# The one-workflow pilot checklist

A practical worksheet for the person commissioning an AI workflow pilot.

Prepared by Leon, Independent FDE Engineer / AI implementation consultant.

This is an open resource. You can download and use it without joining a mailing list. Sending a project inquiry does not create a marketing subscription.

## 1. Name the handoff

**Recurring task:** ____________________

**Input and source version:** ____________________

**Person who approves the work:** ____________________

**Destination and next owner:** ____________________

Describe one completed item. The result should be inspectable in the destination, not only in an AI conversation.

## 2. Bound the permitted operations

| System | Data permitted for the pilot | Reads allowed | Writes allowed | Access owner |
| --- | --- | --- | --- | --- |
| First named system | | | | |
| Second named system | | | | |

Record the approved environment and the named pilot users. Confirm actual API or export capabilities. Mark an unverified connection as unverified.

## 3. Define review and evidence

- Keep the input ID and version with the draft.
- Record the reviewer decision for the exact output version.
- Require separate authorization for any destination update.
- Retain the actual destination URL or record ID after the write.
- List open questions instead of generating unsupported facts.

## 4. Include four acceptance cases

| Case | Expected behavior | Evidence to inspect |
| --- | --- | --- |
| Normal input | The approved result reaches the intended destination | Input version, approval, destination record |
| Missing required input | The item is held with a clear owner and reason | Missing field and visible held state |
| Write succeeds, response is lost | Reconcile the actual destination before retrying | Attempt ID and destination lookup |
| Owner takes over | The owner runs a normal case and one exception | Observed handover and operating guide |

Where available, test the destination's idempotency mechanism. Otherwise, agree a lookup and a decision for unresolved attempts before allowing automatic retries.

## 5. Measure the whole cycle

Capture a baseline from representative work. Compare handling time, review effort, corrections, missing handoffs and recovery work. Record software, model and infrastructure costs separately from consulting fees.

Keep observed results, estimates and unknowns separate. Evidence from a small pilot does not establish company-wide savings.

## 6. Make the next decision explicit

Choose **build, adjust or stop**. Record the reason, remaining uncertainty and the next owner. Any production release or broader rollout needs an agreed scope, permissions and acceptance decision.

## Discuss one workflow

Describe the recurring handoff and the two main tools involved:

https://leon-ai-workflows.pages.dev/start-project

Or email m15178824683@gmail.com. Redacted examples are enough for an initial discussion. Scope, fees and dates are agreed before paid work starts.

This checklist is general operational guidance. Its cases are illustrative, not client outcomes or proof that a specific integration is available.