# AI Pilot Decision Worksheet

Use this before approving or extending an AI experiment. Revised 2 October 2026.

## The decision

Write the decision this trial should inform. Identify who can approve the next step and what evidence would lead them to stop, change or expand the trial. A review date is a date for a decision, not an unconditional promise to deploy.

## Evidence to collect

- The current process: completed cases, elapsed time, quality, rework and operating cost.
- The proposed change: the tasks AI may perform, the tasks people retain and the permitted systems and data.
- The result: the same measures on representative cases, including review and repair effort.
- The consequences: errors that matter, how they are detected, who can stop execution and whether the fallback works.
- The operation: a named owner, support capacity, required approvals and a credible cost to run the workflow.

## At the review

Record which assumptions held and which did not. Choose one of four responses: stop, revise the experiment, expand testing or approve a bounded deployment. State the evidence and conditions for that choice. A financial gain does not compensate for unacceptable quality or risk.

Do not retire the existing process, change headcount or commit a long contract simply to make the experiment irreversible. Make those decisions when the evidence and operating readiness justify them.

Source: [AI Pilots Are a Form of Procrastination](https://juanbeltran.ch/blog/ai-pilots-are-a-form-of-procrastination)