A Case for Generative AI? (medium.com)

A skeptical startup CTO recounts a pragmatic turnaround: after earlier failed experiments, he gave generative AI (Claude) another try to tackle concrete engineering tasks and saw measurable results. His product is a two-repo stack (frontend: TypeScript/React/Redux/Tailwind/Vite; backend: vanilla Java), ~150k LOC, with strong test coverage, strict TypeScript and -Werror. Over months he merged 9 AI-generated PRs that now account for roughly 4% of the frontend. The most visible win was a safe refactor that extracted a reporting component from an object view, separating state and data loading; builds and extensive Playwright tests passed and the change has been running in production for two months. Other PRs improved readability, component granularity, centralized layout calculations and introduced patterns like memoization and custom hooks. The story highlights a practical sweet spot for generative models: they excel at discrete “nail-hammering” tasks and filling knowledge gaps (e.g., hooks/memoization, OIDC hardening, Azure integrations), boosting productivity for small teams that can’t afford specialist hires. But limitations remain — models behave like junior engineers with imperfect risk awareness, so robust tests and human oversight are essential. Economically, AI can deliver outsized value for lean teams, though its cost-effectiveness and market sustainability at scale remain open questions.