How AI Is Redefining QA: Lessons from an On‑Site Recycling Team

Two fatal defects of AI‑generated code

Even though AI‑written code is syntactically perfect, it suffers from two critical shortcomings that easily slip past human code review:

Blind confidence : the model lacks business knowledge and often imposes legacy validation rules on new scenarios.

Local view : the model focuses on the current code block, missing the overall system context, which leads to omitted business scenarios or hidden trigger points such as asynchronous compensation tasks.

Consequently, the bottleneck in software engineering has shifted from "how to write code" to "how to convey business intent accurately to machines".

1. Abandon traditional test cases – adopt Intent‑Driven testing

Pain point : step‑by‑step test scripts are a "list of actions" that AI cannot understand, resulting in massive junk test cases.

Our approach : replace low‑level scripts with high‑level "intent documents" that capture three core elements:

Intent – the business rule to verify.

Risk – the consequence of a failure.

Acceptance criteria – the boundary of correct behavior.

Human experts define the intent; the AI uses it as a perfect prompt to generate executable API and UI automation steps, embodying the principle "people define rules, AI writes steps".

Past: step‑by‑step test cases overload AI with detail and hide business intent. Now: Intent‑Driven testing aligns high‑level semantics with large‑model comprehension. 2. Upgrade test gate – Business‑Perspective AI Code Review Pain point : tools like SonarQube catch syntax and style issues but cannot detect business‑logic drift introduced by AI‑generated code, causing critical workflow failures after test submission. Our approach : at the merge‑request stage, feed the generated intent document (treated as a VibeContract) and the code into a large model that performs a bidirectional comparison. The model automatically evaluates whether the implementation satisfies the acceptance criteria, flagging missing pre‑checks or omitted scenarios before the code reaches the test environment. Figure 1: AI matches QA intent cases against code to pinpoint high‑risk logic bugs. Figure 2: When backend smoke tests are missing, AI issues a clear “do not merge” verdict and produces a remediation checklist. 3. Automate infrastructure – focus on testing the “deep water” Pain point : with AI doubling code output, manual regression and fragile automation scripts cannot keep pace, leading to untested critical paths. Our approach : shift from blind full‑coverage to risk‑matrix‑driven precision. Let AI handle basic happy‑path verification and routine rule checks, while human experts concentrate on high‑risk, AI‑blind zones. Basic verification by AI : integrate existing ApiTest platform and UI automation layers so AI generates, executes, and maintains happy‑path tests. Interface automation report: massive regression frees daily QA effort. UI automation report: intent‑generated scripts run stably, replacing manual UI traversal and assertions. Human experts now tackle the "deep water" scenarios that AI struggles with: Scenario‑combination explosion testing : detect state conflicts and over‑privilege when multiple processes intersect (e.g., order timeout label combined with expired contact window still triggers a transfer). Extreme states and negative exceptions : Complex state‑machine omissions (e.g., missing coach‑machine return order in a specific assessment stage). Boundary‑value failures (e.g., price‑difference = 300 causing incorrect risk downgrade). Silent data filtering (e.g., empty adminNumber silently dropped by the backend, resulting in successful but unrecorded transactions). Conclusion – From test executor to AI testing co‑pilot In the AI wave, a QA team’s competitive edge is no longer the number of test cases it can run, but its ability to decompose vague requirements into quantifiable, automatable rule engines. By front‑loading risk through intent‑driven design, guarding the gate with AI code review, and freeing human effort to focus on deep‑water risks, the team has transformed from a pure test executor into the co‑pilot of an AI‑augmented testing toolchain.