Summary of Key Points
While AI can automatically generate 60% of the code, it has not significantly improved corporate R&D efficiency. The root cause lies in the fact that writing code itself is not the most time-consuming step in the development process; AI does not address critical preparatory tasks such as requirement analysis and technical solution design. The code generated by AI requires manual review and cannot be used directly. AI only speeds up individual coding tasks, not the entire cycle from requirement formulation to product launch, which is what companies are really concerned about. Moreover, the cost of maintaining AI-generated code is high. The core bottleneck in R&D remains human abilities, such as understanding problems and designing system architectures, which AI cannot yet replace.
Detailed Explanation
1. Writing code is just the “last step” in the development process; preparatory work takes the most time
Many people mistakenly think that R&D consists solely of writing code. However, it’s more like cooking: when a client asks for a membership system, the team needs to discuss various details (number of membership levels, benefits, user transition, payment integration, point calculation, and whether the system should be integrated with apps, websites, or mobile apps). These discussions can take multiple meetings before everything is clear (similar to preparing ingredients, washing, cutting, and seasoning vegetables). Additionally, the team must design the technical solution (e.g., how to set up the database and define APIs). Preparatory work accounts for 80% of R&D time, while AI only helps with about 60% of the actual coding. Without proper preparation, overall efficiency does not improve.
2. Unreviewed AI-generated code is useless
No matter how fast AI generates code, it still needs to be manually reviewed. If issues arise online, the responsibility lies with the developers, not the AI company. For example, if AI produces thousands of lines of code, the team must verify its correctness and whether it can handle high user loads. Unreviewed code is like a draft; you need to check every detail, just as you would proofread a written assignment, otherwise, you’ll lose time correcting errors.
3. AI improves “local speed,” not the “overall cycle” that companies need
A GitHub study claims that AI can speed up individual coding tasks by 55%, but what companies care about is the total time from requirement submission to product launch. For instance, a project might take 2 weeks for requirements discussion, 1 week for design, 3 days for coding, 2 weeks for testing, and 1 week for deployment. Even if AI reduces coding time by half, the overall cycle only shortens by 2 days, which is minimal. Companies want projects to be launched faster, not just certain parts of the code to be written more quickly.
4. Maintenance of AI-generated code is a challenge
Although AI writes code quickly, it does not understand the rationale behind it. If changes are needed after the project goes live, the design may be unclear, and you might have to rely on AI again, with uncertain results. This is like using someone else’s messy notes for review—you can’t understand them and must rewrite everything, increasing maintenance costs and offsetting any initial efficiency gains.
5. The core bottleneck in R&D is not code writing; it’s human capabilities
Software development is about transforming vague business goals (e.g., “making users more likely to subscribe”) into functional systems. This requires understanding customer needs, analyzing business logic, designing system architectures, managing risks, and collaborating as a team—abilities that AI currently lacks. AI is a tool; the value of a system ultimately depends on human expertise.
In conclusion
AI can help with coding faster, but it cannot solve the most time-consuming and thought-intensive parts of the R&D process. Code is a tool, and the true key to success lies in human capabilities.