modelnote

Pilot desk / 04

Make the first run
easy to inspect.

A first AI workflow should teach your team something useful. The aim is not to prove that a system is autonomous; it is to make a small piece of work more legible, reviewable, and reversible.

The working rule

Do not begin with a system nobody can explain.

Start with an existing handoff: a recurring request, an input someone already checks, a decision someone already owns. The smallest useful pilot shows where the model helps, where it fails, and what a person needs to review.

A five-part pilot

Use the sequence as a planning aid. Keep a note of every unknown it exposes.

01

Name the job

Write the trigger, expected output, source material, and a named owner. If those are fuzzy, the first pilot is still a discovery exercise—not an automation.

02

Set the operating boundary

Make a short list of what the workflow may touch, what it may draft, and what must stop for review. Prefer a narrow, reversible environment for the first run.

03

Define evidence

Decide what a reviewer needs to see: the output, the source material, a change record, a test result, or an explicit escalation. Confidence is not evidence.

04

Run a small case

Use a representative task, including one awkward case. Record the outcome, the correction, and the point where the workflow became unclear.

05

Keep or change one thing

Turn the finding into a better brief, boundary, check, or regression case. Expand scope only when the current version is understood.

The review packet

A reviewer should not have to reconstruct the work from a cheerful summary.

  • The original request and the boundary of the task.
  • The output and the evidence that supports it.
  • The tools or systems touched during the run.
  • Failures, unresolved assumptions, and required follow-ups.
  • The named person who decides whether to use the result.

Start with the brief

Write the job, boundary, and evidence request

Use the local brief builder before choosing tools or promising outcomes.

Then test the work

Build a five-case evaluation suite

Turn an encouraging demo into a small, repeatable learning loop.