The agentic COGS stack
As head of AI R&D, I spend a number of time with architects and CTOs, and the dialog virtually at all times lands on a COGS breakdown that mirrors the agent’s structure:
- Mannequin inference: Tokens throughout planner/executor/verifier calls, often the biggest contributor to COGS of agentic software program
- Instruments and unwanted effects: Paid APIs (e.g., internet search), per-record automation charges, retries and idempotent write safeguards.
- Orchestration runtime: Employees, queues, state storage and sandboxed execution for code and paperwork.
- Reminiscence and retrieval: Embeddings, vector storage, index refresh and context-building or summarization checkpoints.
- Governance and observability: Tracing, analysis suites, security filters and audit retention.
- People within the loop: Assessment time, escalations and help load created by agent errors.
How does FinOps assist standardize unit economics when outcomes span actions, workflows and duties?
Gartner has cautioned that value strain can derail agentic packages, which makes unit economics a supply requirement.
In the case of most SaaS merchandise, clients don’t purchase uncooked tokens; as an alternative, they purchase progress towards finishing their work, e.g., instances resolved, pipelines up to date, reviews produced or exceptions dealt with. Unit economics turns into actionable once we measure on the boundary the place that worth is delivered, and that boundary expands as your agentic SaaS matures: from solutions within the UI, to a single accepted operation, to a multi-step course of and finally to a recurring duty the agent runs end-to-end. Within the following desk, we lay out this construction and the corresponding unit metric and final result to meter at every stage of scope.
