Let's get the objection out of the way first: yes, you can tell when creative is lazy AI output. Plastic skin, cursed hands, that weirdly smooth "everything is an airbrushed render" look. When people say "AI ads don't work," that's what they're picturing.
But that's not an AI problem. That's a judgment problem. AI didn't make those ads bad. A human shipping the first draft did.
The principle: AI for speed, humans for judgment
My whole workflow is built on one split: machines do the production, humans do the decisions. AI never decides what to make, which angle matters, or whether an asset is good enough to carry a client's money. It just collapses the time between "brief" and "testable asset" from days to hours.
Pool rule 008AI for speed. Humans for judgment.
The pipeline, step by step
1. Angles from the data. Before anything gets generated, I read the account: which hooks are holding attention, which angles are fatiguing, what competitors are running repeatedly (if they keep paying for it, it's working). The output is an angle map: problem-aware vs product-aware, comparison, social proof, credibility, offer. This step is 100% human and it's where the ads are actually won.
2. Brief with one job per ad. Every creative gets a single job: stop the scroll with X, land claim Y, push offer Z. Briefs with three jobs produce ads that do zero jobs.
3. Generate in batches. This is where the tools earn their keep: Fal, Higgsfield, Nano Banana Pro, depending on the job. Product scenes, lifestyle contexts, background swaps, texture work. I generate wide: many options per brief, knowing most get binned.
4. The human pass. Real product renders (never AI-invented products), real claims, real typography. Photoshop for compositing and cleanup, brand palette enforced, compliance checked. The AI look dies in this step. What survives reads like brand creative, because at the layers that matter, it is.
5. Modular variations. Winning concept → new hooks, new crops (1:1, 4:5, 9:16), new headline stacks. Variations are nearly free once the concept exists, and they're what keep an account fed between big swings.
What I refuse to let AI do
- Invent the product. The product in the ad is always the actual product. Non-negotiable.
- Write the claims. Claims come from the brand's reality and compliance, not a model's imagination.
- Approve itself. Nothing ships without a human deciding it would stop them mid-scroll.
Why bother? The economics.
Testing is a volume game: most ads lose, and the winners pay for everything. If producing one polished static takes your designer two days, you can't afford a real testing cadence; you'll protect every asset emotionally because it was expensive. When production is fast and cheap, you can let the data be brutal. That's the whole unlock: AI doesn't make better ads. It makes honest testing affordable.
Never run an ad you'd scroll past, no matter how fast you made it.