Selected model
Naive implementation
Every step uses expensive defaults, retries are broad, and verification lacks risk targeting.
Cost Anatomy
AI unit economics as an architecture problem
Toggle between representative workflow designs and inspect where cost units move when routing, sandbox reuse, retries, and judge coverage become explicit.
Representative normalized model. Exact company costs omitted. Representative sanitized scenario. Customer data, private prompts, internal traces, and exact company costs omitted.
100 units
LLM planning12 units
No task classification.
LLM code generation24 units
Premium route used by default.
Sandbox execution14 units
Setup repeated too often.
Judge verification22 units
Broad judge coverage.
Chart rendering8 units
No batching.
Deck rendering6 units
Late failures are expensive.
Retries10 units
Retry policy is not task-aware.
Observability4 units
Basic traces only.