# Cross-border AI pilot worksheet — public planning template

Adapted from my cross-border AI consulting materials. The interface, accounts, amounts and receipts in the videos are synthetic. They explain a proposed workflow, not a deployed client system or measured business results.

A sanitized input example, the current handoff and the person responsible are enough for an initial fit discussion. We can define a diagnostic or pilot scope before any production access or commitment.

## Scope agreement

- One task / market / account / action:
- Existing tool gap and manual baseline:
- Business, IT, approval and maintenance owners:
- Read-only scope, data use and retention:
- Approved objects, limits, cooldown and stop conditions:

## Earn the next level of access.

### 01 / Diagnose · read only

Confirm the tool gap, source dictionary, baseline and owners using sanitized samples.

Exit: agreed task, data scope and acceptance criteria.

### 02 / Draft · shadow mode

Generate candidate actions and compare them with the operator’s judgment.

Exit: reviewed rules, representative cases and workable stop conditions.

### 03 / Approved · limited writes

Use one designated writer, bounded objects/limits, action IDs, readback and escalation.

Exit: written authority and demonstrated exception handling.

### 04 / Review · hand over

Assess net value; transfer Skills, regression cases, runbooks and team practice.

Exit: named maintenance owner and a continue/change/stop decision.

## The delivery is a set of usable operating assets.

- Data contract: Input schema, source revision, freshness, timezone/currency and allowed data use.
- Rule book: Version, observation window, sample minimum, limits, cooldown and conflict policy.
- Action ledger: Evidence, old/new values, approver, action ID, execution status and platform readback.
- Acceptance pack: Representative cases, failures, denominators and operator review—not only a green demo.
- Recovery runbook: Stop, reconcile before retry, restore where permitted and hand over to an owner.
- Team capability: Versioned Skills, dependency list, training exercise and a maintenance/change checklist.

## What a useful acceptance test looks like

- Stale advertising data: Hold the candidate; keep the reason and route it to the operator. No bid write.
- Submission timed out: Reconcile the actual platform state before retrying. No duplicate task.
- Unsupported product claim: Flag the phrase, link the approved source and return the draft for review.
- Creator opts out: Suppress further contact and synchronize the status; keep only necessary records.

## Measure the work that remains.

Net time benefit = comparable manual baseline − manual operation, review and exception-handling time after automation. Record model, tool, implementation and maintenance costs separately. Compare advertising with the same attribution and observation window; do not attribute all revenue or traffic changes to AI.

This public template contains no client identity, live business data or production credentials. Scope, fees and dates require separate agreement.