Abstract: In response to national priorities and the frontiers of academic research, the development of digital and intelligent commerce in the era of large models depends not only on breakthroughs in key technologies such as artificial intelligence and their applications in the industry but also on the reengineering of traditional business processes and innovations in management theories. Deeply integrating data intelligence with business scenarios, overcoming core technological bottlenecks, fostering new industries and models, and achieving a new paradigm of intelligent, efficient, and collaborative digital and intelligent commerce represent crucial future directions. Based on the 406th “Shuang Qing Forum”, focusing on the theories and methods of digital intelligence business in the era of large language models, this study delves into three aspects: key technologies and methods, intelligent collaboration and autonomous decision-making, and new models and forms. It further proposes a research agenda addressing advancements in AI methods and large model technologies, multi-agent decision-making and cross-agent collaboration, and human-AI integration in process-organization-market reengineering, along with funding recommendations prioritizing interdisciplinary research in management science and computational science for digital and intelligent commerce.
Introduction: The "Last Mile" of Digital Intelligence
While general-purpose Large Language Models (LLMs) have shown incredible potential, their direct application in business is hindered by a "last mile" problem. Businesses face heterogeneous data, complex logic, high-risk environments, and the need for strict compliance—areas where standard LLMs often stumble. This paper proposes a theoretical framework that integrates management science, computer science, and economics to move from simple "digitization" to "digital intelligence" (数智商务)—a state where autonomous agents and humans co-create value within a redesigned market ecosystem.
A Comprehensive Research Framework
As a synthesis of the 406th "Shuang Qing Forum," this research outlines a high-level roadmap categorized into three pillars: key technologies, intelligent collaboration, and new business models. This framework addresses the entire lifecycle from identifying needs and challenges to implementing governance and evaluation across various organizational layers.
Key Technologies and Engineering Strategies
The path forward focuses on "Model Deployment → Domain Enhancement → Risk Control." This involves bridging the gap between raw AI capability and practical business value through specific engineering strategies.
As illustrated in Figure 2, these strategies include:
- Model Structure & Parameters: Utilizing techniques like low-rank adaptation (LoRA), knowledge distillation, and parameter-efficient fine-tuning to adapt general models to specific business domains.
- Inference Optimization: Implementing quantization (e.g., INT8/INT4) and batch scheduling to ensure efficient deployment in resource-constrained environments.
- Domain Knowledge Integration: Integrating structured industry rules with unstructured data to overcome the probabilistic nature of LLMs and ensure rigorous business logic.
Key Insights and Organizational Impact
- The Automation-Augmentation Paradox: Automating tasks with AI provides immediate cost savings but can erode "organizational learning." Conversely, augmenting humans preserves learning but requires higher training costs. Finding a "dynamic balance" is a key research priority.
- Performance Polarization: AI acts as a "skill-leveler" in creative tasks, helping low-skill workers more. However, in complex strategic tasks, high-skill workers who know how to "steer" the AI see significantly higher gains, potentially widening the performance gap within organizations.
- The "Double-Edged Sword" of Anthropomorphism: Giving AI a human-like persona can increase customer retention by 20%, but it also makes customers more price-sensitive as they perceive the AI as a negotiable entity rather than a rigid system.
- From Two-Sided to Triadic Markets: Traditional platform theory focuses on two sides: buyers and sellers. We are entering a "Human-Machine-Platform" era, requiring new economic theories to handle value distribution when an AI agent performs the work or makes the purchase.
Key Results and Evidence
- Productivity Gains: Generative AI has been shown to reduce professional writing task time by 40% while increasing output quality by 18%.
- Financial Inclusion: AI-driven credit scoring using "weak signals" (e.g., mobile usage patterns) can accurately assess risk for the "unbanked," breaking the cycle of needing a credit history to get credit.
- The Skill Decay Warning: In medical diagnostic studies, while AI advice improved accuracy by 19%, doctors who relied solely on AI without performing their own initial analysis saw their independent diagnostic skills drop by 31% over six months.
Limitations and Future Work
The authors acknowledge several bottlenecks defining the next 5–10 years of research:
- The "Black Box" Logic Gap: Business requires rigorous logic, but LLMs are probabilistic. Bridging the gap between neural networks and symbolic industry rules is a major technical hurdle.
- Trust and Resistance: Employee resistance to "algorithmic management" is a growing issue. Future work must focus on "algorithmic co-governance" to ensure AI is seen as a partner rather than a digital overseer.
- Governance and Ethics: Moving beyond vague ethical principles toward enforceable technical standards for data privacy (e.g., federated learning) and copyright protection.
- Cross-Disciplinary Funding: The paper concludes with a call for the NSFC to prioritize interdisciplinary grants that force computer scientists and management researchers to work together on Digital Intelligence problems.
Summary updated