Summary of Key Points
This news report focuses on a closed-door meeting organized by Huxiu Think Tank in Suzhou, which discussed the implementation of AI in the manufacturing industry. Practitioners from over a dozen companies, including GCL, Zhuimi, Inovance, and Laiyifen, shared real-world cases of how AI is being applied in production, research and development (R&D), and supply chain management. However, the meeting highlighted the key challenges faced in implementing AI: it's not that the technology itself is inadequate, but rather that "infrastructure" issues such as data, systems, and organizational structures are holding back progress. The approach to the meeting shifted from simply showcasing successful cases to collectively evaluating potential solutions and identifying pitfalls. Three major bottlenecks that prevent AI from being integrated into core production processes were identified, along with strategies proposed by the companies to overcome these barriers.
I. AI is already in use in manufacturing, but mostly in peripheral areas
Although AI has been applied in various manufacturing scenarios, it has not yet penetrated deeply into core production processes. For example:
- Supply Chain: Zhuimi Technology uses AI for end-to-end supply chain management, with successful sales forecasting and intelligent scheduling. However, key aspects like product quality control are hindered by a lack of negative samples (too few instances of defective products) and the inability to measure critical parameters.
- R&D: Using AI for coding has significantly improved efficiency, but integrating R&D decisions with production processes (such as using design parameters directly in manufacturing) remains a challenge.
- Fast Moving Consumer Goods (FMCG)/Retail: Laiyifen uses AI to inspect store displays, and Shanghai Jahua uses it for skin analysis to recommend products to consumers. However, the vast number of FMCG products and fragmented distribution channels limit the accuracy of sales forecasts.
- Energy: GCL uses AI to predict electricity price trends, but factors like competitive pricing strategies in the power market and unexpected events (e.g., weather affecting power generation) make predictions difficult.
In short, while AI can assist with auxiliary tasks, it has not yet reached the core of production processes.
II. Three major barriers prevent AI from being integrated into production
The meeting identified three fundamental obstacles that hinder the adoption of AI in manufacturing:
1. Uncertainty in key data: Industrial settings do not provide easy access to user data like in the internet environment, making it difficult to measure critical parameters accurately. For instance, sensors can only provide surface-level readings for hot workpieces and catalysts, rendering models ineffective without comprehensive data.
2. Lack of confidence in AI outcomes: Although AI can automate tasks like coding and forecasting, companies are hesitant to rely on its results directly in production. For example, code generated by AI must be manually verified, and process optimization suggestions may not be implemented due to concerns about potential issues.
3. Data security and system compatibility: Integrating AI with internal systems requires addressing data security concerns. There are also compatibility issues between large-scale AI models and existing enterprise systems, creating significant barriers to implementation.
III. Solutions: Addressing technical and organizational challenges
Companies proposed the following approaches to overcome these obstacles:
- Technical solutions: Reconsidering approaches to break down the "black box" of complex systems. For example, Inovance is moving away from pure simulation and using causal inference to understand the underlying relationships between data, while Zhixingyi advocates for making industrial systems more transparent (i.e., "white-boxing") to reduce uncertainty.
- Organizational changes: The implementation of AI requires organizational adjustments. For instance, Pake New Materials has automated 70% of its product design processes, but issues with measuring manufacturing parameters have led to a disconnect between design and execution. Adjusting organizational structures to align technology and production departments is necessary.
- Gradual progress: Starting with small, achievable goals and gradually expanding the scope of AI applications can help avoid failure. A leading new energy company used this approach by first solving a specific problem (e.g., improving process quality) before scaling up.
IV. The meeting shifted from showcasing successes to jointly evaluating challenges
Unlike previous meetings where companies focused on their achievements, this closed-door meeting focused on identifying and discussing problems:
- Highlighting difficulties: Companies directly shared their current challenges in implementing AI, such as poor data management and security concerns.
- Cross-disciplinary collaboration: Different companies exchanged insights on how to address these issues. For example, Zhuimi's supply chain challenges were compared with Inovance's causal inference methods, and GCL’s energy-related issues were discussed in context with other companies’ experiences.
- Clarifying boundaries: The meeting helped establish clear understandings of when AI works well and when it fails, such as the limitations of sales forecasting due to the high number of FMCG products.
Huxiu Think Tank plans to hold more offline meetings and online discussions to continue exploring topics related to manufacturing supply chains, marketing, and operational automation, with the aim of sharing actionable insights among companies.
In summary, the success of AI in manufacturing depends not just on the availability of technology but on the solidity of foundational elements such as data quality, system transparency, and organizational readiness. The value of this meeting lies in bringing these critical issues to the fore for collective solutions.