I shipped an AI-generated PR that violated four of our own architecture decisions. Nobody caught it

Short summary

An AI-generated PR reintroduced a banned dependency six weeks after removal because architectural decisions lived in ADRs and Slack, not in the AI system or review checklist. AI tool adoption has increased per-engineer PR throughput from 1 to 4-6 daily, cutting reviewer attention by 80% and making review the only enforcement point unsustainable. Real governance requires typed, enforced policies upstream of code generation, not advisory IDE config suggestions.

• •AI models correctly follow their prompts; the failure was insufficient architectural context and governance mechanism placement
• •Review bottleneck is structural: throughput increased 4-6x while reviewer capacity stayed fixed, breaking the single-checkpoint model
• •Current safeguards (pre-commit hooks, CI checks, CODEOWNERS) sit downstream of AI generation and cannot prevent architectural drift

Generated with AI, which can make mistakes.