AI Ad Creative Generation for E-commerce Marketing Teams

At a glance

A US performance marketing organization asked Brainy Neurals to take the queue out of its ad production.

Engagement facts

Why couldn’t manual ad creative keep up?

Manual ad production could not keep pace with campaigns that need fresh hooks & new cuts every week. Audiences stop responding to an ad they have already seen, which is why this US performance marketing organization turned to AI ad creative generation.

Each ad was made one set-up at a time, & the queue was the product. A marketer wrote the brief while a designer built the frames. An editor then cut the video, & a voice artist recorded the read before anything could reach a campaign.

Nothing started until the person before had finished, & the delays stacked up at every handoff.

• Four people worked on one thirty-second video
• A hook idea waited on a designer
• Every variation restarted the whole chain
• Past winners were reviewed by eye
• Campaign tests waited on the edit bay

Slow creative carries a real cost, because ads wear out. A meta-analysis of advertising repetition found that an ad’s effect on attitude & recall decays over time^[1]. The creative has to keep moving, which is where generative AI development earns its place.

What do teams try before this?

Teams usually try four routes before building an engine, & three of them still suit some jobs well.

The last route pays off because generative tools save real time. An experiment with 453 professionals found that a generative writing assistant cut task time by 40 percent^[2].

How we designed the creative engine

Brainy Neurals built the creative engine in two halves that share one store of advertising insight. The analysis half reads the ads this client already knows worked, one at a time.

For each ad, the engine samples frames from the video & transcribes its speech, then reads picture & words together.

Mood boards are older than the internet, & agencies pinned them to studio walls in the sixties. The generation half is a mood board that writes back.

That half works from a product page & a campaign angle aimed at one audience. Out come hooks, a full script, product images, a voiceover & a finished cut.

A model with no reference writes fluent advertising that has never sold anything.

The technology stack we used

The technology stack had to cost less per creative than a designer, or the engine would save nothing. The same layers show up in our computer vision development work.

Ingestion & analysis

Generation

Delivery

How does one brief become a video?

One brief becomes a finished video in six steps, & the engine runs them in this order.

1. The engine reads the product page & pulls out the claims & benefits it may use, along with the images.
2. Next, the campaign angle & audience are matched against the insight store built from the client’s winning ads.
3. Hooks & a full script come next, with every claim traced back to its source page.
4. Product images & testimonial-style frames are generated for each scene the script calls for, in the brand’s own tone.
5. A voiceover is recorded from the finished script, so the spoken read & the written words never drift apart.
6. Finally, the cut video goes to the marketing platform ready to test, & every asset is stored alongside it.

The six steps run in one pass. The marketer’s part is the brief at one end & the pick at the other.

What went wrong before it worked?

Three problems showed up before the creative engine worked, & they arrived in this order.

The first scripts were fluent & wrong. The model wrote benefits the product page never claimed, in the confident voice of a good copywriter. An invented claim in an ad is a legal problem rather than a typing mistake.

Next, the engine wrote thirty variations of one idea. Every hook opened the same way & every script hit the same beat, so all that volume tested nothing.

Turning the variety up made the writing worse. Research on ad text generation reports the same trade-off between diversity & ad quality^[3].

The testimonial-style creatives also read as real customers. The federal endorsement guides treat a promotional review as an endorsement, & a fabricated one is deceptive^[4].

These are the weeks when clients ask us about specialist engineers instead of learning the slow way.

What we changed to fix them

We changed three things, one for each problem, & every fix is short to describe even though none was quick to find.

Finding problems like these is the job an AI proof of concept does, before anything is promised.

What changed after it went live?

After go-live, creative production moved from weeks to hours, the client reports, because AI ad creative generation replaced seven handoffs with one pass.

We have not published a cost per creative or any click-through & conversion figures. The client reports a large drop in manual effort, & a number we have not measured is a number we will not print.

Day to day, a marketer with an idea in the morning can have it running by the afternoon. Because the engine writes from the product page, the claims survive review without a rewrite.

Writing many angles at once also means a losing hook costs one run rather than a week.

What is running today

The creative engine runs in production today, across this client’s live advertising campaigns. A marketer briefs the engine, then pushes the chosen set straight into a live test after review.

Since handover, the client has pointed the engine at more of its retail accounts, one brand at a time. The insight store grows with every advertisement the team marks as a winner, so the engine keeps learning.

The review step never moved, & a named person still signs off before any of it runs.

What we would do differently

Four lessons from this build would change how we start the next one.

A model will happily write the same advertisement thirty times & call it thirty variations.

Where else does AI ad creative generation fit?

A creative engine like this one fits wherever ads are tested faster than a team can make them. Brands in five other industries face the same problem, each with its own twist.

Porting the engine takes a new claim source & a brand voice pass, plus a review step that somebody in the business owns.

Services behind this case study

Six Brainy Neurals services came together in this build.

An AI proof of concept is the fastest way to try this on your own catalog. AI consulting helps pick the formats, & an AI readiness assessment shows whether your content can carry it. The industries hub shows where the approach already runs.

Similar case studies

Three more Brainy Neurals builds shipped into real use rather than into a demo.