Advancing AI Integration for Real-World Applications

With the rapid evolution of artificial intelligence technology, generative AI has achieved significant breakthroughs in model capabilities. The current focus is no longer on whether computational power, models, or data are sufficient, but rather on whether real needs exist and if the efficiency gains from AI can outweigh its costs and risks.

AI is transitioning from being technically feasible to being value feasible, consistently creating incremental value in real scenarios and transforming into a new productive force. Meanwhile, AI continues to advance, with ongoing progress in multimodal systems, world models, and embodied intelligence, supporting mid- to long-term capability evolution. AI must shift from being an auxiliary tool to becoming the main agent for task execution and result delivery, at the core of change.

From an implementation perspective, domestic AI applications show a trend of “C-end first, B-end gradual”. Applications targeting individual users are growing rapidly, leveraging a mature mobile internet ecosystem, with leading products achieving monthly active user numbers in the hundreds of millions. In enterprise scenarios, large-scale implementation will not happen overnight; it will undergo repeated validation and iteration. As model capabilities and tool utilization improve, AI agents are evolving from merely “answering questions” to “completing tasks,” gradually acquiring the ability to execute complex tasks across multiple steps and systems. Significant progress has been made in long-term tasks, tool utilization, planning, and error correction, with attempts at large-scale deployment in controllable scenarios such as e-commerce, customer service marketing, and content production. As foundational model capabilities cross application thresholds, inference efficiency is becoming the focal point of computational power competition. With conversational AI, code generation, and image/video generation applications entering large-scale deployment stages, the volume of inference calls is growing exponentially; the multi-step reasoning characteristics of AI agents further amplify the demand for computational power on the inference side. This shift means that chip dynamics, cloud service pricing, and enterprise procurement logic will revolve around inference efficiency, transitioning from “which model is the strongest” to “which offers the best inference cost-performance ratio”.

Currently, challenges remain in advancing the implementation of “AI +”. The core bottleneck for large-scale AI deployment is shifting from the supply side to the demand side. Most enterprises experimenting with AI agents have only a small fraction achieving large-scale deployment. The reasons are not due to insufficient technology, but rather unclear goals, limited integration readiness, and difficulties in proving commercial value. In many enterprises and scenarios, the efficiency gains from AI are insufficient to cover overall costs. AI is taking on more functions in business, but its reliability, interpretability, and accountability mechanisms are still immature, posing significant institutional constraints on large-scale deployment. For instance, edge AI disrupts traditional privacy protection and data security orders, while AI agents’ mixed data usage exacerbates data security risks. A singular business model and low-price competition hinder sustainable development. The deployment of enterprise-level AI faces structural contradictions between customization and scalability. Currently, about 70% of AI solutions require customization, while only 30% can be standardized, with deep business integrations varying across enterprises. In the short term, enterprise-level AI deployment primarily relies on API calls and customized services, with many projects depending on custom development, making it difficult to form replicable product capabilities.

Looking ahead, to comprehensively advance “AI +” and expand the breadth and depth of applications, the following recommendations are worth noting:

1. Establish Effectiveness Evaluation and Dynamic Screening Mechanisms for AI Applications: The 2026 Government Work Report emphasizes creating a new form of intelligent economy, deepening the expansion of “AI +” and promoting large-scale commercial applications in key industries. Establishing measurable effectiveness as the core metric for evaluating AI applications is crucial. In public service sectors like government, healthcare, and urban management, exploring the establishment of industry-specific evaluation frameworks for AI technology applications is essential, setting quantifiable indicators around efficiency improvements, cost savings, and service quality enhancements. Given the rapid iteration of AI technology, solutions deployed six months ago may already be surpassed by better alternatives, making regular reviews and dynamic screenings of demonstration projects particularly important. Initial trials can be conducted in government services, establishing a minimal viable evaluation framework centered on efficiency and public feedback, gradually expanding to other fields. Proven scenarios that generate incremental value should be promoted, while resources should be reallocated for scenarios with unclear cost-benefit ratios or difficult-to-measure outcomes, moving away from the practice of completing projects and immediately closing them.

2. Cultivate an Industry-Level Middleware Ecosystem: Encourage leading enterprises to promote industry experience from customized projects into reusable middleware products, following a path from customization to modularization and then to platformization. This can draw on established practices in the internet industry, transitioning from business middle platforms to capability APIs and then to open platform ecosystems. The construction of AI middleware in traditional industries can leverage experiences in capability encapsulation and platform operations, gradually transforming one-time project investments into sustainable reusable product capabilities. This also addresses the current challenge of AI business models converging on low-cost APIs. In procurement mechanisms, explore shortening the bidding cycles for AI projects, allowing phased acceptance and iterative delivery to adapt to the rapid pace of technological evolution.

3. Improve Governance Rules Covering the Entire AI Application Chain: As AI applications evolve from single model calls to complex collaborations across multiple steps and systems, the responsibility chain lengthens, necessitating an extension of the existing institutional framework. In terms of accountability, it is advisable to explore establishing a tiered responsibility framework covering all aspects, including model provision, application orchestration, and terminal services, clarifying boundaries regarding data security and output quality. This will promote traceability and auditability throughout the AI application process. In safety management, for critical decision-making processes involving citizen rights and public safety, the “Human-in-the-loop” principle should be upheld, ensuring that ultimate decision-making authority remains with humans. In terms of institutional integration, leverage existing legal frameworks to address privacy protection and data security challenges posed by AI, promoting deep integration of AI tools with current compliance and risk control systems in high-risk areas such as finance, healthcare, and law.

4. Activate Demand-Side through Procurement and Subsidies: To promote large-scale deployment of AI applications, efforts must be made from both the supply and demand sides. The 2026 Government Work Report proposes measures such as “supporting the construction of open-source AI communities” and “supporting public cloud development,” clearly signaling the intention to lower barriers to AI application on the supply side. Local government practices indicate a shift from early broad subsidy models to precise support combining specified platforms and vertical scenarios, concentrating limited resources on directions with clear application scenarios and verifiable effects. In line with this trend, further institutional arrangements should be made at the procurement level. In public service, urban management, and public health sectors, pilot a composite pricing model that combines basic service fees with performance-based payments, using large-scale procurement to drive the industry towards sustainable pricing and delivery standards. Additionally, guide governments and enterprises to increase budget proportions for cloud computing, software, and service subscriptions in AI project procurements, shifting from supply-side support to demand stimulation, fostering a robust commercial ecosystem through real orders. On the individual usage side, the paid subscription of AI tools is forming new capability thresholds, suggesting the exploration of appropriate consumer subsidies for individuals purchasing AI productivity and learning tools.