A randomized trial involving 4,867 developers found that using AI code completion increased the average number of completed tasks by 26.08%. Our judgment is that AI has shifted engineers’ scarce capabilities from “writing code quickly” to “defining it correctly, controlling risk, and proving the result.”
“AI can already write 80% of the code, so why does your company still need you?” The point of this interview question is not to debate which methodology produces that percentage, but to answer again: when implementation costs are falling rapidly, what should an individual still deliver to a team?
What this is
AI coding is moving from generating disconnected code snippets to serving as execution systems that participate in the entire R&D process. Products such as Codex, Claude Code, Gemini, and GitHub Copilot can already study code repositories, break down tasks, modify code, run tests, review changes, and diagnose faults. In China, DeepSeek, Qwen Code, and Kimi Code are expanding with similar capabilities.
These Agents—AI systems that understand objectives and invoke tools step by step to execute tasks—are absorbing work that humans previously handed off among requirements analysis, coding, testing, and code review. However, vendor demonstrations and self-reported test results do not equal reliable delivery in production systems, and automatically generated commits do not mean enterprises have accepted deploying code without human review.
Industry view
The optimistic view is that AI expands what engineers can deliver. In the 2025 DORA survey, 90% of technology professionals surveyed used AI, more than 80% reported higher personal productivity, and 59% believed code quality had improved. Although these judgments are subjective, they show that “AI-generated code is basically unusable” is no longer an industry consensus.
We agree with a deeper change: human value should not rest on AI’s current inability to write code. Engineers need to turn organizational constraints that were never documented into explicit rules—for example, legacy interfaces must not change, financial workflows require human approval, and sensitive data must not enter third-party models. They must also establish boundaries that make automation failures discoverable and recoverable.
The risks are equally clear. Faster code generation amplifies unclear requirements, permission failures, and insufficient verification. If enterprises assess performance only by commit volume, AI may produce more unmaintainable code. What the industry truly needs is traceable decision records, test evidence, access controls, and rollback mechanisms—not the mistaken assumption that “automatically completed” means “ready to deploy.”
Impact on regular people
For enterprise IT: Teams will shift from hiring execution-focused developers to acquiring technical leaders who can design workflows, review results, and take responsibility. They will also need to establish permissions for code, data, and model access, and review critical systems in which AI participates in generation.
For individual careers: The ability to write code remains foundational, but the value of defining business problems, decomposing acceptance criteria, and judging technical trade-offs is rising. When job hunting, it is more important than merely showing experience with AI tools to explain how an automated delivery was verified and how it was recovered after failure.
For the consumer market: As software delivery accelerates, users are more likely to receive frequent small features and faster service fixes. They will also face a new problem: release speeds that outpace quality assurance. To judge whether an AI product is mature, look beyond its generation results and examine whether errors can be detected and whether its processing can be traced.