean meme in which early ChatGPT confidently made up a historical story about King Sejong throwing a MacBook.) For sales, that data is your revenue data, customer data, calls and emails, and whether the CRM you use can connect via MCP or be pulled via API. Turning unstructured data into structured data, and structuring it so AI can learn from it, is one of the ways to use AI well.

A pushback: skepticism about training

The second question was a pushback. Someone who works on AI solutions said up front that they tend to be negative about training, especially AX training. They agreed with training that raises literacy, but were skeptical that training reduces people's workload. What people want is for someone to build an agent that solves the problem they're facing. So rather than teaching people how to build tools, isn't the best approach for someone to build the tool that solves their problem and just show them how to use it? They gave an example: if you're assigned to summarize YouTube videos or papers, you don't build a tool, you use something like LilysAI and pass along the result, done. Thanks to this question, the session didn't end as an ad.

Jaekyung agreed that approach is great for building processes set to the top tier. Companies have several problems that aren't clearly solved, and having the 10 to 15% of champions who are good at AI build and spread processes is necessary and may matter more than training. But one thing bothered them.

AI is developing pretty fast. There's no guarantee the agents or skills those people built will last forever. So in the end, the people in the lower tier need to adjust the agents and skills they receive to fit their own work.

UNESCO defines digital literacy as more than reading and writing. It includes using, integrating, modifying, and managing. Building and distributing processes is very important, but to adapt what you're given to yourself at a very small scale, literacy training is essential for the lower tier. Jaekyung added that it might come down to how broadly you read the word "literacy."

Development was 15%; defining the problem and the data was 85%

The third question was about the whole company. People often say you shouldn't bolt AI onto existing processes but redesign work with AI as a given. Was that happening inside Teamsparta, and could they share a detailed example? Teamsparta has a separate organization for AX and spends a lot of time redesigning and changing how work is done.

Take just the training team. Building courses, managing them, meeting clients, recruiting instructors, handling instructor pay and tax invoices were all split up into fragments. They wanted a SaaS product that tied it all together, but no such SaaS exists, no CRM fit, and nobody could build it for them, so they decided to build it themselves. They started by defining the work by part and mapping how the parts connect, and redesigned the workflow. When they taught Claude that, it put the system together faster than expected. Company-wide, that led to finance management, a hiring system, and document automation connected to Google Workspace.

I think the actual development took less than about 15% of the time. About 85% went into defining the problem and gathering data.

The key isn't which tool you use. It's the step before that: looking at the whole job again end to end. To plan company-wide AX, building your data and organizing it in a form AI can understand is what matters. Jaekyung closed by saying AI handles the rest extremely well.

3.6% actually used assessment results to design training

I also pulled a few numbers from the second report we got through a QR code, the 2026 AX Training Trends Report. It's from the same 330 respondents. 56.3% of organizations are implementing AI, and 54.5% feel their company is behind other Korean companies in AI adoption. 74.6% said their executives are interested in AX, but only 15.8% of executives personally drive it as a top priority. The most common trigger for feeling AX was necessary was AI news and content at 40.6%, followed by executive orders at 22.7% and competitor examples at 17.9%. As difficulties, 54.8% said employees' AI levels vary widely, and 50.9% each said there's no curriculum fitting their job and industry, and that training doesn't carry over into real work.

The report also shows why Jaekyung talked about assessment first. 51.5% of organizations feel they need skills assessment but don't know how, and 3.6% have actually used assessment results to design training. For measuring results, 29.7% look at productivity and time saved, 28.8% at frequency of tool use, and 20.3% want to measure but don't know how. 61.5% plan to complete company-wide AX by 2027, but the largest group, 27%, has a plan without a dedicated budget. For desired training, job-specific practical training led at 60.6%, well ahead of company-wide common literacy at 33%.

Demand is changing too. In inquiries Teamsparta received, the share asking for general training fell from 96% in the first half of 2025 to 51% in the first half of 2026, while job-specific training rose from 4% to 41%. The number of training types requested per inquiry rose from 1.24 to 1.89. That means requests are shifting from "please do AI training" to "design general training together with job-specific training and leadership training for managers." In a prompt challenge with 262 HRD managers, the basics were solid: 93% specified the task topic and 85% specified the output format. But task decomposition, breaking work into steps, was weakest at 42.4 points, and only 37% asked for the results to be verified. The report's conclusion is that people need to learn the flow of breaking things down, assigning them, and verifying, instead of "just do it all for me."

The person with the priciest AI plan ran each table

For the last hour, we sat in groups of six. The table lead was whoever at each table subscribed to the most expensive AI plan. Each table discussed two questions: where is our organization's AX, and what's the evidence? What's the wall that keeps training from carrying over into real work? Then the six table leads came up front to share what their tables talked about.

One table was led by the CEO of a one-person company founded four months earlier. With them were a member of Teamsparta's training team, an engineer at a large manufacturer, and a solo entrepreneur who teaches corporate communication. The Teamsparta member confidently said the company had reached the cost-cutting stage, with real deliverables and concrete numbers. The manufacturing engineer was getting results gathering field data, including from overseas plants, into one place and processing it and training on it in an on-premises environment.

The communication trainer recently vibe coded a website for a doctor client. As a non-developer, the coding worked surprisingly well, they said, but outside of that, AI was still very hard. The most candid story came from the table lead. They'd started their company after watching a lecture by someone saying they'd build a company that runs AI agents like employees, but doing it themselves was different. They come from engineering, so they're confident they can build all the way to a finished product. But a business owner isn't a builder. They have to revise their own strategy and do sales and marketing too, and lately they've been feeling the limit: running agents like employees in those areas is still hard.

A PO who leads an AX task force at a services company had started ambitiously, wanting to help non-developers use AI. But the business teams were busy, and learning AI took extra time, so the reaction was that it felt like double the work. They couldn't roll it out as much as hoped and came to figure out how to lower that hurdle. The answers from the table went two ways. Mature organizations had such wide variation in the skills each person built that they were moving toward managing them in one place. Another organization, driven by its CEO, holds a hackathon every two months and adds extra training for people who have trouble keeping up, and keeps pushing.

At the table led by a new-grad developer, the conclusion was company age. At their company, individuals don't build their own tools. Everyone shares one thing built by the AX team. Seeing it as a process was good, but the company has been around a long time, so it's hard to suggest new things. Other team members use AI well in parts of their work but struggle to see their whole job as a process. The table agreed that the more a company has built up over the years, the harder an AX that tears it all up becomes, and that permission from above is the biggest wall. A founder at the same table said new companies can boost productivity with AX easily, but organizations with a lot built up need to think harder about how.

The table led by an MLOps engineer at a game company was a mix of a CEO, a CEO who teaches, a content marketer, a PM, and engineers, all with different interests. The CEO wanted the output of several employees. The content marketer was thinking about how to work AI's uncanny valley into content. The PM was thinking about how to share prototypes. The engineer's own concern was this:

The bottleneck in AX for engineers is people, in the end. AI is so fast and so smart, and I'm slow and dumb.

Then they asked whether they really had to say something, and Junsu said he wasn't sure either, which made everyone laugh.

A person from the Korean branch of a Japanese IT company became table lead after mentioning that most employees use an enterprise plan and the company pays substantial token costs every month. Their AX is at the early stage where the dev team leads and its methods spread to other departments. But work has shifted toward one senior handling the full stack, so it piles up on one person running four or five projects alone, with no capacity left to onboard other departments. Key people are overloaded. Individual productivity went up, but to use AI at the team level, they concluded you have to redesign the org structure first.

The last table lead handles AX as the only AI specialist at a long-established animation IP company. A patent attorney and two interns were at that table. The shared sense was that every organization has huge individual variation and different attitudes and willingness toward AI, so AX still has a long way to go. At content companies, many people ask why they should do this at all and worry their IP will get stolen. The lead wished the report had included content as an industry, and wondered if there were cases where AX gets blocked by practitioners afraid for the security of their IP and resources.

The patent attorney talked about knowledge workers worried that tools like Astra will take all their work. The interns felt that the way you get recognized for your abilities inside a company is changing. One person at that table said their company lets them use Claude or GPT however they want but offers no training, and they learned that night for the first time that something called Claude Code exists.

"There's no loss in not using it" keeps coming back to me

In closing, Junsu said that after roundtables there's always a lot of off-the-record talk and people saying they wish it had gone longer. That's why Bloom runs Mixer, groups of four to six continuing these roundtables in Gangnam, and we keep thinking about what can only be made offline.

On the way home, I thought about how the two firesides met in the same place. Yein's team of five could run a hundred things at once but was only running one or two, because they couldn't find the work. The companies of hundreds that Jaekyung sees were stuck too, not for lack of tools, but because people didn't know where in their work to use them. Whether it's five people or a thousand, they get stuck in the same spot. Yein solved it with homework: 20 statistics a week. Jaekyung solves it by writing the workflow out from A to Z and finding C and D.

And the line "they don't use it because there's no loss in not using it" stays with me. Our company has unlimited tokens and I hand off everything starting with email, so this was the first time I'd heard the math of people who don't use AI laid out so clearly. I still don't know whether what changes that math is training, process, or the approval chain. But next week I plan to ask two or three people on my team to name five things they did with AI last week.

Originally written in Korean on Brunch, September 16, 2026.