Producing Ad Creatives with AI: the 2026 Workflow

In 2026, the bottleneck of a Meta or TikTok campaign is no longer budget, and it isn't targeting either. With Advantage+ and the Andromeda era, the algorithm decides who sees what: your job is to feed it enough different creatives to test. And this is where almost every company gets stuck, because producing fifteen or twenty ad variants the classic way (brief, designer, revisions, exports) eats up days and several hundred euros.

AI image generation flips the equation. Not because it "does everything on its own" (it doesn't), but because it moves people's work to where it matters most — concept and offer — and industrializes the repetitive part: composing the scene, adapting formats, multiplying variants. Here's a concrete workflow for producing ad creatives with AI systematically, not through improvised prompts.

Why creative production became the bottleneck

For years the competitive edge was targeting: whoever profiled better, won. Today that advantage has nearly disappeared. Platforms moved the intelligence inside the algorithm, and creative has become the main signal the system uses to figure out who to target. That's the core of how Andromeda changed the role of creative: the creative is no longer just "the ad", it's the input that steers distribution.

The practical consequence is stark. The algorithm learns fast, and just as fast it "tires out" the audience: a creative that was performing well can saturate within a week or two. To keep your cost per result from climbing, you need a steady stream of new, varied material. Depending on spend, an active campaign can need anywhere from a handful to several dozen new creatives a month just to stay fresh. With the traditional method, every extra creative costs an extra hour: production scales linearly and becomes the ceiling on how far you can grow.

The 5-step workflow for producing creatives with AI

A good AI workflow isn't "open a tool and type a prompt". It's a repeatable chain, where each step sets up the next. Here are the five steps I use as a structure.

1. Start from real assets, not from nothing

Mistake number one is generating the ad from pure text and letting the model invent a fake product. For a real campaign you need your product, recognizable. So you start by gathering assets: high-resolution product photos (multiple angles), brand palette and fonts, offer and price, the main selling point. The difference from a few years ago is that today's models don't just "draw" — they edit: they take the real product photo and build a scene, a background, a layout around it. That's what makes the result usable in advertising.

2. Lock the concepts before the prompts

You don't ask the AI "make me an ad". You decide the concept: collage with call to action, "us versus them" comparison, UGC style, lifestyle, before-and-after, benefit list. Turn these into reusable templates, so quality stays consistent even when you produce at volume. If you're short on angles, start from a proven playbook: our complete guide to ad creative collects the patterns that convert most often. The concept is the part where human judgment matters most: define it before touching any generator.

3. Structured prompts, with the copy in Italian

Here you turn concept and product into a structured prompt: layout, palette consistent with the product (no electric purple on a pastel cosmetic), and exact text values. The rules that make the difference: copy in Italian, short headlines (a few words, not walls of text), price written as 39,90 and not $39.99. Text rendering has historically been the weak point of image models: it improved a lot in 2026, but large, clean lettering works while small paragraphs don't. Design your creatives accordingly and keep a human eye on every word.

4. Generate with the right model

There's no single "best" model — there's the right one for the job. It's worth knowing the image generation tools available for marketing and understanding what each one is good at.

For volume production, today's workhorse is Gemini's image model family, the one the community calls "Nano Banana": it's fast, costs a few thousandths of a euro per image, and is strong at editing real products. We covered the case in depth in Nano Banana 2 applied to advertising. For hero concepts or trickier situations, you can step up to a premium model — pricier, but more faithful to the prompt.

5. Variants at scale (where AI really wins)

From a validated concept, you generate many variants by changing one lever at a time: hook, background, angle, color and, above all, format (1:1, 4:5, 9:16). What used to take a designer half a day now takes minutes: from a single concept you can pull 8-15 publishable variants in a few clicks. The trick is to generate them in batch and name them consistently (concept, variable, format), so once they're in the campaign you know exactly what you're testing. This is how you cover how many creatives you need each month without blowing up costs and timelines.

From generation to testing: fueling campaigns that perform

Be careful: volume is not a strategy. Churning out a hundred random images is pointless if you don't then put them to the test with a method. The next step is structured testing, isolating one variable at a time and watching the metrics that actually matter (hook rate in the first seconds, CTR, cost per acquisition), not vanity metrics. If you don't have a process yet, start with a structured creative testing method before hitting "publish".

From there the loop closes itself: you measure, you kill the losers, you push the winners, and you feed the angles that work back in as the base for new variants. It's the creative loop: production, testing, learning, more production. AI exists precisely to keep this loop fast and cheap, not to churn out the one "pretty" image and stop there.