Could you launch an actual service with a company where the developer, the travel editor and the marketer are all AI? That question produced an AI travel service. Planning, development, deployment, social login and SEO setup were finished in one week, and it's now a live service receiving real users from Google and Naver search. It's a clean look at where an AI agent organization works and where it stops.
An organizational form where AI agents perform the core functions of running the business — development, content production, marketing.
It's different in kind from automating tasks people used to do. The org is designed around AI agents from the start, with humans concentrating on direction and quality control.
AI editors each carry a persona and publish a travel plan and article once a week. One editor writes as a 28-year-old curator living in Kyoto, producing itineraries for small Japanese cities. Images, route construction and article text are all AI output.
Users log in with a Kakao account to see the detailed plan, and a "take this plan" button connects to booking through a travel commerce partner.
An AI company is still a company, so it needs an org chart. A CEO agent orchestrates at the top with four C-levels beneath.
The CBO function was, honestly, a failure. Neither the logo nor the color work was satisfactory. Running design through one particular model produced steadily worse visual quality, and the takeaway was that requesting it from a different model from the start would have gone differently.
Planning, design, development, deployment, Git integration, social login and SEO setup all completed in a week. The AI wrote the spec, generated the code, deployed, wired up Kakao social login and registered with Google Search Console. Humans steered direction and requested design revisions.
The original business model was a paid subscription to travel content. Stepping back, the question became how many people would pay monthly for AI-generated content. The value of travel content comes from who recommends it, and trusting an AI editor's recommendation enough to subscribe is still a long way off.
So it pivoted to a referral model: affiliate links to flights, lodging and experiences embedded in the articles. Content stays free and revenue appears when a booking happens. No revenue yet, but the entry barrier is clearly lower than a subscription.
The biggest lesson was model selection. The model strongest at writing code and planning articles was not the one better at image generation and product data collection. The branding failure traced to the same cause.
No single AI is good at everything. Splitting models by role was what raised quality.
The AI produced the basic UI structure in one pass. But design colors and fonts required explicit revision requests. The itinerary detail page took four or five rounds; a single booking button took two or three.
UX judgment, business model decisions, brand direction — the AI couldn't settle any of the three alone.
For a solo operator or side project, this structure can build a product and run its marketing. But turning it into revenue requires handling marketing, SEO and business model design alongside development.
The conclusion — that AI can execute while humans set direction — matches the four points where AI waits on humans. The faster execution gets, the more clearly judgment becomes the bottleneck. And as AI starts retaining work context, the lock-in that memory creates becomes another variable worth watching.
Automation replaces tasks people used to do. An AI company designs the organization around AI agents from the outset, with humans focused on direction and quality control.
Planning, design, development, deployment, Git integration, Kakao social login and Google Search Console registration. Humans steered direction and requested design revisions.
Branding and model selection. The best model for code and planning wasn't the best for image generation, so roles had to be split — and UX judgment, business model decisions and brand direction couldn't be settled by AI alone.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
Waiting at the booth rarely works at B2B trade shows. How to control what you reveal, book meetings in advance, stop passersby cheaply and follow up the same night.
Can employees take prompt libraries built at work when they leave? How U.S. work-for-hire and trade secret law apply to AI prompt ownership, and what firms should do.
Four teams named four different core customers, all based on gut feel. A step-by-step way to define your core customer from repeat behavior and map your beachhead.