A client came to me recently wanting an entire product campaign generated with no real photography or video at all, purely to test how far that approach could actually go against their usual production budget. Worth writing up honestly, because the result wasn't the clean success story either AI skeptics or AI boosters would want to hear.
What actually worked without a real shoot
Establishing and lifestyle-adjacent shots, environments, mood pieces, generalized scenes without the specific product needing to be held or interacted with precisely, generated convincingly enough that the client approved them without hesitation. That's a genuinely large chunk of a typical campaign's shot list, and removing the real production cost for that portion alone was a meaningful budget win they were happy with.
Where we quietly brought in real photography anyway
Any shot requiring the actual product to be shown with total accuracy, exact packaging text, exact proportions, a hand interacting with it in a specific functional way, kept breaking down under close review even with reference-anchored generation. We ended up shooting those specific frames for real and compositing them into the otherwise AI-generated campaign, which the client hadn't originally planned for but agreed was necessary once they saw the alternative up close.
- Environmental and mood shots: fully AI-generated, approved without hesitation
- Precise product interaction shots: still needed real photography
- The final campaign ended up hybrid, not purely one or the other
The actual lesson from this project
"Fully AI-generated" turned out to be the wrong framing for the brief from the start, and the honest, useful version of this pitch going forward is "AI-generated where it genuinely works, real photography where it doesn't," presented upfront rather than discovered partway through a project after a client's expectations were already set differently. I've changed how I scope these conversations because of this one job, and I think that's a more useful lesson than either "AI can do it all" or "AI can't be trusted," which is closer to how this actually gets pitched in a lot of coverage I read.