OpenAI Korea Chief: Enterprise AI Adoption Hinges on Data Integration, Not Just Models

Kim Kyung-hoon stood at the Samsung SDS REAL Summit 2026 in Seoul's Gangnam district on a September morning and offered a metaphor that cut to the heart of why most companies are failing at artificial intelligence. Hiring a smart employee and giving them no information, he said, is exactly what happens when a business deploys an AI model without connecting it to email, documents, or work systems. Kyung-hoon, OpenAI's Korea country lead, was making a case that would reshape how enterprises think about their AI investments: the model itself is not the bottleneck. Integration is.

The gap between companies that have figured this out and those that haven't is now staggering. According to an enterprise AI usage analysis OpenAI released in August, the top 10% of companies by AI usage generated 8.3 times more output tokens per active user than median companies as of June 2026. Five months earlier, in January, that gap stood at 2.6 times. In manufacturing alone, the disparity reached 5.3 times. The widening chasm reflects a single, measurable difference: leading companies connect their AI to internal data. At top-tier firms, 21% of weekly AI users utilized plugins—features that let AI access internal documents, customer information, and work tools. At median companies, that figure was 9%. Within OpenAI itself, 95% of employees use plugins weekly.

Kyung-hoon drew a parallel to the industrial revolution. A century ago, factories swapped steam engines for electric power but kept their machine layouts and work sequences unchanged. Productivity barely budged. Only when factory owners installed motors on individual machines and redesigned production flows from the ground up did electricity's transformative power become real. "If you simply bolt AI onto existing workflows, you can achieve small efficiency gains across various tasks, but it's hard to translate that into enterprise-wide productivity change," he said. "Companies need to pick their most important work and redesign it from scratch with AI."

The nature of enterprise AI work is shifting rapidly. The question-and-answer chatbot era is giving way to autonomous agents that handle multi-step tasks. Codex, OpenAI's coding agent, accounted for 64% of all output tokens generated by enterprise customers as of June 2026. But Codex has moved far beyond software development. Since February, weekly Codex users among enterprise customers grew 108 times in legal work, 41 times in sales, 41 times in recruiting, and 26 times in marketing. The pattern is consistent: AI is now performing research, analysis, document creation, and system verification in sequence—work that once required human hands and hours. Cisco, the world's largest networking equipment maker, deployed personalized AI agents called "MyAgent" to roughly 90,000 employees worldwide, finishing rollout in late August. The agent goes beyond chatbot Q&A. It accesses internal information and work tools available to each employee, handling research, data analysis, report writing, and email management—essentially covering most desk work responsibilities. When Cisco applied Codex to its large-scale software development environment, the company raised defect-fix throughput by 10 to 15 times. By analyzing build processes across multiple code repositories, Codex cut build times by about 20% and saved more than 1,500 hours of engineering time per month.

Yet the promise of AI agents has collided with hard reality. Meta pursued a secret reorganization called "Project OT" earlier this year, reviewing scenarios that would cut some teams by up to 60%. The company cut 10% of its workforce in May and planned additional layoffs in November before scrapping the plan. According to internal Meta documents obtained by Reuters, code changes resulting from AI agent adoption surged 220% year-over-year as of June. But new services or improved features actually delivered to users rose only 36%. During the same period, major technical and security incidents—including service outages and data leaks—increased 40%, and the time employees spent on incident remediation jumped 70%. The gap between raw AI output and actual business value proved cavernous.

Mark Patterson, Cisco's chief financial officer, noted that AI agents are already drafting 80 to 90% of financial disclosure documents. But as the scope of AI-assigned work expands, side effects mount. An AI industry insider observed that Meta's case "paradoxically demonstrated that the more you increase AI agent utilization, the more skilled humans you need to evaluate the results." The security risks are real. In the chatbot era, a wrong answer was harmless if the employee ignored it. In the agent era, AI can make mistakes like unauthorized access to customer information or leaking security files externally. Experts increasingly advocate for assigning separate identities to AI agents, creating unique accounts for each one, designating a human manager responsible for its work, and granting access to only the minimum data needed. High-risk actions—fund transfers, purchases, file deletion, external transmission—should require human approval. The entire work process should be logged to enable clear cause identification and accountability if something goes wrong.

Kyung-hoon's closing argument echoed his opening: "Connect AI more deeply to your organization's data and systems, confidently delegate real work, and turn the results into business performance. Just as factory owners a century ago decided what kind of factory to build, now it's time for companies to decide how they will use AI." The companies winning at AI adoption are not those with the best models. They are those willing to rebuild their workflows around what AI can do—and those disciplined enough to keep humans in the loop where it matters most.