# One-task AI pilot

Import task-pilot.csv. Replace the teaching row. Choose one job, one permitted input, and observable done checks. Record a comparable manual baseline, including checking and corrections. Mark estimates rather than describing them as measured.

```text
Perform only this bounded task using the permitted input below.
Task: [FILL]
Reader or user: [FILL]
Input: [PASTE PERMITTED MATERIAL]
Expected output format: [FILL]
Acceptance checks: [FILL]
Prohibited data or actions: [FILL]

Use only supplied facts. If needed information is missing, state the gap and ask a focused question. Do not invent a quote, event, result, capability, or source. Do not send, publish, change settings, or use private accounts.
Return the draft output and the inputs that remain unknown.
```

Run an ordinary and a sparse case. Inspect each output against its original input and checks. Record test time, checking, corrections, errors, and fallback. Keep capacity and cash separate. A one-hour difference doesn't establish revenue or reduced spending.

Continue only if the observed result is useful and the remaining risks and work are understood. Otherwise change the method, choose another task, or keep the manual process. Set a review date based on the task, not a universal pilot duration.