AI Can't Save Software Engineering (medium.com)

A recent critique argues that the current rush to build "AI for Software Engineering" (AI4SE) startups misunderstands the problem: software engineering is fundamentally a socio-technical system driven by people, incentives and org structure, not just code. While AI can flag bugs, surface technical debt, or suggest architectures, it cannot change who is responsible for fixing things, nor can it rewrite KPIs or inter-team incentives—Conway’s Law means architecture mirrors organization. That mismatch makes many AI4SE products nice-to-have analytics rather than must-have interventions, because the real bottlenecks live in meeting rooms and spreadsheets, not in the codebase. The piece also highlights two practical barriers: measurement and business model. Proving an AI tool’s impact is hard—undetected bugs may never cause outages, the causal chain to revenue is fuzzy, and purchasers care about profit metrics, not cyclomatic complexity. And most ventures are “experts selling to experts,” which limits scale because engineers already judge tools by expert norms. The clear implication for the AI/ML community: focus on products with immediate, tangible developer value (the Cursor example) or on solutions that sell to outsiders with clear ROI, rather than trying to use models as a managerial hammer to fix organizational problems.