MetaCon: AI Builders Face Rising User Acquisition Costs in Attention Economy

As AI tools make it easier than ever to build new products, the harder problem is increasingly finding users—and keeping their attention long enough to prove value. That was the central message at MetaCon 2026 in Seoul, where speakers argued that in an era of infinite content and near-zero production costs, 'attention' has become the scarcest resource.

Lee Joo-hyung, a manager at Infograb, delivered a session titled “How do AI-built products reach users?” on Friday UTC (July 3–4 in Seoul), during MetaCon 2026, a major South Korean AI conference hosted by TV Chosun with Tokenpost as co-organizer. Lee framed the challenge in blunt terms: the democratization of product creation does not translate into democratized distribution. “As more people can build, the cost of earning attention rises,” he said, pointing to the growing gap between building quickly and validating demand in the real world.

Lee anchored his talk in the concept of the 'attention economy'—the idea that when information is abundant, human attention becomes the limiting factor. In practical terms, he argued that successful products in the AI era require two parallel disciplines: defining a user problem precisely enough to create meaningful value, and designing an effective way to 'hook' users amid relentless noise.

To operationalize that thesis, Lee introduced an approach Infograb calls 'EAM (Engineering as Marketing)'. Rather than leading with promotional content, EAM treats engineering output itself as the marketing surface: a team releases a single, narrowly scoped feature as a standalone free tool, then measures real user interest and feedback before committing to a broader product roadmap.

The advantage, Lee argued, is that the method proves capability through working software rather than claims. By splitting the product lifecycle into smaller iterations—weeks or even days rather than months—teams can observe market pull quickly, discard weak ideas early, and double down on the ones that generate strong organic traction. Users, meanwhile, receive a useful free asset, creating what Lee described as a mutually beneficial exchange.

Infograb runs this pipeline through a virtual agent it calls “Gary Agent,” which is designed to identify and validate potential product ideas from within the organization’s own workflow. According to Lee, the agent analyzes internal context stored across Notion-based wikis, Slack conversations, customer feedback, and code repositories to surface recurring friction points, high-leverage AI skills already used internally, and unmet customer requests that could be turned into tools.

Before any build proceeds, the team applies a structured validation process inspired by Y Combinator-style office hours. Lee said an idea must clear four questions: whether users truly experience the pain point, how they currently solve it, what the smallest viable unit of functionality would be, and whether there is supporting market signal or data. If the team cannot answer all four, the concept does not advance.

As a case study, Lee pointed to a tool built to diagnose issues inside an internal automation system, M8N. The team observed common operational problems—unused or abandoned nodes, insufficient error handling, and workflows that continued running even after their original creators had left the company. Using Gary Agent, Infograb produced an “M8N workflow complexity analyzer” and released it publicly, where it drew strong user response. That validation, Lee said, helped justify development of a larger commercial product called “Napper.”

Beyond product validation, Lee also described how Infograb uses agent-based workflows to scale marketing production—particularly AI-generated brand video. The company operates pipelines for automatically producing product demos and AI-based user-generated-content-style testimonials, organized as a multi-agent “team” rather than a single general agent. Lee argued this structure matters because video work requires heavy context management and multiple specialized tasks; concentrating everything in one agent can degrade both quality and speed.

In Infograb’s setup, a “CEO agent” coordinates task allocation, maintains shared context, and performs quality checks, while final approval remains with a human—an approach commonly referred to as 'human in the loop'. Lee said the pipeline integrates multiple tools for search, generation, editing, and rendering, including Higgsfield for video generation, Gemini TTS for voice, Playwright for automation, and FFmpeg for post-processing.

To improve output quality, the team pre-defines references for characters and environments, analyzes product strengths and market pain points to generate storyboards and copy, and—rather than relying on script-only prompts—creates scene-by-scene still images first and feeds them into the video model as additional guidance.

Lee concluded that shipping AI products is increasingly less about raw building speed and more about proving relevance in the marketplace. Whether or not teams adopt complex agent systems, he urged builders to test demand by extracting a small capability into a free tool, learning from real usage, and competing directly for user attention—the decisive currency of the AI era.

MetaCon 2026 was held July 3–4 in Seoul at the COEX Grand Ballroom under the theme “AI Makers Rise,” bringing together speakers across technology, business innovation, marketing, and investment to examine how AI is reshaping industries and day-to-day work.