97% say they need AI training. 82% still can't use it at work

What was discussed

Wonjun took the mic and opened with this: everyone says they are good at AI these days, and he has yet to meet anyone who says it against a standard, so let us build that standard here tonight. Then he asked for hands. Plenty went up from people leading AX adoption at their companies. Almost none went up from people in HR who had run training and felt their colleagues still could not use it. The gap between those two questions followed us all evening.

The opening fireside was Kim Ye-in of LilysAI. It is a tool that helps people who have to read a lot understand faster: throw in YouTube videos or PDF papers and it summarizes and answers questions. The figure of one million we had seen in a recent article was corrected on the spot to 1.3 million. And she said that is not the number she watches. The public number is one thing; internally they track retention and paid conversion, so when 1.3 million came in she was happy for a moment and moved on.

What she brought up instead was Readray, a new product aimed at North America. Seventy percent of users come back a week after using it, far higher than LilysAI. The reason she gave was not intelligence but usability. An extraordinary knife has arrived, and because it is so good people are cutting their hair with it along with their fruit. Hair is much easier with scissors. AI's penetration into daily life is still at one percent, and for reading, the best experience is information augmented right in front of you like Iron Man's HUD, so Readray adds research alongside you without being asked.

We had to hear why they built a new product while the existing one was doing well. The answer was a dream. She and her co-founder started the company dreaming of building software used worldwide, like Notion, Canva, or Grammarly. The numbers were striking. Karrot, considered Korea's most successful startup, does 270 billion won a year, while Grammarly built a trillion out of grammar correction alone and Canva more than four trillion. Building that kind of product means getting far sharper. It took a year to see it. People keep warning that each new model is a threat, and retention has never moved, because their customers watch enormous amounts of video to make investment decisions and cannot decide off a five-line summary.

Asked whether failing to penetrate the US was a product problem or a marketing problem, she said both. North American UI is far more extreme in its simplicity. In Korea people need a label on a button before they will press it; in North America, labels on every button draw complaints about too much text. Conversely, Koreans will not press an unlabeled button because they are afraid of what it might do. Community worked differently too. In Korea they grew by summarizing good material and sharing it in open chat rooms. On Reddit the filtering is severe, and the first response to anything is isn't this an ad, who even are you.

They have never raised. They crossed breakeven within a year and have been profitable since, and since it all comes out of your own pocket they learned to manage costs from the start. The team is five people. How those five build the product was what we most wanted to hear. The answer was authority. One person carries something from A to Z, so the front-end engineer also does product planning, marketers use Claude Code Max alongside engineers, and each engineer runs three instances. It was not always like this. You can run a hundred things at once in this era, and at first only one or two were moving, and looking into why, the problem was finding the work. Her prescription was homework. Every team member had to find twenty new statistics a week and post them to Slack, and twenty new facts produced twenty new things to do.

Only two things are left that they do not hand to AI. One is writing. However hard they tried, the output stays average, so they ideate with it and write anything going out to users themselves. The other is talking closely with users. They do a lot of interviews and still handle customer support themselves, and while AI could do it, she said it matters that she feels and keeps that sense herself. Her closing advice ran the same way: products that go ta-da go viral and end up loved by nobody, and it takes constantly asking what users actually want and doing a lot of unglamorous work before something usable comes out.

The main fireside was Hwang Jae-kyung, who leads the consulting practice on Team Sparta's AX training team. What sat on every desk was the 2026 Corporate AX Benchmark Report he built. Two numbers stood out. Ninety-seven percent said AI training is necessary, and 82 percent of those trained said they still cannot use it at work. Wonjun compared it to studying English: everyone knows they should, and after the lessons, being told to go and speak English still gets you a no.

Half of companies said they were at par with or ahead of competitors, and 29 percent of those had not started adoption at all. Asked where the gap comes from, he said they were not lying; the AI you see on the ground and the AI you see from a desk at headquarters are simply different. The comparison is not big tech but a friend. For people planning from a desk, everyone within ten meters uses AI. So we asked what a company should look at to measure where it actually stands, and three things came back. Can you list five things you produced with AI last week? Can you name someone on another team who uses it well? Has any task actually disappeared because of AI?

What does that 82 percent look like on the ground? The biggest problem is that nothing produced in training makes it back inside. A skill, a file, whatever it was, stays in a personal folder. What blocks it is document templates and approval chains. You learn AI and the report format and the sign-off chain are unchanged, so nothing changes, and writing with AI adds the manual work of transcribing it back, until people decide they may as well do it the old way. The top barrier in the report was not technology but people: differences in AI proficiency between employees, cited by 55 percent. That gap first breaks meetings. One person prepares in 30 minutes while another spends from Monday doing it the old way, so the thinking behind what they bring differs, and work piles asymmetrically onto whoever quietly uses AI well.

So where should a company set the bar? His answer split in two. Process should be set to the top. Pick the roughly 15 percent who are champions and encourage them to build the process that fits your company; that is fastest. Training should be set to the bottom. Aim training at the top and the people at the bottom give up within an hour. We also asked why some people will not use it no matter what. The answer was neither inertia nor industry. It is that there is no penalty for not using it. They are not short of time and it is not difficult, and with no gain and no downside, there is no reason to bother.

Differences by size and by industry split as well. Adoption sits at 76 percent for large corporations, 50 percent for mid-sized, and 46 percent for small. The size gap is money: 41 percent of large corporations spend more than 50 million won a month on AI, against 9 percent of small ones. The industry gap, though, is data rather than money. That is why IT at 78 percent and manufacturing at 44 percent pull apart. IT already works with data organized in spreadsheets and CSVs, while manufacturers often have digital data nobody has ever opened. The steel mill story stayed with us: helmet on, following a line, dusting yourself off, and being told that if you plug a USB stick in here the data comes out.

The last hour seated everyone in groups of six, with the table lead assigned to whoever paid for the most expensive AI plan. One lead admitted he had founded a company meaning to run AI agents like employees and found it different in practice. He can see a product through as an engineer, and a founder has to revise the strategy themselves and handle sales and marketing too. A line from an engineer at a game company stayed as well: the bottleneck in AX is ultimately the person, because AI is so fast and so smart while I am slow and stupid. On the way home, the thought was that both firesides met in the same place. Five people or a thousand, the sticking point was identical. Not a lack of tools, but not knowing where in your own work to put them.