AI often feels inexpensive to the person using it because the bill is hidden. The cost becomes visible when a team moves to metered APIs, adds multi-step workflows or reaches a license renewal that finance can no longer ignore.
With token-metered APIs, providers commonly charge for model input and output, and output is often priced higher. Cached input, reasoning tokens, retrieved context and other features may be treated differently by each provider. Long conversations and repeatedly supplied context can increase input usage quickly. Some platforms use caching, retrieval or context compaction to change those economics, so measure the behaviour of the actual API rather than estimating from what the chat window shows.
That leads to three practical consequences.
- Context size is a cost lever, not just a capability. Sending substantially more context than the task needs increases usage and can sometimes reduce answer quality by burying the relevant evidence.
- Reasoning modes and multi-step agent workflows can increase model usage and cost substantially. Budget per completed task, not per call.
- Flat-rate subscriptions hide this. When a team moves from a consumer subscription to an API, the same behaviour becomes a metered bill.
Running a model locally moves the bill rather than removing it. Hardware, electricity, administration and someone's time replace part of the provider charge. Part D examines that trade in more detail.
The useful comparison at any scale is cost per successfully completed task, including human review. Compare complete outcomes, not isolated model calls.
Part C follows Northstar Services from a simple policy question to a controlled employee service. Each new technology appears because the previous design reaches a limit. That sequence matters more than the individual product or framework choices.
- 10 Let the company problem drive the design
- 11 Retrieve evidence with permission
- 12 Test confidential retrieval end to end
- 13 Give tools narrow authority
- 14 Understand the MCP boundary
- 15 Control an agent outside the model
- 16 Walk through an attempted overreach
- 17 Evaluate the task and its failure modes
- 18 Decide what evidence permits release
- 19 Design for hostile and unreliable input
- 20 Give risk decisions an owner
- 21 Assemble the production service
- 22 Operate for failures and change