> `YOUR GRACE AGENCY  /  AUTOMATION LAUNCH KIT  /  SETUP  /  08 OF 11`
>
> **AUTOMATE. DELIVER. EMPOWER.**
> *A launch is a decision supported by test evidence—not a switch you flip.*

# Automation Launch Checklist

**Use this checklist from selection through limited live launch.** Complete it for one workflow before starting another.

| Gate | What must be true before moving on |
| --- | --- |
| Select | one clear trigger, owner, and outcome are named |
| Design | required fields, destination, and approval boundary are visible |
| Test | representative records prove the happy path and failure path |
| Launch | a limited scope, daily review, and fast rollback are ready |

> **OPERATING RULE**
> If a consequential decision cannot be safely described, tested, and paused, keep that step manual.

## 1. Choose the workflow

- [ ] The work happens at least twice a month or causes a repeatable handoff problem.
- [ ] The workflow has one observable start event, not a vague intent.
- [ ] A named owner is accountable for reviewing exceptions and improving the process.
- [ ] The first version can avoid autonomous decisions about money, contracts, access, privacy, security, employment, health, or legal matters.
- [ ] I can state the desired outcome in one sentence: “When `[trigger]` happens, `[owner]` can reliably see or do `[outcome]`.”
- [ ] I chose the matching template in `01-workflow-templates/`.

**Decision gate:** If any box is unchecked, keep the process manual and document what is missing.

## 2. Define the operating design

- [ ] I named the workflow: `YGA | [workflow] | v1`.
- [ ] I listed the source of truth for each input field.
- [ ] I created a simple field schema with only the information needed to route, fulfill, or measure.
- [ ] I chose the destination record where the result will be visible.
- [ ] I wrote explicit allowed labels for any classification step.
- [ ] I documented what should never trigger the workflow.
- [ ] I identified one manual-review queue and the person who monitors it.
- [ ] I wrote a rollback action: disable the trigger, label affected records, and notify the owner.

## 3. Protect the data and people

- [ ] Test data uses fictional names and contact details.
- [ ] Passwords, API keys, payment-card data, government IDs, health details, and unnecessary private data are excluded.
- [ ] Any AI prompt receives minimized fields only; the original record remains in the source system.
- [ ] Client-facing messages are drafts until a human approves them.
- [ ] Sensitive, billing, security, privacy, refund, account-access, and dispute cases route to a human.
- [ ] Collaboration permissions were checked before creating shared folders, projects, or notifications.

## 4. Build the minimum viable version

- [ ] I configured one trigger, one validation, one branch, one destination action, and one owner notification.
- [ ] I added a duplicate check using the safest available unique key.
- [ ] I use a test destination or a visible `TEST` label.
- [ ] I made every action idempotent where possible: a second run updates or alerts rather than blindly creating a duplicate.
- [ ] I added a failure alert that includes record ID, workflow name, and failed step.
- [ ] I kept irreversible or external actions (sending, charging, deleting, publishing) manual in v1.

## 5. Run the test set

Create at least 10 records with Prompt 16 or your own examples. Include the following:

| Test | Required result |
| --- | --- |
| Standard valid record | reaches the destination with correct fields and owner |
| Missing required field | enters manual review; no silent failure |
| Duplicate event | updates/links the existing record; no second message/task |
| Invalid format | shows clear error or review path |
| Off-policy / sensitive item | reaches human review; no AI/external send |
| Wrong route attempt | is caught by allowed labels or validation |
| Delayed event | does not create an outdated or misleading action |
| Source unavailable | alerts owner and preserves recovery instructions |
| Destination unavailable | captures failure and prevents accidental repeats |
| Human rejection | lets the owner stop or edit the drafted output |

- [ ] I compared each actual result to its expected result.
- [ ] I fixed every blocker: duplicate creation, lost data, wrong owner, external action without approval, or missing rollback.
- [ ] I reran failed tests after each change.
- [ ] I saved screenshots, run IDs, or notes for the final 10 tests.

## 6. Limited live launch

- [ ] The owner approved the exact trigger and scope.
- [ ] The workflow starts with a narrow segment, label, or manually selected records.
- [ ] The owner will review outputs daily for the first seven days.
- [ ] The success metric and guardrail are recorded before launch.
- [ ] The review queue has a response target (for example, within one business day).
- [ ] The rollback action can be completed in under five minutes.
- [ ] The next review date is on the calendar.

## 7. Final sign-off

| Check | Owner initials / date |
| --- | --- |
| Workflow purpose and scope are clear | |
| Test evidence reviewed | |
| Human approval boundaries are active | |
| Failure path and rollback verified | |
| Metric baseline recorded | |
| Limited launch approved | |

## What “done” means

The automation is not done because it is switched on. It is done when the owner can explain what starts it, where results go, how duplicates are handled, which actions remain human, and how to pause it safely.

---

`YOUR GRACE AGENCY  ·  SETUP / LAUNCH CHECKLIST  ·  COMPLETE ONE WORKFLOW BEFORE STARTING ANOTHER`