saas-agent-unit-economics-and-cogs
SaaS Agent Unit Economics & COGS Skill
Overview
Generic AI unit economics treats one user query as one LLM call. Agent unit economics is fundamentally different: one user request triggers a planner step, N worker steps, M tool invocations, possible retries, possible parallel branches, possible critic / supervisor calls, possible human-in-the-loop (HITL) escalations. A single resolved customer-service ticket can be 15-40 LLM calls and 5-15 tool invocations. A single "deep research" outcome can be hundreds of LLM calls. The cost waterfall is multiplicative, not additive, and investors will not accept generic AI cost models for agent products.
This skill installs the agent COGS discipline:
- Agent direct COGS — LLM tokens at each step (planner + workers + critic), tool invocations, external API cost, retry overhead, branch cost, HITL escalation cost.
- Agent indirect COGS — eval-loop cost, observability / tracing infrastructure, audit-log retention, sandbox / staging cost, supervision UX runtime.
- Agent reserves — irreversibility-incident reserve, regulator-engagement reserve, model-migration reserve.
- Agent overhead allocation — share of Agent Architect / Tool Engineer / Eval Engineer / AI Safety Lead / HITL Designer payroll.
Together these produce cost-per-task, cost-per-resolved-task (the true unit because unresolved tasks still cost money), agent gross margin, intervention-cost overhead, and the agent-cost-as-%-of-agent-revenue diagnostic.