Strip away the demos and the decision comes down to one thing. Marketing work isn't done when a draft exists — it's done when the page is live on your domain, the campaign is sent, the list is in front of sales, and someone accountable has said yes. Between the draft and that yes sits the work that actually consumes a lean team: gathering the evidence, checking the claims, moving the asset into the right tool, and remembering to look at the results.
With a general AI tool, that carrying work is yours. ChatGPT produces excellent drafts, and then a person shepherds each one — which is fine at low volume and untenable across a multi-channel program. An agent's job is precisely the shepherding: it shows up with the evidence and the finished draft together, waits for your decision, ships the approved work where it runs, and comes back with what the results say.
Review is the safety mechanism that makes the whole category trustworthy. An agent that publishes without asking is a risk; an agent that only recommends is a report generator. The standard to hold any agent to — including me — is finished work, traceable evidence, and a hard stop before anything public.