Yang Soo-yeol, Chief AI Officer at South Korean AI Data Company CrowdWorks: Pilot Failures Often Stem from Vector Database Data Gaps, Not Agent or Routing Issues

In a series report on enterprise AI agent platforms published by Byline Network (South Korea) on October 23, 2026, Yang Soo-yeol, Chief AI Officer at South Korean AI data company CrowdWorks, shared a specific debugging finding: when teams troubleshoot the reasons for pilot project failures, the issue is rarely with the agents or routing, but rather that the vector database simply lacks the data required to complete the answer. He stated, "When debugging the reasons for pilot test failures, the problem is more likely not with the agents or routing, but with the vector database lacking the actual needed data, leading to an inability to provide answers."

He also emphasized that both structured and unstructured data should be uniformly converted into machine-readable formats through a single pipeline before entering the vector database; if governance and permissions are handled entirely at the agent layer, development complexity increases. Therefore, permission information should be pre-embedded when data enters the vector database or search engine.

He further pointed out that generating answers through multiple requests during agent orchestration consumes significant tokens. To reduce token consumption at the source, data engineering must come first: reducing garbage data and establishing cleaned data portals so that the inputs for search and reasoning are inherently smaller.