Plan a small AI pilot your team can evaluate

A small AI pilot should answer one question: does this workflow produce acceptable work with less total effort? Give it an owner, a bounded task, and a way to stop.

Write the pilot brief

Task and current process:
Input source and permitted data:
Expected output:
Actions requiring a person’s approval:
Reviewer:
Acceptance rules:
Examples for testing:
Costs and active work time to record:
Conditions that stop the pilot:

Do not begin with a target number of agents. Begin with a task whose result someone can inspect.

Use a review queue first

A proposed support pilot might take one incoming question and an approved FAQ, then create a draft for a person. It should not issue refunds, change account records, or send replies during the initial exercise.

The reviewer checks policy facts, cited FAQ entries, missing information, and whether the draft needs account access. The support draft exercise includes a sample FAQ and five acceptance tests.

Separate checks from predictions

Before the pilot, prepare a few representative inputs and their expected outcomes. Include an incomplete request, a request outside the task’s scope, and an instruction embedded in the input that attempts to change the rules.

Keep the original cases when revising the prompt. Count incorrect outputs and rejected drafts. Do not describe a successful sample as proof that the whole process is reliable.

Record the work that remains

Measure preparation, review, corrections, and repair after errors. Keep tool costs and setup effort separate. The measurement worksheet shows how to calculate capacity value without calling it automatic cash savings.

Make the decision explicit

Continue when the accepted outputs meet the rules and the total effort improves. Revise when a recurring failure has a testable fix. Stop when the workflow creates more work, reaches data it should not access, or takes an action without the required approval.

Expansion is another decision. A second channel, more data, or new tools change the scope and require new tests. This plan is a pilot worksheet, not a claim of a tested enterprise deployment.