Start an AI infrastructure decision
A six-part sequence for defining an AI workload, data boundary, evaluation, unit economics, reliability plan, and exit trigger.
1. Bound one useful task
Name the user, permitted input, expected output, acceptance rule, review owner, and failure response. “Add AI” is not a workload. A reviewed support draft or accepted document extraction can be.
2. Map the data and responsibility boundary
Trace what enters the system, where it travels, what is logged or retained, and who owns access, deletion, incident response, application security, model evaluation, and user communication.
3. Build a representative evaluation set
Use authorized cases that reflect normal work and important failure modes. Write the rubric before comparing models or architecture paths. Keep the set and configuration versioned.
4. Translate billing into unit economics
Include model use, orchestration, retrieval, storage, data transfer, retries, rejected outputs, human review, support, and idle capacity. Divide by accepted useful tasks, not raw generations.
5. Design the degraded mode
Set deadlines, bounded retries, queues, manual fallback, stop conditions, observability, and recovery ownership for both dependency failures and unacceptable model behavior.
6. Record an exit trigger
Name the cost, capability, reliability, contract, or risk threshold that forces a review. Preserve the evaluation set, prompts, schemas, data exports, and operating evidence required to test another path.