AIM Intelligence × AWS × Classmethod: AI Workflows That Work at Scale

We ran a Bloom event with attorney Hyunseo Lim and Haon Park, CTO of AIM Intelligence. Yongho Choi from AWS joined us too, along with an executive from the Japanese headquarters of Classmethod, a Japanese AWS Premier Partner (the top partner tier). It was a strong lineup. Turnout was far beyond what we expected, so we had to open the room next door at the last minute. In the end more than 150 people came.

I think it's fair to call it our most successful event so far. With four speakers, I worried it might drag. Instead, not one person left. People stayed nearly 30 minutes past the two-hour mark, listening closely and networking. The balance of the whole program was about as good as it gets.

Thank you again to our speakers. Here's my recap of the day.

Intelligence doesn't matter much anymore

In the AI era, intelligence doesn't matter as much as it used to. A high IQ, knowing a lot, a long career: AI has already started to replace those. Building one good open source project and getting it to go viral now counts for more than a good resume. So what matters? Insight into the market, good instincts, ideas, and above all the ability to execute fast.

Solo builders and small startups can do that. Organizations have a harder time. There are internal processes and approval chains, and nobody works alone. That's why you hear "people are the bottleneck" so often these days. When I asked the people from large organizations to raise their hands, they related so hard they almost looked upset about it. The bigger the organization, the bigger the security risk, and an agent that works fine for one person gets full of holes once you run it across a team. So this time, instead of focusing on how individuals can use AI well, we invited four companies to talk about building workflows that hold up in real, day-to-day work. The venue was AWS Center Field in Yeoksam, Seoul, with the space sponsored by AWS.

How Amazon uses AI agents

The first session was from Yongho Choi, a tech evangelist at AWS, on how Amazon uses AI agents. He checked the room, found that a third of the audience were developers and two thirds were not, and kept it high level.

His first example was delivery. Amazon.com gets millions of orders a day, and first-attempt delivery failures were common. Even at the same address, one apartment building lets you go to the door, another wants packages left with the security desk, and an office can't receive anything after business hours. People used to keep notes on all this and handle it by hand. Amazon replaced that with an AI agent. A person asks a question, and the agent acts as the orchestrator: it searches the internal database first, and if the answer isn't there, it pulls in public address records, geographic data, and even company business hours, puts it all together, and gives guidance. Manual work dropped sharply, and the first-attempt delivery failure rate fell by 74%.

The point he stressed was that building one agent isn't the end of it. For agents to get reused inside a company, you need a platform. So Amazon built an internal platform called Agent Z, where each team's agents pile up over time. As of May, there were more than 23,000 of them. Search for what you need and it usually already exists, which cuts developers' workload a lot. Non-developers can type what they want to build in plain language and get guided through building it.

He introduced Kiro, AWS's coding tool, and said the strength of these tools is that they're spec-driven. Normally when we prompt "build me a shopping site," the AI makes a huge number of decisions for us. What to use for the frontend, what to use for the backend. The result looks plausible but may not be what you wanted, and every round of back-and-forth burns tokens. So these days you have the AI write a document first. The document holds the AI's plan for how it will build the thing, a person edits it, and once both sides are in sync you start the big job. The results are better and the token efficiency is better. Amazon built Spec Studio to share these documents, and about 15,000 specs have been created there. Other teams use the well-written ones as references, which speeds development up even more. The common thread was clear: don't keep it to your own team, spread it, and to do that, turn it into a platform.

Another example was a new framework for Amazon Bedrock. It was originally estimated as a job for 30 developers over 18 months. Using spec-driven tools, 6 developers finished it in 76 days, and it's now deployed and running in production. The interesting part is that for the first few weeks they barely felt any productivity gain. Productivity only jumped once they realized that using AI wasn't enough and they had to change the work process itself. They took away three lessons. Giving the AI a goal and letting it run on its own worked better than ordering it through each task one at a time. A multi-agent setup, with several agents working together, beat going back and forth with a single agent. And running for hours or days produced better results than exchanges of a few seconds, because the AI runs a lot of tests and checks along the way. He said recent features in Claude Code and Codex are evolving in exactly this direction.

Then came questions. A non-developer said their agents kept failing to work properly. He said it isn't about whether you're a developer. AI just doesn't do everything perfectly, and you need prompt engineering, context engineering, and a process of analyzing why a result was bad and building up guidelines from that. Asked whether Amazon has internal token limits, he said every model has rate limits. Asked which model he prefers, he said personally he likes Claude best, though a lot of people are switching to Codex lately. AWS has strong partnerships with both OpenAI and Anthropic, both models are on Bedrock, and people use whichever fits the purpose.

Where Japan and APAC are

I set up the second session because I was curious about what's happening outside Korea. We invited Akimasa Omori, an executive officer at Classmethod's Japanese headquarters who also leads Classmethod Korea and Classmethod Malaysia. Classmethod is an AWS Premier Partner in Japan. He spoke in Japanese, with live interpretation from Tiro. We've done plenty of English sessions, but this was our first in Japanese, and the oddly global feel of it gave me chills.

He started with Malaysia and Thailand, where he'd been the week before. Those markets are still one step short of adoption, and cost and security worries keep things from moving. Japan isn't as far along as Korea but is adopting bit by bit, and he named four recent trends: AI-driven development, security, reverse engineering, and company-wide rollout. AI-driven development has gone past coding help into things like preparing materials for regular meetings. On security, companies have AI read large amounts of source code and config files to detect risks. Reverse engineering has gone beyond converting languages like COBOL to having AI read tangled legacy systems whose internals nobody understands anymore and document them. The hardest is company-wide rollout. A chatbot is fine, but getting internal knowledge into a state where it can be shared takes a long time.

He asked the Korean audience to raise their hands if AI had made them more productive, and quite a few hands went up. Ask the same question at a Japanese company, he said, and almost no hands go up, so Korea really is ahead. Then he talked about "work slop" in Japan. AI produces lots of documents that look nice, but when you read them you often can't tell what they're saying, and people spend an average of two hours a day interpreting and fixing them. Productivity actually goes down. When he meets with executives, he often hears, "We brought in AI too, but we're not sure we're using it well." So he stressed understanding context. Japan is a high-context society, and Korea probably is too. If you don't understand the context before you ask, you won't get the right answer, and things drift further off course.

The biggest problem at his own company, he said, is that ideas pour in and nobody executes them. So he keeps telling people to execute and get results instead of just having ideas, and to take one step, even a small one. He thinks the future needs people like the "expert generalist" that's been getting attention lately: someone with depth in one area who can execute and drive things across areas. He also gave AI two meanings. Besides "Artificial Intelligence," there's "Actual Intelligence," meaning your own intellect, which is the most valuable part. Work goes well when you use AI and your own intellect together to execute. That's why communities like this matter, he said: when we share best practices and understand each other's context, we get a world that uses AI better. He closed by saying Classmethod Korea would keep supporting communities like this.

Asked how he defines intelligence, he laughed and said it was a hard question. Studying at a desk builds knowledge, he said, but it turns into intelligence through actually using it. Asked where AI is used most efficiently in Japan, he said that since many companies there are engineering-heavy, demand is high for AI-driven development and reverse engineering, and there are many cases where work that took six months now takes a few weeks.

When a lawyer meets an AI safety expert

The third session was the main event of the day, a fireside chat. Our guests were Hyunseo Lim, an attorney who specializes in bankruptcy and insolvency and calls himself self-employed, and Haon Park, co-founder and CTO of AIM Intelligence, a Korean startup focused on AI security and safety. One came from a deep domain, law, and the other from the side that builds AI, so their views collided in a good way.

I started by asking about a moment AI made them feel like Superman, and a moment AI made them think they were in real trouble. Park said he's had both. His company's tool automatically finds a client's website chatbot before a meeting, finishes the assessment during the meeting, and delivers a report as soon as it ends. That level of automation was impossible before AI. On the other hand, he used to read a lot of papers while leading a research team, and now AI summarizes dozens of them and even applies what's needed, so reading them one by one by hand feels inefficient. Lim said he does about 2.5 times the work he did a year and a half ago, yet he used to pull all-nighters and now sleeps through almost every night. When he gave a talk at Seoul National University, he recorded everything from the outline to the Q&A and published it on the web the moment it ended, and people loved it. Because of that real-time speed, he said, he's seeing a new world, to the point that sleeping feels like a waste.

Some say heavy AI use actually wears people down, so I asked about that. Lim said his work-life balance has gotten much better. He said AI's strength is taking on emotional labor for you. Groveling, asking favors, the unpleasant stuff you don't want to do yourself: AI is good at it. When he replies to a client, he deliberately asks it to write something detailed, kind, and long, it does exactly that, he tweaks it a little and sends it, and the other side can't tell. He laughed that if he keeps telling people this, he'll get caught one day.

Written filings matter a lot for lawyers, so I asked how much AI-written material actually goes out. He said each office keeps that to itself. But ChatGPT output comes with double asterisks, and he has actually seen another lawyer's brief submitted to court without them removed. It was a panel dedicated to intellectual property cases, so the judge probably noticed, he said, laughing. The legal world, he said, is a place where you'd be a fool not to use AI, and yet it's full of fools, so the gap between people who use it and people who don't is suddenly getting very wide. The lag is so bad that some older lawyers can't turn on their camera for a video hearing even when asked to.

I asked Park to explain more about what his company does. They handle AI security on both offense and defense. When clients build AI services, AIM Intelligence supplies a red-teaming solution that automatically assesses the risks, and a guardrail solution that blocks the vulnerabilities it finds. Because AI moves so fast, they also work with big tech companies abroad and test models before they're released. For example, Anthropic cares a great deal about making sure its models won't give out instructions for making biological weapons, so AIM checks for that in advance. They also test and help improve guardrails from Meta and OpenAI.

The NASA story stuck with me. Since red teaming is their core technology, while looking for customers they found an AI service connected to NASA's Parker Solar Probe and tested it. Even though it was an official chatbot, some dangerous responses came out with no filter or guard. They reported it to NASA, NASA took it seriously and shut the service down entirely, and AIM got letters of recognition from NASA's chief innovation officer and its security chief. The chatbot wasn't connected to any classified database, but NASA saw the risk in an official chatbot presenting wrong information as official information with no way to verify it.

I asked whether hacking really is surging in the AI era. Park said yes. Even before quantum computers arrive, models have gotten better, so they've gotten better at cyberattacks too, and institutions and companies no one would have imagined are getting breached. Unlike before, ordinary people can now very easily do expert-level hacking, which has raised the risk, and he said it's no surprise governments are regulating. On global expansion, they're going after the US, Europe, and the Middle East in parallel. Saudi Arabia and Dubai have huge state-owned AI companies like Humain, so landing those could win the whole region. Europe has plenty of regulation but few companies building technical solutions for it, and there's a trend of moving away from US solutions, so there's a lot of room for a Korean company.

The conversation turned to the future of the legal profession. Lim said he uses AI anywhere he can, from research to communication, and that work written by people actually seems worse. When a new GPT came out, he used it for two days and thought, "Humans are done." In general reasoning and intellect, he thinks it's better than over 90% of lawyers. The limit in Korea is that court rulings aren't publicly available, so combining AI with domain knowledge is weak. Still, he said, all of it will be replaced. People used to say an AI judge made no sense, but he's changed his mind. In Korea, which judge or prosecutor you draw matters a lot, because everyone's bar for reasonable doubt is different, 95% for one, 90% for another. If you can keep that intellect more consistent and manageable, why leave it to judges with individual variation? The process of arguing it out will just take a long time, and the system isn't ready to be fully computerized yet.

His take on copyright was vivid. He called the current state of copyright for AI-generated work "a free-for-all." Work made purely by AI generally isn't recognized as creative, but it is if a person contributed enough, and that's hard to verify. In a recent case, someone claimed they drew a picture themselves when anyone could see AI made it. When he asked what evidence he had, the answer was "six fingers." The picture really did have six fingers, and the other side admitted it. Copyright only protects the form of expression, so copying an idea isn't infringement, and if the copier says they used AI to copy it, there's no way to know. Everything is up in the air. His conclusion: the paradigm needs to change completely, and right now nobody takes responsibility.

I asked whether coexisting with AI means people should step back or take more control, and whether there's anything they'd never hand over to AI. Park said if you grip it too tightly there's no point in using AI, but as a security company control matters too, so finding the balance is the key. He has his employees use AI a lot, but any email that goes out in his own voice, he checks himself before sending. Lim said he doesn't auto-send either, but that the speed itself already seems to be beyond human control. There are 7 billion people who all think differently, and with things moving this fast, he said, it feels like a huge sweep of history is already underway.

Finally, I asked who survives in the AI era. Lim's answer was "diligent clickers," the people who keep hitting the button and letting AI do the work. The world changes fast, people change slowly, and plenty of people won't use AI for no reason even when told to. The day before, he'd sent his whole firm a one-line notice: "Please make good use of Claude." There's no reason at all not to use it, he said, and inertia is a strange thing. Park said the survivors will be people with the skill to take responsibility and verify. AI is still a while away from taking responsibility, so we still need people to check its work, and you only develop that eye by doing a lot of work in the domain. So he thinks becoming a specialist in one field is a better way than being a generalist to push back the day you get replaced. Lim added honestly that because AI's general reasoning is so strong, he's afraid only a small number of people who can keep up with it and verify it will be left.

An audience member asked what qualities people will need in the future. Lim picked the ability to understand people. AI still doesn't get the subtle reading of the room and strategic judgment that takes. He does a lot of litigation and criminal complaint work, and some people scare more easily under pressure while others do something unexpected. Understanding people well enough to weigh the odds each time and come up with a reasonable plan is hard to teach. That animal instinct is worth a lot, he said, and the higher you go, the more the complexity comes from managing relationships inside and outside, so it will stay valuable for a while.

The questions kept coming

The audience questions were heavy too. An in-house lawyer at a securities firm said they use AI for everything, but whenever they get proper legal advice, issues come up they hadn't known about. How do you use AI so it produces something like a lawyer's opinion letter? Lim said the paradigm is shifting. Billing by the hour is getting hard to justify, and if an AI with strong reasoning that knows the regulations can find the answer, you no longer need advisory lawyers who are basically paid searchers. The collapse of the advisory market has already started. What remains is consulting, strategy, and the hands-on expertise only firms that have been through a process can offer. Registering a fund overseas is one example: no amount of searching will get you the answer, and only firms that have done it know.

There were also questions revisiting responsibility and copyright for AI-generated work, and one about the fatigue of switching contexts while running several agents at once. One attendee shared how they ran hundreds of agents in Claude Code for eight hours to map thousands of PDF pages. With ten windows open and jumping between them, they said their brain wears out fast, almost like ADHD. Lim told them to accept it as a lifestyle: mental fatigue is still better than physical fatigue.

Toward the end, I brought up UBI, universal basic income. It can sound like saying we'll graciously adapt as the system collapses, but won't the world actually swing to extremes? Maybe that anxiety is exactly why we gather at Bloom and put our heads together.

What came out of the roundtables

After the fireside chat, we split into roundtables of four. It can feel unfamiliar to first-timers, but hearing what the other tables talked about is where most of the insight comes from. Each table lead shared for a minute, and the points fell into a few groups.

The topic that came up most was generalists versus specialists. One comparison stuck with me: right now AI fills in the empty areas, so generalists have the edge, but once all the areas are filled, the next moves will belong to specialists. Along the same lines, now that knowledge is a baseline for everyone, what matters is the domain expert who can point out where to go next within that domain.

How people use AI split roughly into automation, life coaching, and doing what used to be impossible. Automation covered code review, writing documents, cutting Shorts out of videos, making card news (Korean-style slide carousels for social media), and even having an agent built on someone's own persona review hundreds of Jira tickets. The life coach examples were fun. One person had AI plan their weekly routine and coach them on diet, supplements, and exercise, and went from squatting 65 kg to 120 kg on a 5x5 program. For doing the impossible, two stories stuck with me: someone who used Claude Code to build a Korean interlinear Bible that matches the original language to Korean word by word, and someone who made a meal plan for their mother, who has leukemia. Another person canceled Netflix and subscribes to a music AI instead.

Several tables also talked about building a second brain by collecting their own knowledge in an LLM wiki, and said that with so much data, documenting and accumulating it is the real challenge. The limits were clear too. Work involving sensitive personal data and accountability, like HR payroll, is hard to fully replace. And while things start with one click, nothing is done in one go; you have to keep iterating to get something usable. Many tables landed on the same idea: AI works better when you use it on a foundation of understanding people.

People stayed late. Between the anxiety that AI might replace everything and the hope that there will still be room for people who verify, take responsibility, and understand others, we were each moving forward one step at a time, clicking away.

Originally written in Korean on Brunch, June 30, 2026.