Summary of Key Points
The rate of AI-generated code has reached 60% (as seen with Kuaishou), but the overall R&D efficiency of companies has not improved accordingly. The problem lies not in the AI's coding capabilities, but in the fact that individual efficiency improvements do not equate to organizational efficiency improvements. This is due to traditional production and research models, the loss of collaborative context, and insufficient knowledge accumulation. At the same time, AI tools are evolving from being mere assistants (such as Copilot) to supporting team collaboration (Agent Teams), with non-R&D personnel (from operations to product management)也开始 using AI to develop software. In the future, AI will transform the entire corporate structure, but human judgment and ability to understand user needs will remain core competencies.
1. AI Makes Individuals Faster, but Organizations Are Not Keeping Up: Three Major Reasons for Efficiency Stagnation
Many companies have found that while engineers can write code 40% faster with AI, project cycles have not shortened, and overall team output has not increased. The issues stem from three main areas:
- Outdated Organizational Models: Traditional divisions of front-end and back-end development, as well as product-R&D-test processes, are not adapted to the AI era. For example, although AI can write code across multiple files, teams are still structured according to old roles, leading to repeated coordination and time wastage.
- Loss of Collaborative Information: When requirements are communicated from business to R&D, context (such as business background and system rules) is often lost. Even if AI can write code, it may not meet actual needs if it does not have all the necessary information, requiring rework.
- Disconnected Knowledge Systems: Business knowledge (user requirements), domain-specific knowledge (e-commerce system rules), and R&D knowledge (code standards) are scattered among different individuals, making it difficult for AI to effectively utilize this information. This can result in non-standard code or duplicate development efforts.
2. Pitfalls of Using AI for Coding in Companies: How to Avoid Deviations and Unreliable Code?
Although AI looks impressive in demos, practical use often encounters problems such as code deviating from requirements, unnecessary architectural changes, and security vulnerabilities. Large companies address these issues in the following ways:
- Preventing Deviations on Long Tasks: They have AI ask clarifying questions before starting work. For instance, Kuaishou ensures that AI understands the requirements thoroughly before executing tasks and creates a detailed plan (SDD) first; only after human validation does it proceed with coding.
- Ensuring Code Reliability: Multiple layers of verification are implemented, including providing AI with code standards and architectural requirements in advance, having multiple AI agents review each other's work, and conducting automated testing. Finally, human confirmation is required before the code goes live.
- Managing Context Loss: This is achieved through dividing complex tasks into smaller, manageable parts, with different AI agents responsible for specific tasks (e.g., one for requirement analysis and another for code generation) to prevent confusion.
3. Can Non-R&D Personnel Also Develop Software? Is This a Trend or an Illusion?
Currently, operations, product management, and even finance departments are using AI to create small applications (such as registration systems and data analysis tools). This is not just a demo; it's a reality:
- Real-World Examples: Kuaishou has equipped non-R&D personnel with the necessary infrastructure, allowing them to develop registration systems; REDnote's Muse tool enables non-technical staff to create functional applications directly. These tools include database and AI capabilities.
- Controversial Aspects: Security and quality are concerns, as non-technical users may produce code with vulnerabilities. However, these issues can be mitigated by human oversight (e.g., in database operations and payment interfaces). For demos and automated tools, this is entirely feasible.
- Future Outlook: This trend is likely to continue. Just as the invention of movable type printing made writing more accessible to everyone, AI will enable more people to develop software. The focus will shift from coding to identifying needs and understanding users.
4. Should Companies Develop Their Own Tools or Buy Existing Ones? The Real Barrier Lies in Organization, Not Technology
External AI tools (such as Claude Code and Cursor) are updated frequently, leading companies to question whether to develop their own:
- Small Teams/Startups: It is more cost-effective to buy ready-made solutions. Community-developed tools often meet needs quickly and can be used immediately (e.g., with minimal setup costs).
- Large Companies: A hybrid approach is adopted, where core components (such as knowledge integration with internal systems and security controls) are developed in-house, while general-purpose tools are purchased from the community.
- The Real Barrier: The key challenge lies not in the technology itself, but in the organization's ability to effectively integrate AI into its systems and processes.
5. Looking Back at 2028: Will AI Coding Revolutionize Corporate Structures?
Two experts agree that AI will bring about significant changes:
- Corporate Structure: There will be no longer purely software companies; all businesses will be considered AI-powered. Teams will become more self-contained and flexible, with smaller units capable of completing the entire development process from requirement gathering to deployment.
- Changing Roles for Developers: Programmers will no longer just write code; they will evolve into “requirement discoverers” who collaborate with AI (e.g., using AI to create product requirements documents, design prototypes, or conduct competitive analysis). The distinction between junior and senior developers will blur, but those with discernment and decision-making skills will remain crucial.
- Cost Trends: If AI model costs become as low as electricity, more consumer-oriented applications will emerge (e.g., individuals using AI to develop their own mini-programs).
In summary, AI coding is not about replacing humans with better tools; it's about redefining how teams work, how knowledge is utilized, and how humans and AI collaborate. The company that best adapts its organization and systems to AI will gain a competitive advantage.
In one sentence: AI coding is not about adopting a more advanced “tool”; it's about redesigning how work is done, how knowledge is managed, and how humans and AI complement each other. Those who successfully integrate AI into their operations will be the fastest to adapt and thrive in the future.