After the era of AI Coding, how do mature software companies transform their product R&D processes?

From AI Coding to quality engineering, product collaboration and customer delivery, a number of mature software companies are promoting AI to evolve from a personal tool to a full production and R&D process.

In an R&D team, what does it imply if several people are highly proficient in AI Coding?

It can prove that individual efficiency has improved, but can hardly directly demonstrate faster requirement delivery, more stable software quality, or better resolution of customer issues.

Code generation is only one link in the software production process. Whether the requirements are clearly stated, whether AI understands the business rules of legacy systems, who reviews the generated content, how tests are verified, and how on-site customer issues are fed back to product and R&D teams, all these issues jointly determine whether AI can be integrated into the formal production process.

Explorations by mature software companies have gone deep from tool usage to the entire production and R&D system.

Before AI Coding is adopted by teams, the production process must be fully supplemented and improved

Individual users usually start using AI Coding from code completion, feature development, unit testing and documentation generation. At the team level, the situation is far more complicated.

A set of ToB software that has been running for many years often has legacy code, customer customization, multi-version maintenance, private deployment and implicit business rules at the same time. A lot of information is not documented, and only exists in the experience of product managers, architects and senior employees. AI can read code, but it is difficult for it to judge on its own why these rules exist, let alone accurately identify which parts can be modified.

Therefore, according to practical cases, some teams are supplementing three categories of information:

Convert verbal requirement descriptions into executable specifications, and clarify the objectives, scope and acceptance criteria;

Document architecture constraints, interface specifications and historical decisions to reduce AI's guesswork on context;

Integrate code review, testing and release check into a unified process, so that AI-generated content can pass through a stable quality gate.

This change also brings new measurement methods.

The code generation ratio is very intuitive, but it cannot represent the overall efficiency. More worthy of attention are requirement response time, R&D cycle, rework rate, defect rate, regression testing time and customer acceptance cycle. If the code generation link becomes faster but subsequent rework increases, the overall delivery efficiency still cannot be improved.

According to the on-site sharing and survey at the Beijing ToB AI Production & R&D Conference hosted by CNIU Club on July 31, we found that: AI Coding has received the highest attention, followed by testing, quality, security and go-live acceptance. Enterprises have begun to pay attention to production issues after code generation.

After development speed is increased, quality engineering must be moved forward

AI increases code output, and also expands the scope of quality management.

In the past, testing was usually concentrated after the completion of development. After AI speeds up the development process, if the testing team still receives deliverables at the final stage, it will easily become a new bottleneck. Deviations in requirement understanding, architecture conflicts and missing business rules will be intensively exposed in the testing, go-live and customer acceptance stages.

Quality work needs to be moved to earlier stages.

In the requirement stage, it is necessary to define in advance what results are considered correct; in the product design stage, business boundaries, exception paths and acceptance conditions must be clarified; after entering the development stage, AI-generated code needs to go through independent review, and test cases must cover real user paths, instead of only verifying whether the functions can run.

The on-site discussion also touched on a key issue: For AI-generated test cases, who judges whether the test coverage is complete?

This work still requires human involvement. AI can help generate test cases, analyze defects, perform regression testing and inspect code, but business correctness needs to be jointly defined by product, R&D and testing teams.

Especially in ToB scenarios with high requirements for security, compliance and customer acceptance, permissions, security, audit and customer acceptance cannot be left to the model for autonomous decision-making.

The role of the testing team has also changed accordingly. It starts to participate in requirement and solution reviews at an earlier stage, helping the team establish verifiable goals, quality standards and release thresholds. The faster the development speed is, the more important this set of mechanisms becomes.

Real changes take place in collaboration relationships

When AI is integrated into requirement, product, R&D and testing workflows at the same time, the original handover method between different posts will also change.

Product managers can use AI to sort out customer interviews, work orders and meeting minutes, and also get assistance in generating PRDs and prototypes. The premise of efficiency improvement is that product managers clarify the objectives, boundaries and acceptance criteria first. When the requirement definition is vague, AI will only speed up the generation of wrong solutions.

The work focus of R&D personnel will also change. As the proportion of code writing decreases, the importance of architecture design, context building, code review and quality control will rise. R&D teams need to help AI understand the system, constrain the generation scope, and be responsible for the final results.

Testing is further moved to earlier stages, and the delivery team also needs to be integrated into this set of processes. Customer issues of ToB software often occur on site. When customer feedback is passed from sales to delivery, then to product and R&D teams, information loss is very likely to occur.

AI can help sort out on-site information, identify common issues and form requirement clues, but enterprises still need to clarify: which issues are included in the product roadmap, which belong to personalized delivery, and which can be precipitated as standard capabilities.

A more complete production and R&D link is thus formed:

Customer issues are converted into requirements, requirements are converted into executable specifications, R&D personnel generate code under constraints, testing teams verify according to acceptance criteria, and delivery results are fed back to the product team.

Whether this link can operate normally determines whether AI will eventually stay at the level of personal tools, or become a stably reusable production capability for the whole team.

This article is from the WeChat Official Account "Neuters" (ID: Neuters), author: CNIU Club, published with authorization from 36Kr.