AI marketing automation approval checklist

Set the boundary before the prompt.

An agent may have the technical ability to perform an action. That does not give it business authority.

Write down who owns each consequence before connecting the account. The rule should cover five things: the action, the allowed scope, the approval point, the activity record, and the rollback.

One account. Four levels of authority.

Copy this record before any live action.

A usable approval gate records the proposed action, the authority granted for this action, and the person accountable for the consequence. Tool access alone proves none of those things.

Request ID and versions

Assign one request ID. Record the agent, model or workflow version, input version, proposed payload version, and the last known good live version.

Action and system

Name the exact draft, edit, publish, send, spend, delete, or permission change and the account where it will occur.

Data and scope

List the data class, audience, locations, exclusions, assets, and settings the action can touch.

Impact and recoverability

Record possible customer harm, claim or legal risk, data sensitivity, audience exposure, money at risk, permission scope, and whether the action is reversible, only stoppable, or remediable after release. The organization sets its thresholds before the request.

Limits and expiry

Set the spend ceiling, audience boundary, frequency, stop condition, and date when the permission expires.

Reviewer and evidence

Name the reviewer and attach the rendered payload, sources for claims, final audience, links, budget, schedule, and version.

Approval and payload proof

Record approver, time, approved payload identifier or hash, and decision. Approval applies to that exact version; any audience, link, price, budget, timing, or permission change reopens review.

Execution and live-state proof

Record the platform receipt, executed payload identifier, actual-versus-approved differences, and a live-state check that confirms what customers can see, receive, or be charged.

Rollback owner

Name the person who can pause or restore the action and identify the last known good version before release.

This control structure applies the govern, map, measure, and manage functions in the NIST AI Risk Management Framework Core. Claims and endorsements still require substantiation and clear disclosure under FTC advertising guidance and the FTC Endorsement Guides.

Every live change needs an owner, a verified result, and an honest recovery plan.

A screenshot of a finished screen is weak proof. Compare the executed payload and live account state with the approved version. A successful API response does not prove that the intended audience, message, budget, price, or permission is live.

Prevent

Use sandbox or draft mode, validation, least privilege, separate approval, and bounded audience or spend before execution.

Pause or revoke

Stop future sends or spend and remove the agent's access. This limits further impact but does not retract a delivered message or viewed ad.

Restore

Return a reversible setting, template, segment, or campaign to the recorded last known good version, then verify the live state.

Remediate

For actions that cannot be undone, name who corrects the claim, notifies affected people, restores eligibility, refunds, or handles the incident.

Least-privilege examples

An email agent can receive draft access without send authority. An ad agent can analyze and propose changes while a buyer-owned administrator retains publishing, billing, and user-management control. A catalog agent can edit a staged feed without changing live price or availability. Grant only the action, account, data, duration, and volume required; use separate identities and revoke access when the task expires.

Measure approval quality, not just approval speed

Review instant approvals, reviewer disagreement, exceptions, rejected payloads, near misses, incidents, and time-to-decision. Repeated instant approval or high reviewer load can signal rubber-stamping. Narrow or automate only recurring low-impact actions whose error, exception, and recovery records support the change; reapprove the new boundary and keep sampled human review.