AB180 × Bloom: AI for Growth Marketing

We held a Claude Bloom event with Sungpil Nam, CEO of AB180, and Jehim Choi of Eungdapara Marketing, a Korean marketing newsletter and community.

Jehim said AI trends move so fast that any special know-how you share is out of date within a week. So Jehim focused on cases from their actual day-to-day work, which made the talk land much harder. Two things stuck with me: all five winners of a Claude hackathon came from non-developer backgrounds, and what Jehim said after recently winning first place at the OpenAI Codex Skillathon. AI is only a tool, and you create much deeper business value by running straight into the domain problems that come up in real work. I strongly agreed.

Sungpil shared the story of running a startup for 11 years and growing it to around 40 billion won (about $29 million) in ARR. When the mobile wave began, he focused on performance advertising, and to buy time to build and launch his own product, he won the exclusive Korean distribution rights for the global marketing tools Amplitude and Braze. He also landed MMP partnerships with Meta and Google after many meetings, the first and only company in Asia to do so. I never imagined how hard and grueling that was. For anyone running a startup, it was packed with insight on strategy, development, team building, and more.

I'm really grateful that Claude Bloom keeps getting to work with such a range of people who are working hard in the field.

The person who defines the problem wins

On Thursday, June 18, Claude Bloom × AB180 × Eungdapara Marketing took place at AB180's office in Seolleung, Seoul. The theme was AI for Growth Marketing.

More than half the people there were at Claude Bloom for the first time. About 30% came through Eungdapara Marketing and about 30% through AB180's newsletter. A good mix. I checked that everyone had their coffee and dessert, looked at the people standing and chatting in the back, and started the session. AB180 sponsored the event by letting us use the shared space in their office.

Claude Bloom is an AI community we started two months ago. We've covered all kinds of topics, and everyone who comes is an AI early adopter, though their jobs and industries vary. This time we made it domain specific. We narrowed it to growth marketing and marketing, so most people there worked in marketing, and we set it up so we could go deeper into their actual work and roles.

I have history with both speakers. Sungpil Nam of AB180, who joined me for the fireside chat, stepped in after my first event. That first event honestly didn't go well, and he gladly sponsored the space and catering so we could hold our second and third seminars in his nice office. AB180 has supported Claude Bloom almost from the start, with a founder's mindset. I also come from marketing, so I'd wanted to do this topic, and the collaboration happened naturally. My history with our special speaker, Jehim Choi of Eungdapara Marketing, goes back five or six years. When I was doing marketing at my previous company, HYBE, Jehim reached out wanting to write about it in the newsletter. There was a lot of sensitive material, but that was the first time I ever made newsletter content. Both of them connected with me through marketing, so it became a three-way collaboration.

Last week we worked with AWS and LG Uplus on the sales profession. That came from wondering whether there are parts of sales that AI can replace or do better. Picking marketing this time came from the opposite direction. There's so much you can do in marketing with AI, and I wanted to know how far it can go. Sungpil's AB180 is a company more than 12 years old. It runs the marketing software Airbridge and hosts Korea's largest marketing conference. Jehim does marketing at T'order (a Korean tabletop ordering tablet company) as their day job, and recently won first place at a hackathon hosted by OpenAI. I wanted to hear how a non-developer who knows nothing about coding used AI to build a campaign for their actual marketing job and took first place.

1. So, what's the problem?

Our special speaker, Jehim Choi, who runs Eungdapara Marketing, gave the first session.

Jehim said AI is a field where it's hard to give a right answer. After winning first at OpenAI's Codex Skillathon, Jehim got talk requests from several companies and told them they couldn't make a talk. If you talk about a feature, a week later it's outdated or doesn't work anymore. So Jehim decided sharing cases was better. Not a feature demo like "Codex can organize your desktop," but showing that it can solve a certain problem. When people who haven't tried AI yet see that, they feel AI has come a long way and can solve their problems.

The talk was titled "So, What's the Problem?" Jehim has been working for about six years, and an executive at their first company, a large corporation, loved that question. He was big on the five whys. When you reported a problem, he'd ask, "So, what's the problem?" You'd answer, and he'd ask again, five times, and if you couldn't get to a solution in that time, he'd chew you out. So Jehim sees the key skill in the AI era as defining the problem and explaining it well to AI. That's how marketers and non-developers will grow from here.

Eungdapara Marketing aims to be an all-in-one marketing brand. Jehim has made the newsletter for about five years, and it has about 20,000 subscribers. It's core and niche. A subscriber survey found that 77 out of 100 are marketers. There's also an open chat room where people can ask anything, from dating advice to how marketers should grow to how to use AI. Gathering marketers in one place made Jehim very interested in what marketers are worried about in the AI era, and recently they ran an AI sharing session for marketers' growth. People met at a coworking space like SparkPlus to share how they use AI and how they solve problems. Features go stale in a day, but cases last. That's how Jehim sees it.

Jehim used to lack confidence with AI. As a marketer, you'd watch YouTube and hear things like "use AI and make 2 million won a month automatically" or "drop in one link and all your content comes out automatically." Seeing that, Jehim felt they just couldn't use AI. Then at OpenAI's Codex Skillathon, Jehim simply showed the work they actually do. For T'order, the soft landing right after a store adopts it matters. You have to show the owner the technical value, so within a week of signing, they send stats on how much revenue came through T'order and how much extra revenue it's generating. Say a malatang (Chinese spicy soup) shop adopts T'order. The report shows the spicy malatang was the top seller and the milder one came second, how long customers stayed, and the most popular menu items and times of day. Jehim used to download store data with the sales reps and run pivot tables in a spreadsheet. That alone ate up the whole day. Talking through the process of the problem with Codex, Jehim turned that work into a report-making skill. Put in T'order order data and the AI sends a web report to the customer. The room loved it. It wasn't a flashy AI trick. It made people think, I could actually use this in my own work, I could do this too. Jehim thinks that's why they won.

What matters, Jehim said, isn't learning features. It's learning how other marketers and non-developers use AI to solve their work problems. That's why Eungdapara Marketing joined this event: when people share how they use AI and what cases they have, it multiplies. Jehim asked the audience how many of the five winners of a Claude Code hackathon were non-developers. The answer was all five. First place went to a doctor, because the doctor had domain knowledge. For a developer to build a dashboard for doctors or lawyers, they'd need medical or legal knowledge. Doctors and lawyers already have huge amounts of diagnostic data and tacit knowledge from practicing and teaching, so they could solve the problem. In the end, Jehim said, what matters in the AI era is domain knowledge, and the core of it is defining the problem.

These days Jehim talks with AI all day at work and does four things with it. First, define the problem by talking with the team and the people involved. Next, work out how to solve it: when the data gets entered, and what criteria get the first sign-off. Then solve it using AI building blocks like MCP, vibe coding, and skills. Last, ask the AI to turn the whole conversation so far into a skill, and that task is automated. That's how Jehim is clearing work bottlenecks and problems at both the company and Eungdapara Marketing. Jehim uses AI in three categories: vibe coding, automation, and skills. Vibe coding means taking something that used to live in a spreadsheet and turning it into a website, web app, or app. Automation covers recurring tasks that run on their own at a set time. Skills cover repetitive tasks where a given input should always produce the same kind of output.

Jehim showed cases in screenshots. When Jehim ran promotions, stores that signed up got a cashback benefit that had to be paid out three months after the contract. Before, they'd pay the cashback and log it in an Excel spreadsheet. When customer service said an owner wanted to know when their cashback was coming, Jehim would Ctrl+F the store name, copy, paste, and reply in Slack. Vibe coding cut that down. Now you enter the store name and branch, and an Apps Script shows every discount promotion for that store. Referral coupons work the same way. Like Baemin or Coupang Eats (Korean food delivery apps), they create referral coupons per customer. Before, they'd ask a sales rep every time who referred a store, get a call transcript attached, and then approve it. Now you enter a store's info, issue a referral link, and anyone who comes in through that link gets the referral benefit.

Jehim also automated competitor monitoring. Monitoring is one of a marketer's core jobs. What competitors and accounts you take inspiration from did for marketing that week matters. Every Monday, Jehim would go to competitors' Instagram accounts, take screenshots, put them in PowerPoint, and tally the likes and comments. Now Codex crawls competitors' Instagram posts and engagement numbers every Monday and sorts them, and takes snapshots of the company's own social accounts, with the results sent by email into Confluence. It shows how many followers they gained today compared with yesterday. If a marketer's work splits into planning and operations, the goal is to spend less time on operations and focus on planning. Recently Jehim has been using skills to produce SEO content in a consistent format. Store owners are very interested in government support programs for small businesses. They want to know how to apply for things like the high fuel price relief payment or the small business lifestyle and culture innovation program, and how much they can get. Every Monday, or every day, Jehim crawls government sites for newly posted programs, and when the official notice is a Word file, runs a skill.md to keep producing content in the same format. There are limits, though. Image generation in tools like Codex still doesn't meet the standards of the in-house designers, so it's hard to post those images as-is on the company's social accounts.

Jehim wrapped up with a few points. Joking, Jehim said that as an employee, it's important never to brag about the productivity you gained with AI. If you cut a five-hour task to one hour, those four free hours just get filled with other work. Also, cutting a five-hour task down to one doesn't mean the task is gone. It needs maintenance. Support programs change their criteria, and content posted today might not fit a year from now because the logo or criteria changed, so it has to be refreshed. The bragging part was a joke, Jehim said; in reality they brag a lot. What really matters is separating what to hand to AI from what not to, and precisely defining the problem for what you do hand over. For a marketer, running operations well matters, but in the end you have to make an impact through planning. When looking at competitors' social media, the insight ("they make content this way, so let's try this") matters more than making a pretty report.

Jehim also shared a current struggle. If a marketer's work is planning and operations, they've automated operations to some degree with AI, and now want to take on planning. Planning comes down to decisions. Many people think humans should make decisions, but Jehim thinks AI could help. That takes context. AI is great at execution, so Jehim gives it MD files covering why we're doing this, what kind of company we are, and which customers we target, and the question is how to do that well. Jehim is exploring LLM wikis, the second brain built on Obsidian plus an LLM wiki that Karpathy proposed, and Hermes Agent, which is said to improve itself well, all as ways to keep their work from standing still and keep learning.

2. Questions for Jehim

We took questions. One person said they mostly use frontier models through things like Hermes Agent and Codex, and asked whether Jehim had tried running efficient Chinese models or small models locally on something like a Mac mini, and whether Codex was their main tool or they also used Claude. Jehim said they got the 170,000 won (about $120) a month plan and are going all in on Codex, moving as much work as possible from ChatGPT to Codex. Claude seems better at context-heavy work and long text, though, and Jehim is still figuring out how to handle that. For local, they use it as a hosting server. On the team, they're discussing putting a Mac mini in the office to run Hermes Agent, since beyond each person's own second brain, the team needs a shared one.

Another person said AI is great for SEO optimization but seems bad at producing content that's actually valuable to consumers. Jehim agreed. When card news agents (tools that make Korean-style slide carousels for social media) were trending, a marketer could tell at a glance which results were made by AI. AI output doesn't meet consumers' standards, so they can't post it on the company's social accounts. Talking with a designer, Jehim learned it's because the AI doesn't understand the company's design context at all and just produces what millions of AI users ask for. Jehim brought up SEO first for that reason: text is where AI writing shows the least. If you keep refining text with an SEO optimization skill, you get good quality. So they apply it a lot to things like support programs, where images don't matter and information does. As for getting AI content onto the company's own channels, they're working on it with the designers by refining a design MD file, but haven't solved it yet. I asked if they used AI for the presentation slides. No, Jehim said, because if it looked AI-made, it would hurt their pride.

3. From philosophy student to martech founder

Next came the fireside chat with Sungpil Nam of AB180. While preparing, I'd heard his whole life story, and it was full of twists and challenges. We followed it back ten years.

AB180 is a marketing technology company, known for its product Airbridge and for Amplitude and Braze. It provides technology and solutions that maximize marketing performance and help companies grow. It supports about 500 companies in Korea, and since going abroad three years ago, about 300 companies in the US, Vietnam, Turkey, the Middle East, and elsewhere. Lately it has been applying AI in its products and using it heavily in internal work.

Sungpil was a liberal arts student too. He entered the College of Liberal Studies planning to study French philosophy, and meant to go all the way to a PhD. Then he went through a change of heart during his military service and came back wanting to do something more practical. While he was wondering what to do after leaving philosophy behind, an older friend from the same school told him about a Google keyword marketing competition. College students could enter, and Google paid for it: you picked a local small business, ran keyword ads for a month, wrote up the results in a report for Google, and the best teams won prizes. He picked a small business that sold Korean tour products to foreign tourists and ran keyword ads for them. That's how he discovered marketing. He ran Google Translate in five languages to research keywords, set up the campaigns, and improved the landing page based on quality scores. He realized that he, someone who didn't know the first thing about marketing, was enjoying it. The results were good. His team didn't win a prize, but unofficially it came first in Korea.

That's when he got interested in marketing. He joined the company founded by the friend who'd invited him to the Google competition, and started startup life in 2013. AB180 is his third startup and his first as CEO. Over two startups, even without a technical degree, he learned to code and took on a developer role, and 11 years ago he founded the company with co-founders he met at Software Maestro (a Korean government program for young developers). Interest in marketing led him to startups, along the way he learned to code and became a developer, and with an identity as a marketer-developer, he built a marketing technology company.

I asked why he got into marketing when philosophy is so different. It's all connected, he said. What he studied was a Korean field called imagination studies, which looks for patterns in what people imagined throughout history and records and studies them. It was so abstract that, during a hard stretch in the military, he decided that if something this abstract couldn't make his life better, he shouldn't keep doing it, and went the other way, toward something practical. The Google competition was in 2012, just as mobile and performance advertising were taking off. Lock screen apps were appearing, and Facebook CPI was 100 or 200 won (well under 20 cents). He got into the field right out of the military, just as mobile and performance ads were rising.

He looked back on why his team came out on top. It was relentless testing, he said. Performance marketing has a lot of room to improve if you put in the work. Beyond the good brand keywords, how do you find long-tail keywords, and what quality score do they get given the landing page content and CTR? He tested that relentlessly. Fourteen years later he still remembers: the company had a tour of Gwangjang Market (a famous traditional market in Seoul), and he agonized over how to express Gwangjang Market as keywords. You could say food tour, street food, or Korea Gwangjang market. He thought hard about keyword combinations like that and tested their volume and quality scores. It's an approach that still works today.

4. From search engine to Airbridge

At the founding there were four developers and one designer. I asked what it was like. AB180 didn't start as Airbridge, what people call an MMP (mobile measurement partner). It started as a company building a search engine. Back then, search engines only indexed websites, not apps, but a lot of services existed only as apps. So they built a search engine that automatically crawled the content inside apps and turned it into websites that Google or Naver (Korea's biggest search engine) could index. He said marketing has two rivers. The Han River is the one between offline and online. Cheonggyecheon (the stream in downtown Seoul) is the one between the web and mobile apps. The gap between apps and the web was so big that they thought their search engine would solve it. It needed deep linking technology to take people from the web into apps. But when they asked clients, the clients didn't care about search engines or SEO. They wanted deep linking that worked more smoothly and reliably, with tracking. So they pivoted to a deep linking tool, and a year later changed again. Ads were a key part of how users arrived, and clients wanted to see ad ROI, so they switched to an attribution tool. Over two years, working with about 50 clients, they pivoted from search engine to deep link tool, and from deep link tool to today's Airbridge.

The story of their first paying customers stuck with me. When they'd just built Airbridge and had no customers, the first paying ones were Baemin (Baedal Minjok, Korea's biggest food delivery app) and Gmarket, then part of eBay Korea. The service was new and had plenty of gaps, and the final bosses walked in. He compared it to a One Piece story where the Four Emperors show up right at the start. How did it happen? Because he had a developer background and understood a bit of marketing, he gave a lot of lectures early on to get the word out. A director from eBay Korea came to one. Afterward the director invited him to the office, realized he was running a company and not just a lecturer, and said, let's try it. A month later, that director left the job and moved to Jeju Island to raise their kids, but first introduced him to Gmarket. At Baemin, the CMO at the time wanted to work closely with a trusted partner to do bigger marketing, and adopted it. Three years of back-and-forth with those two clients pushed the service a long way.

I asked what that back-and-forth looked like. In mobile marketing you install a module called an SDK in the app. Baemin and eBay are used by the whole country, including grandparents, so there was a huge range of devices. Some people were on Android versions two or three years old. So he went to Yongsan Electronics Market (Seoul's big electronics bazaar) hunting for old phones that never got OS updates. He'd find an old device, see if it turned on, wait three or four minutes to check if it was Android 6.0 or 7.0, then go to the next shop, buying devices to test on. All to make sure the SDK collected data reliably. They also learned the quirks of Korean marketing. Korean commerce sites are good at getting users to sign up, so companies want analysis based on user IDs, not device IDs. They built that kind of localized functionality and stability, and once those two services came on board, they were suddenly tracking 30 million devices, which pushed their big data and machine learning technology forward.

5. Selling someone else's product while building your own

While building Airbridge, AB180 also distributes Braze and Amplitude. I asked why a company that builds its own product does distribution and sales, and how it became partners with such globally known analytics tools.

To be an MMP, you need partnerships with companies like Google, Meta, and TikTok. That's how you receive the touchpoint data on who viewed and clicked an ad, combine it with conversion data, and analyze it. When they started, they didn't know that. They set out confidently to reinvent analytics and attribution, and when Baemin and eBay came on, they found out they couldn't get data from Meta or Google. Without it, the value they could offer was less than half of what clients expected. They had two choices: pivot entirely or keep going.

Until they became partners, Airbridge was hard to sell, so they needed to make money. As late as 2018 and 2019, retention and the full funnel were weak concepts in marketing. The focus was on getting installs, raising DAU, raising MAU. One conglomerate chairman went as far as ordering his team to spend 10 billion won, no questions asked, and bring back 10 million installs. It was all about user acquisition. In 2018 they happened to meet Braze and Amplitude. At the time, the only company in Korea using Braze was Yogiyo (a Korean food delivery app), because its headquarters was Delivery Hero, so it was a foreign-owned company. They saw how well Yogiyo used it, made the connection, and started selling Braze. Once they had 20 clients, Amplitude's headquarters reached out, saying they'd heard about them from a friend in San Francisco, and asked them to sell Amplitude too. Three things lined up. These were services that fit the shift from user acquisition toward retention and repeat purchases. And Korean CRM at the time meant scheduled messages blasted to everyone at 10 a.m. and 2 p.m. with no personalization, so the timing was right for something better.

I asked if reselling other people's products while building your own caused an identity crisis. The developers pushed back, he said. But he believed digital growth can't come from any single element. Marketing has to be done well, messages have to be sent well for customer engagement, and in-product behavior has to be analyzed and improved through A/B testing. He saw these as steps, and as long as they were helping companies grow in a real way through those steps, he could wait.

He also explained why MMP partnerships are hard to get. It's personal data. An ad ID is sensitive information, a kind of pseudonymized data. Meta giving ad IDs to a third-party company is like a bank giving information to a credit bureau. So Meta checked from every angle why this company should have it, what the impact on Asian customers would be, and whether it could really grow, had the technical capability, and had airtight security. It was a harder process than he expected.

6. Why Vietnam and Turkey?

We moved on to going global. Korea is home base, so working hard in Korea makes sense, but at some point growth in Korea alone can't keep the team motivated or meet investors' expectations. That point came about three years ago, and they went global. Instead of the markets everyone goes to, they started with China, Vietnam, and Turkey. China didn't work out. Now they're in Vietnam, Turkey, and, by way of Turkey, Egypt, Saudi Arabia, and the UAE. For the past year they've been focusing on the US. It gives the team new challenges, pushes him too, and brings in revenue, which is a joy. He said there's a lot to learn from marketers abroad. They're really good at using AI and at short-form video. Korea, where a fixed market has a few dominant platforms competing, has lots of know-how on reactivating users and analyzing by user ID. Each side has something to trade.

I asked about best practices abroad. He's a performance guy at heart, he said, and from that angle the hottest technique lately is web-to-app marketing. Korea has had web-to-app, but it's different abroad. In Korea, you send users to a website, they browse, and they hit a CTA button to download the app or go in through a deep link. Abroad, iOS has a bigger share, and iOS app campaigns have gotten less efficient. So marketers run web campaigns on iOS but set the destination to the App Store, then combine data through a conversion API to build the campaign. On the surface it's web-to-app, but in practice it's using web campaigns to get app campaign results. It's becoming close to the new standard for iOS marketing among subscription and consumer apps abroad, but you don't see much of it in Korea.

The choice of Vietnam and Turkey was data-driven too. Korea is strong in some industries and other countries in others, and Vietnam and Turkey are strong in two: casual games, and consumer apps, a group that includes utilities, subscriptions, healthcare, and education. Vietnam is the world's second-largest producer of casual games and Turkey is fourth. Turkey also has many well-known consumer app companies. Korea, by contrast, is known for MMORPGs in games, with strong commerce and finance. So they went to those two countries to sell to casual game studios and subscription apps. It wasn't luck. It was a decision based on data.

7. 90% of his time with AI

I asked how AB180's team, especially marketing, uses AI. Sungpil said they're a company that takes AI seriously. He spends 90% of his own time working with AI, to the point he worries he's addicted. Once AI arrived, they rebuilt all the company's websites with it. airbridge.io was originally built with a builder called Webflow, and they switched it all over to AI-built code. The website for their marketing conference, Modern Growth Stack, was built entirely with AI too. The design team and marketers learned Codex together and built the public-facing sites. Along the way, marketers and BX (brand experience) staff also submitted PRs for the internal back office and shared company agents, and there have been more than 1,000 PRs so far. The second piece is an internal agent called A-Bot. He presented it at an AWS conference and wrote about it in detail on the blog. They connected it to internal knowledge sources and to tools like Slack, GitHub, and Jira, like an LLM wiki, and usage exploded. It keeps gaining features and now handles 400 to 500 queries a day.

There was a lot of trial and error. He once wrote a blog post on what non-developers should watch for when working with AI, and highlighted two points. The first is the principle of clarity. If your requirements or problem aren't clear when you work with AI, it won't do the job well. One way to get clearer is to have the AI interview you. Throw out your first thoughts, then tell it to keep asking you questions until things feel clear. The more you define what "clear" means, the better. When the AI asks questions that sharpen your thinking, quality goes up a lot. The second is that individuals working with skills isn't enough; those skills need to become assets of the organization. If everyone builds their own skills, at some point the synergy hits a wall. Beyond personal productivity, it matters to find what the skills have in common, productize that into assets the organization can trust, and keep improving them.

8. SaaS doesn't die, it splits

I brought up the SaaSpocalypse. How does he see the idea that we're entering an era of AI agents, where software gets built and customized by users themselves and pricing and structure change? Sungpil said it isn't a SaaSpocalypse. As the CEO of a SaaS company it does put him on alert, but he thinks SaaS won't die. It will split. The essence of software is the essence of agents, and that's results. One direction is evolving into agents that deliver results. The other is becoming the infrastructure those agents use, because agents can't do everything themselves. He compared it to a robot kitchen. For a robot to make fried food, it needs a fryer. That's infrastructure. Building the fryer into the robot would be inefficient. Instead, the fryer signals the robot when the frying is done, and the handle a person used to grab becomes a handle that's easy for a robot to grab. That's what CLIs and MCP are, he said. The job is the same as before, but the interface changes to talk to agents in a way that's easy for them. So software has to strengthen its infrastructure side and also deliver results as an agent. Software used by humans may shrink, but software used by agents can grow even more.

I also asked what AI marketing is. It's a hyped term, he said, but one he's thought about a lot, so he gave a condensed answer. He remembers when the term "growth hacking" appeared. Dropbox gave free storage when you invited a friend, people invited each other on their own, and it grew explosively. The essence of growth hacking is producing big output from small input. He thinks people's expectations of AI marketing reflect the same hope: put in a few resources and AI will amplify them into great results. That hope will stay as technology advances, but the gap will be widened by people who learn faster and use AI in more varied and relentless ways. In the early days of Facebook ads, D2C commerce companies like Blankcorp and Mybag dug relentlessly into creatives, copy, and hooks to maximize their technical edge. AI, for now, is a game won by people who learn fast and push it relentlessly. Over time, though, the marginal utility drops. He agrees with Airbnb's Brian Chesky, who said performance marketing is a laser pointed at one spot and brand marketing is a chandelier that lights the whole room. He thinks this is a good era. People who can build brand value from the inside and people who pick up technology fast can both get the most out of it. In ad creatives, campaign structure analysis, SEO, every field, the people who apply AI fast and push it relentlessly will come out ahead.

9. How far has AX come?

We took questions. Someone said they think a lot about AX (AI transformation) and asked how you judge how far your company has adopted AI compared with other teams or other companies. Sungpil said AB180 has internally defined stages. Using GPT or Anthropic's website comes first. The next stage is moving to coding tools like Claude Code or Codex and doing agent-based work. After that, turning everyday workflows into skills. Then, the organization defines what those skills have in common and productizes it into assets. After that, you strengthen those shared assets with harnesses, context engineering, and environment variables, pull in more APIs to orchestrate, or attach more information, so the core assets keep getting stronger. The more that leads to bigger use cases across the organization, the better the AX. He thinks Microsoft CEO Satya Nadella was right to introduce the idea of AI tokens as assets. The difference between companies that do AX well and those that don't, he said, is whether it's just you who's good at it, or whether it becomes an organizational asset that creates value across many more cases.

10. What came out of the roundtables

After the fireside chat, we ran roundtables in groups of six to eight. This is the best part of Claude Bloom. Each group shared names, companies, titles, and why they came, discussed, and then the table lead shared insights.

One table, led by a headhunter at a recruiting firm called AP Search, had people running operations at a game company, a computer company, and an AI company. Their summary: in the AI era, marketing results come from people who define problems well, not people who are good at ads. Marketers are changing from people who execute directly to people who design operations and automation, and AI is a tool, not the goal. Going forward, the marketers who matter will cross the lines between brand, performance, operations, and PM, raise productivity with AI, and bring new perspectives. They agreed the key skill is looking at data and finding insights others miss. The talent that will shine most, they said, is not the person who's good with AI but the person who understands people well. AI helps with data and execution, but human emotion is what drives a purchase. On workflows, people shared using Perplexity to monitor competitors, industry trends, and regulation, using Claude to build up company info and channel guides, and using AI to produce concepts, copy, images, and video. Marketers will become less like content producers and more like strategists and orchestrators.

A table led by someone doing GTM marketing at ESTsoft brought together marketers from startups to large companies. GTM, go-to-market, means creating touchpoints with overseas clients and getting the product embedded in their work. Rather than sticking to the discussion questions, they talked about how marketers at companies of different sizes use AI tools. Interestingly, large companies weren't using AI agents much, and startups were using them more actively. The lead said they felt that using AI well in their own work takes a lot of organizational support.

A table led by someone doing global influencer marketing at Fullseat read out a summary Claude had written for them. Token efficiency came up as the most important skill. Tokens are a measure of how precisely you can instruct AI, so using them efficiently matters. Domain knowledge mattered too. AI gives general answers, so if it doesn't know your revenue structure, database, or target criteria, it produces off-base analysis, and you need domain knowledge to tell whether it's right or wrong. Some said their workload actually grew after adopting AI, which they put down to poorly designed workflows. Separating what AI should do from what people should do, and designing the flow, is the most important skill. Musinsa, Topmint, and healthcare companies were using AI for automated competitor monitoring, automated weekly reports, multilingual copy, and creative planning. What they had in common was treating AI as something you plant in the company: organizing SQL logic, work standards, and history into shared assets so the whole team uses the same AI.

Table 3, led by someone in charge of user growth at Next Securities, had people from a brokerage, consulting, a publisher, and startups. They agreed that AI helps with analysis but the final judgment and planning are theirs. When you keep talking with AI, the conversation often wanders off, so the ideal person keeps sight of where they're heading and pulls everything back onto one canvas. On workflows, they shared a bot at a brokerage that pulls competitors' keywords and promotions and sends alerts, and a setup that turns content into templates and auto-publishes it through social media APIs. The standout was a flow built by someone at the publisher that sends an alert whenever review scores drop and suggests how to respond.

A table led by someone who runs Hitchhiker AI had a Tiro team member in the group, so they looked at the notes Tiro had organized. Even in the AI era, they said, it comes down to the people using it; marketing itself hasn't changed. AI can make data and asset production more efficient, but only people can judge whether it fits people, people decide which metrics matter, and in the end it's people who keep pushing efficiency up. That's where they landed.

The presenter from Woogi Inc., which runs digital billboard ads, said they're a small company, so it isn't full AX, but to use AI well they broke all their work into small pieces and attached AI to everything they could. They think data has become more important, and their table talked about people. AI's answers change depending on people's curiosity and questions, so curiosity and questions matter more now, the original idea you start with matters more, and knowing clearly what you want matters more. Others said that trying fast and failing fast leads to fast growth, and that as consumers they dislike content that obviously reads as AI, so the person who double-checks at the end plays a key role. Workflows included news clipping, content generation, results reports, and marketing proposals. The presenter, who has 20 years in audiobooks, said it used to take a long time to add 20 to 30 national media outlets to a proposal. After about a year of training AI on it, the AI now builds it in 30 minutes without being asked, and they joked they'll probably get fired soon. They also told a story about a colleague who knew a little about a lot. It used to be exhausting covering for that shallow depth, but now that AI fills in the depth, that colleague has become an S-tier performer.

The co-CEO of Mir, an AI for social media marketing, said their table talked about other things and shared a personal view. General-purpose LLMs, they said, don't come anywhere close to real marketing. The order in which general-purpose LLMs take market share is developers, then designers, then marketers. Developers work in programming languages, so non-developers fairly quickly became able to build. But even with Claude Design out, no designer has been fired because of it. AI can't do work you don't understand yourself, so you have to know everything. For now, what AI can do isn't turning small input into big output, but turning big input into better output. You can delegate work you could do yourself, but you can't hand off work you can't do and get it finished. They also called out the illusion that AI is making things more efficient. If it really were, someone would have been laid off because of AI, and companies like that are probably the top 1%. Going forward, people who know things more deeply will shine. Their way of using AI: only give AI work you can do yourself. If it works when you do it together with AI, turn it into a skill. If it works with AI alone, start using it internally.

A table led by a freelancer who makes ad videos for game developers focused on two keywords: making it an organizational asset, and clarity. One idea was that it's good to build a single archive and turn work into processes, so new people can pick it up. They agreed with the point about clarity, having AI interview you. Thinking about why clarity matters, they concluded it's because you can't expect AI to hit a high point on any single task. To get good work from AI, you have to string together decent low points into something meaningful, and that means breaking work down well and having the domain knowledge to carry it through to insight.

A table led by BT, a CEO who sold a game agency called GrowthN and now runs Bailey's Growth, summarized with Tiro notes. The group included a brand marketer for the sparkling water brand Voss, two people from Bungaejangteo (a Korean secondhand marketplace app), and a game company marketer. Three things came up on the first question: instant data analysis, using the Braze MCP integration, and a change in how game companies approach content. Since looking at data has become easy, they store large datasets locally and check them daily instead of running them through MCP, and they connect the Braze MCP to cut the unnecessary 70% and customize messages. Game companies now compete not just with other games but with attention-grabbing Reels, so they've started asking whether they're more fun than that. On workflows, a lot of it was about managing each person's work. They connect tasks and projects to Slack automatically, and BT relies heavily on automated context syncing. At an agency, the CEO gets lots of calls and Slack messages. BT syncs those by voice, the right team lead gets tagged depending on whether it's a prospect, a new client, or an existing one, and the conversation syncs automatically. When a client argues about what was promised, they can respond right away, or catch it if someone is gaming the system. Frustrated that outputs didn't come out as wanted, they're also building a shared tool that produces the company's style by connecting skill sets and Claude. After presenting, BT had to leave, emailed the materials to the team, and asked for a group selfie.

A table led by a startup CEO discussed with marketers what kind of talent will shine in the AI era. They see work that used to be divided by job skills shifting to be organized around goals as AI advances. The Plan, Generate, Review, Inspect setup from software development is spreading to business. In larger organizations, one person now owns a whole piece of work around a goal, so communication skills for dividing up goals well matter. In smaller organizations, AI is good at generating and monitoring, so marketers should focus on planning. They raised an interesting point. Workflows focus on raising productivity, and when productivity rises, what companies care about most is cutting headcount. That means business hasn't caught up with productivity yet. In development, you have a right answer and use harness engineering to get the same result every time for productivity's sake. But if the essence of marketing is being different, marketers should focus on outperforming, not on productivity, and society isn't having enough of that conversation.

One person came up from Busan and runs an offline-focused community called South Bridge. Their table had someone doing US marketing at a webtoon company, a business owner doing AX systems for very small businesses, someone offering that as a platform service, the lead (who plans and markets a blockchain business solo), and a performance marketer. The big takeaway was that this group doesn't use AI at as high a level as you'd think, and the general public doesn't either. Since adoption isn't that advanced, what matters before AX is whether your data is in good shape. The webtoon marketer's story was interesting. When you make content, webtoons have an audience that hates anything that looks like AI, so there are cases where people ask for changes because something looks like AI even though a person actually made it. In content marketing, not touching AI has become an unwritten rule. On the performance side, many people use work automation like Braze and Claude Cowork to easily build skills from context, so Claude is common there. The lead felt that if you use a frontier model without training it, the general public still sees Claude as the go-to. The lead uses a headless agent to build a map of their relationships to set the tone of messages, and to find details missing from LinkedIn drafts, but still ends up writing each post by hand for about an hour and a half. With lots of resources, you'd do machine learning on GPU servers to refine it, but some people change prompts, or use a frontier model's memory almost like machine learning, constantly tweaking prompts until the writing comes out in the context they want. Since many people were first-timers, the lead said sharing how each person actually uses AI was more valuable than talking philosophy or workflows.

Originally written in Korean on Brunch, June 21, 2026.