BIG DATA and Deloitte Taiwan Drive Real-World AI Adoption with a Focus on Human-AI Collaboration and Corporate Governance

To address challenges such as scattered internal information, difficulties in knowledge transfer, and cross-departmental collaboration, BIG DATA independently developed KEYLM, an enterprise knowledge AI agent platform that helps businesses integrate information scattered across documents, databases, and various business systems, building an enterprise knowledge foundation that AI can understand and use. Compared with general-purpose AI applications that rely primarily on public information as their knowledge source, KEYLM places greater emphasis on proprietary enterprise data, business contexts, and permission management. It enables employees to search for information, organize content, and handle work tasks based on internal organizational knowledge, reducing the operational burden caused by fragmented information or knowledge gaps.

On the technology front, BIG DATA continues to deepen its Traditional Chinese semantic understanding capabilities. Through its proprietary domain language model, BIG10, and related semantic technologies, the company strengthens AI's understanding of Taiwan's local terminology, industry-specific terms, and enterprise knowledge, and designs AI application architectures tailored to each enterprise's data security and deployment requirements. KEYLM was also recently recognized with The WITSA Global Award for AI Entrepreneurship at the WITSA Global AI Awards 2026, demonstrating BIG DATA's achievements in enterprise AI research, development, and application.

Yu-Ting Yuan, Director of the Innovation Marketing Department at BIG DATA, said that through AI Workshop diagnostics and interviews with companies across various industries, the team found that most businesses do not lack information or experience. The real problem is that this knowledge is scattered across different documents and systems, and even in the minds of senior employees, making it difficult to search, integrate, and pass on effectively. At the same time, although some companies have already adopted AI tools, they have not clearly defined their use cases, resulting in a gap between the technology and actual work needs.

Yuan further pointed out that enterprises should assess their needs along two dimensions: task frequency and complexity. They should prioritize high-frequency, standardized tasks with easily verifiable outcomes, such as data queries, meeting minutes, and routine reporting, to gradually build adoption experience before expanding to workflows that require cross-system collaboration and professional judgment. BIG DATA has also helped companies in industries such as healthcare and cultural and creative sectors explore operational process optimization, enterprise knowledge management, and digital asset applications, designing AI solutions around actual business needs rather than simply pursuing the adoption of more AI tools.

AI adoption is not merely a technology project for the IT department; it also involves management support for organizational change, employee capability development, and cross-departmental collaboration. Enterprises should first identify use cases that offer business value and are easy to execute, build early results and internal consensus, and then gradually expand the scope of application, preventing AI projects from stalling at the proof-of-concept stage due to a lack of governance frameworks and organizational readiness. Tomm Chiang, Executive Vice President of Deloitte Taiwan, stated that with the rapid development of Agentic AI, enterprises are entering a new operating model in which humans and AI agents work together. Future competitive advantage will depend not only on AI technology capabilities, but also on whether enterprises can reshape management mechanisms that balance innovation, governance, and trust. Based on his consulting experience, he noted that enterprises advancing AI applications should simultaneously strengthen their governance frameworks, covering strategic objectives, division of roles and responsibilities, risk assessment, validity monitoring, decision traceability, and regulatory compliance, while using human-AI collaboration mechanisms to ensure that critical decisions retain appropriate human oversight and professional judgment.

Shu-Kai S. Fan, Dean of the College of Management at National Taipei University of Technology, explained from the perspective of AI agent quality engineering that AI development has entered its "midgame," where the focus is on bringing AI off the screen and connecting it with real enterprise systems. He recommended that enterprises evaluate deployment by returning to five first principles: data, mathematical relationships, learning loops, generalization capability, and computing architecture. He also suggested restructuring existing quality management mechanisms such as PDCA, DMAIC, and error-proofing (poka-yoke) as the foundation of AI agent workflows. On the technical side, supervised fine-tuning (SFT) can handle core processes, retrieval-augmented generation (RAG) can support dynamic knowledge, and Function Calling can connect existing systems such as MES and ERP. On the governance side, LLMs, agents, monitoring agents, and humans can be combined to form a "four-eyes" oversight architecture.

In addition, an expert representative from NVIDIA shared AI Agent technology trends at the seminar, outlining the direction of intelligent agents in cross-system task execution, computing performance, and cost management, as well as enterprises' requirements for execution environment isolation, access control, and information security when adopting autonomous AI systems.

As AI Agents enter core enterprise processes, the competitive focus will shift from model capabilities to the use of organizational knowledge and the maturity of governance mechanisms. Moving forward, BIG DATA will continue to deepen KEYLM's enterprise AI applications and work with industry partners to help businesses build AI capabilities suited to their own operational contexts.