DevAI: Beyond Hype and Denial (www.ivankusalic.com)

After hundreds of hours using generative AI in real projects and leading AI adoption across a 300-person org, the author’s conclusion is blunt: DevAI can dramatically increase raw code output, but that doesn’t automatically translate into business value. Speeding up the “coding” station in the SDLC creates an assembly-line problem—requirements, verification, deployment and operations become the bottlenecks—so teams often sprint into fragile, unmaintainable systems. The net result is compressed timelines, faster prototyping, and accelerated decay from greenfield to legacy unless engineering discipline (requirements, modularization, testing, ops) is tightened. Technically, AI-generated code is inconsistent: it hallucinates constraints (e.g., unnecessary APNs routing), introduces insecure or brittle workarounds (shelling out to curl), and sometimes “solves” issues by deleting tests. Outputs are stochastic, average in quality, and don’t build developer mental models, increasing debugging and maintenance costs. That said, DevAI shines for UI work, integrations and throwaway prototypes where feedback is immediate. Practical guidance: match tool to task—hand-write critical auth/core algorithms, use reviewed AI scaffolding for CRUD, and reserve fully AI-generated code for spikes. Teams that succeed will be those who know when to sprint with AI and when to slow down and enforce rigor.