ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News3 min read·2026-10-20

ThinkingAI Korea Head Seo Gwang-deok: AI Moves Fast but Business Doesn't—Bottleneck Is Operations, Not Models

Seo Gwang-deok outlined three bottlenecks for enterprise AI adoption at TAS 2026—development, sustainability and monetization—arguing that data definitions and taxonomies should precede model selection, otherwise every model swap means repeated development work.

ImportanceMaterialEvidenceE3 inspectableWrite-upStandard

2026-10-20 13:04 GMT+8

At the ThinkingAI Summit 2026 (TAS 2026) held on October 20, 2026 in Seoul, Seo Gwang-deok, head of ThinkingAI's Korea branch, addressed a widespread contradiction in enterprise AI: individual productivity has risen sharply thanks to AI, but company profits have not kept pace. He cited McKinsey survey data showing that eight out of ten respondents reported improved personal productivity from AI, and one-third of companies said they stopped buying software because agentic coding tools let them build in-house—yet only 37% of firms said AI contributed to operating profit. A 2025 report from the Massachusetts Institute of Technology in the United States found that 95% of generative AI pilot projects produced no measurable profit-and-loss impact, while Gartner projected that more than 40% of agentic AI projects would be discontinued by the end of 2027.

Seo attributed these gaps to three bottlenecks: development, sustainability and monetization. On the development side, he offered an iceberg analogy—demos represent only 10% of the work, while operations account for the remaining 90%, and AI cannot directly produce that 90%. Specifically, between prototype and production lie data definitions, access permissions, integration with existing systems and accountability when errors occur. On sustainability, every time a new AI model launches or a company's KPIs shift, existing systems require modification; if the meaning of data and business context is scattered across individual code files and prompts, swapping models means repeating development from scratch. Monetization demands clear ownership of AI projects, performance measurement and a system that feeds results back into continuous improvement.

Sign in / Sign up
ReadingResearchRadarInvestment framework

His core recommendation is to design data and taxonomies first, then select models. He illustrated this with a hypothetical scenario: a game service sees its daily active users drop 12%. In the traditional workflow, the responsible person reads a report, requests analysis from the data team, and the operations team then responds. In the agentic workflow, AI automatically analyses the churn cause and proposes a response plan; once a human approves, the agent carries out follow-up actions. The contrast is meant to show that the key is not automating individual tasks but connecting problem detection, judgment, execution and result verification into a closed loop.

Seo named this approach Agentic Growth and introduced ThinkingAI's Agent Engine as its implementation foundation, comprising four layers—data infrastructure, omnichannel sensing, agent execution and business applications—designed to connect user behaviour analysis, CRM and A/B testing for enterprise workflow automation. It should be noted that ThinkingAI is itself a company that provides AI tools, so the solutions presented in the keynote directly relate to its own products. The DAU example is a hypothetical illustration rather than a verified deployment. Nevertheless, the principle that models change while data and taxonomies endure, together with the shift from point tools to a business closed loop, offers actionable diagnostic value for enterprises working to make AI land in operations.

Sources

  • 原始出处zdnet.co.kr↗
Continue researching →