Why the Architect AI Copilot Still Hasn’t Arrived

Alexander Polomodov and Sergey Baranov compare architecture practice with 51 studies from 2019–2025. They cover decisions, ADRs, patterns, and recovery, but not a coherent copilot, long-term quality, or production impact. ArchBench and R2A-Bench also measure artifacts more readily than relationships and trade-offs.

An ADR, diagram, or pattern list is a format, not the decision. Teams first explore alternatives, quality attributes, constraints, and residual risk. AI can collect facts from code and traces, check rules, propose options, and document an understood choice, but not replace judgment.

The gap lies between tools: planning, coding, and operations receive AI support, while modeling, governance, and artifact links remain fragmented. No shared memory connects a requirement, ADR, model, code change, and production signal. A provider that captures it could sell outcomes, but at higher context cost and deeper vendor lock-in.

A real assistant must adapt decisions, maintain traceability, use local knowledge, enforce constraints, prove quality, and observe consequences. Architects decide which uncertainty to investigate first and own the risk. AI accelerates routine work and leaves unusual trade-offs to people.