AI Creator-Agent Acceptance Kit
Use this to decide whether an agent produced a useful research record or merely a list of names. Bring a campaign brief and inspect one candidate from beginning to end. Do not treat a fluent explanation as evidence that an action occurred.
Worked example
Fictional app brief: find a creator who can demonstrate a household-planning task to U.S. buyers, with an option for paid use. These outputs illustrate review decisions; they are not a test of a named tool or actual creator.
| Output | Evidence present | What remains unresolved | Human decision / next action |
|---|---|---|---|
| Search brief | Buyer, task, market and intended platform stated | Usage period unknown | Campaign owner sets usage need before requesting complete quotes |
| Creator A match | Link to a relevant task demonstration | Audience geography unconfirmed | Watch full work; request recent platform Insights if buying distribution |
| Creator B match | Profile contains related keywords | No relevant demonstration linked | Request supporting work or remove from current shortlist |
| Audience report | Dated estimate with a quality warning | Method and recent primary data need review | Keep warning visible; investigate rather than label fraud |
| Outreach draft | Refers to the actual work and asks rates/availability | Recipient and wording not approved | Review draft; no message sent in this example |
| Scope proposal | One video listed | Revision, usage and payment terms missing | Request a scoped offer; do not call it a booking |
| Suggested product claim | Polished sentence without source | Accuracy not established | Product approver supplies evidence or removes claim |
| Campaign status | Names saved to a list | No accepted deliverable or result | Report “research in progress,” not “campaign complete” |
Reuse the inspection record
| Candidate / output | Task requested | Evidence URL / artifact | Source date / limits | State | Human approver | Next action |
|---|---|---|---|---|---|---|
| Enter your record | Brief interpretation | Enter your record | Enter your record | Enter your record | Enter your record | Enter your record |
| Enter your record | Relevant work | Enter your record | Enter your record | Enter your record | Enter your record | Enter your record |
| Enter your record | Audience evidence | Enter your record | Enter your record | Enter your record | Enter your record | Enter your record |
| Enter your record | Rates and scope | Enter your record | Enter your record | Enter your record | Enter your record | Enter your record |
| Enter your record | Outreach draft | Enter your record | Enter your record | Enter your record | Enter your record | Enter your record |
| Enter your record | Claim approval | Enter your record | Enter your record | Enter your record | Enter your record | Enter your record |
| Enter your record | Final booking approval | Enter your record | Enter your record | Enter your record | Enter your record | Enter your record |
Use supported, partially supported, not checked, missing, rejected with reason, or approved for the named action. Approval for one action does not cover another: a useful candidate can be approved for a scope discussion while the booking remains unapproved.
Check operating value
Record the task, time spent checking, corrections needed and usable output. If comparing with an existing process, use a comparable brief and count review time as well as generation time. Do not infer campaign ROI from a faster candidate list.
Your final output: “The agent established ___. It did not establish ___. We approve ___ next, subject to ___. The responsible person is ___.” Another reviewer should be able to open the evidence and understand the same decision.