My AI team built a campaign baseline in a weekend. The ad has never run.

By Sunday night, my AI team had produced a marketing baseline for a new wardrobe-management subscription business.

The business is at the first-engagements stage. The video has never run as an ad. Nobody has clicked, converted, renewed, or proved that the creative works. This is a production result, not a market result.

The team started by extracting the business context and writing a marketing plan. Every forecast in it was marked as an estimate. Two workers independently designed sets of ad scenarios, then the sets were judged against the same brief and reduced to one campaign baseline. From that came two generated hero images and a 22-second, 9:16 silent video master.

Producing the two directions separately mattered. They could disagree before either knew what the other had proposed. The later judgment had an actual choice to make, with the same business facts and the same stage limits on both sides. Speed did not require one worker to draft an idea and then quietly grade its own work.

The same weekend produced a Google Ads proposal through a new reusable workflow, a Blender scene-control workflow with a cryptographically verified handoff, and an encrypted briefing page for a marketing partner. The partner sent due-diligence questions. The page answered them that day. That means the response was fast. It does not mean there is a signed partnership.

The interesting part is where the system stops. After I reviewed the first master, we put a new gate in front of every future video-generation call.

The Blender frame uses no video-generation credits. I can inspect the camera, composition, and scene before a generation request exists. The handoff records that the approved source frame is the one referenced later. Until I approve it, the credit-bearing route is closed. Approval unlocks the route. It does not press the button for me.

The finished master used 166 generation credits across its generated shots. At delivery, its checker ran 193 assertions. We fed the checker a deliberately broken copy and watched it reject the file. Browser checks covered the pages. Self-tests were also forced to fail before their green results counted. After publication, the pages were fetched again from the live site and checked as a visitor would receive them.

Each gate had a narrow job. The plan had to distinguish estimates from evidence. The media checker had to reject a malformed delivery. The Blender handoff had to reject a changed source frame. The search proposal had to remain paused without spending approval. The browser checks had to catch a broken page. Encryption had to keep the partner briefing private. A green result in one lane was never allowed to answer for the others.

Those gates establish that the files have the intended shape, the checks can catch defects, and the published pages are the pages we tested. They establish nothing about whether a person wants the service or whether the ad persuades her.

AI supplied the speed. It could turn the context into a plan, argue two creative directions, make the assets, assemble the master, package search ads, and answer a partner without waiting for me to move each piece by hand.

Taste stayed with me. Money stayed with me. Approval to spend stayed with me. The system can prepare the decision and stop directly in front of it. I can reject the framing without buying another attempt. I can like the frame and still decline the spend. I can approve both, then authorize the request. Those are separate gates because they answer separate questions.

The package exists. The business is in its first engagements. The ad is still waiting for its first audience. The next evidence has to come from a real campaign.