21 Images, One Spot, Four Hours: What a Solo AI Campaign Actually Shows

One person, one afternoon, 21 image motifs, and a 25-second ad spot – produced by Georg Neumann, an AI specialist and marketing bootcamp lead, as a demonstration of what current generative AI can do today. Eight hero images came together in roughly four hours, using a stack of image models, a video engine, AI voice, and AI music, at a cost in the low double digits. For an agency that builds exactly this kind of campaign with teams of photographers, retouchers, and producers, that's not a footnote.

What genuinely impresses about the demo

The honest first reaction: the raw editing speed is real. What used to require a pre-production meeting, a studio day, and several rounds of retouching now happens iteratively on screen – one visual world after another, one variant after another, with no location booking and no waiting between departments for approvals. For concept phases, pitches, and early client presentations, that's a real gain, regardless of what actually ends up in production. I use comparable tools myself during the concept phase, for exactly that reason.

The part missing from the demo

Two details significantly temper the "four hours" figure once you read it toward real agency work. First: the advertised product doesn't exist – it's a fictional vintage dive watch, invented for the demonstration. There's no real reference object whose exact proportions, material surfaces, or brand logo need to be correct. Second, and the article states this openly: real client projects take three to four times as long. That's not a footnote, it's the actual point – between a demo with an invented product and an approval process with a real client, a real brand book, and a real legal department sits exactly the work an agency actually sells.

Where it actually breaks down

More interesting than the speed are the limits the demo itself names: content moderation blocks certain terms in the video AI, spatial logic – left, right, which hand is holding what – frequently gets confused, and detail fidelity on small products needs extra refinement steps before a watch face actually looks legible instead of merely plausible. But the decisive sentence concerns something else: domain knowledge doesn't automatically prevent results that are functionally wrong but visually convincing. That's exactly the error class that makes the biggest difference across seventeen years in image production – not the image that's obviously broken, but the one that looks right at first glance and only on closer inspection reveals a wrong crown on the case or a physically impossible light reflection.

What this means for agency work

For us as an agency, this shifts the actual question. Not: can one person now produce what used to require a team? That question is answered with "partly, for certain formats," and it's the same sober distinction I already drew about AI image quality in Google Ads. The real question is: who catches the wrong crown before it goes live? Execution time can be radically compressed by AI – four hours instead of four days, that's the real number from the demo. Review time – the trained eye that tells a functionally wrong but visually convincing image apart from an actually correct one – doesn't automatically compress along with it. That's the same shift I described about AI agents in a different context: the narrower the frame, the more impressive the demo – and the more the outcome ultimately depends on who sets the frame for the real case.

Conclusion

Four hours for 21 image motifs and an ad spot is a real, impressive number – for a fictional watch, with no client, no brand book, no legal department, produced by someone who knows exactly how to operate the tools. Actual agency work begins right where this demo ends: with real products with real proportions, with approval processes that take three to four times longer, and with the errors only a trained eye catches before they go live. AI compresses execution. Who takes on the review remains, for now, a human question.