AI-generated Meta ads: what to delegate and what to keep

In Meta Ads Manager, a few images and a line of copy are now enough to produce dozens of ad variants. Meta has been pushing in this direction for several years, and the question is no longer whether to use these tools. It is what to hand over to them, and what never to leave to them.

Meta ads: what is automated at the end of September 2026

Under the Advantage+ name, Meta groups its automated campaigns and options. On the creative side, Meta presents generative AI tools built into Ads Manager: image variations based on the original ad, background generation around a product photo, image expansion to fit other formats (the vertical Reels format, for example), and variations of the headline and primary text. Meta has also presented turning still images into short animated videos, initially as a test.

At Cannes Lions in June 2026, Meta announced a further step: creative generation that draws on the brand's identity, text generated directly inside visuals, translations, and a creative approval workflow in Ads Manager. These features are rolling out gradually: check your account for availability in Switzerland.

Finally, according to the Wall Street Journal in June 2025, Meta aims to fully automate the creation and targeting of ads by the end of 2026: the brand would supply a product image and a budget, and the AI would do the rest. This is a reported goal, not a feature we have been able to see in the accounts we manage.

What AI does well

• Producing variants. From a good ad, it creates variants of background, framing and wording, without a new photo shoot.
• Adapting formats. A square visual becomes vertical for Stories and Reels without manual cropping.
• Finding people. With broad targeting, the algorithm identifies who responds, often better than a list of interests chosen in advance.
• Allocating budget. It continuously moves spend towards the ads and placements that deliver the result you asked for.

These are volume and calculation tasks. They used to take time for a designer or a campaign manager, and the machine does them faster. Delegating them is reasonable.

What it does not know

• The offer. What you sell, at what price, on what terms. AI varies what it is given; it does not invent a better offer.
• The angle. The reason someone should choose you now: security, price, proximity, timing. That is a commercial decision.
• The proof. A real customer review, a verified figure, a completed project. AI can write a persuasive sentence with no proof behind it.
• What is true about your product. Materials, finishes, view, floor area, delivery time: only the brand knows them.
• The French-speaking Swiss tone. Formal address, restraint, prices in CHF, local vocabulary. Generic copy quickly sounds like France or like a translation.
• French/German bilingualism. A machine translation is not an ad written for a German-speaking Swiss audience: Swiss German spelling does not use the ß, and references change from one region to another.
• Legal constraints. The Federal Act against Unfair Competition (UCA, LCD in French) prohibits inaccurate or misleading statements about your products and prices. A promise you cannot prove has no place in an ad.

The risks: misleading visuals, unchecked copy, sameness

The first risk is the most concrete. A generated background can add a lake view, a garden or a finish that does not exist. In real estate, that is a serious mistake: never show what does not exist. A buyer who discovers at a viewing that the ad was untrue does not reserve, and says so. For a development, turn off the options that alter the image, or check every variant before it runs.

The second risk is copy published without review. A variation can turn “up to” into a firm promise, add a superlative or distort a condition. Meta itself documents an “AI info” label for some ads created or edited with its generative tools: the ad still runs under your name, and under your responsibility.

The third is sameness. If every advertiser in a sector hands its creative to the same tools, fed with the same data, the ads end up looking alike. In a small market such as French-speaking Switzerland, what sets a brand apart is precisely what AI does not produce on its own: a distinctive angle, local proof, a recognisable voice.

The method: what the brand keeps

1. Write the brief before opening Ads Manager. Offer, audience, available proof, permitted promise, banned words. AI works better within clear limits.
2. Supply strong starting material. Real photos, approved real renders, real reviews. A variation is never better than its source.
3. Choose which options are on. Automatic enhancements are set at ad level: check which ones are on by default and turn off those that alter image content when accuracy matters.
4. Keep human sign-off. Someone who knows the product reviews every text and every visual before it runs, in French as in German.
5. Write German for German speakers. For a bilingual market, have German-language ads written or reviewed by a native speaker rather than publishing a machine translation.

Test angles, not pixels

AI multiplies surface variants: background colour, cropping, a synonym in the headline. These differences matter little. What really moves results is the angle. Let Meta vary each angle, and keep the choice of which angles to test for yourself.

Send back quality conversions

Meta's algorithm optimises for what it is told. If it only sees submitted forms, it looks for easy forms, not customers. Sending qualified enquiries, viewings or sales back from the CRM teaches it to look for the right people. The principle is explained in offline conversions for Google and Meta.

Also watch for creative fatigue: rising frequency, falling click-through rate, climbing cost per result. Generated variations delay the wear, without removing it. When an angle is exhausted, you need a new angle, not yet another variant of the same one.

This is the logic we apply in our Meta and Google campaigns: automation for volume and budget allocation, the team for the offer, the angle, the proof and the sign-off. For Masini Groupe, Meta acquisition feeds a custom CRM that qualifies enquiries automatically; for Arboréa, a bilingual market, the ads rely on a bilingual French/German website.