When AI reaches its limits in asset production: Why standard tools are not enough and how your AI creation can become scalable

Many communications and marketing teams are using generative AI in marketing—with quick "aha" moments. Image generators such as Google's Nano Banana, OpenAI's ChatGPT Image, and moving image models such as OpenAI's Sora or Google Veo 3 are on everyone's lips. A hero motif, a mood clip, a few social variants, and the prototype is ready. But then production often comes to a standstill. Not because there is a lack of ideas, but because implementation in real brand work collides with real requirements: confidential product information cannot be stored in the cloud, content policies block legitimate motifs, and "a good prompt" is not yet a repeatable process.

What does AI-powered asset creation really mean?

When people talk about AI, they often refer to text. In practice, however, assets have a greater impact: key visuals, campaign variants, e-commerce clips, moving image sequences, local adaptations, and formats for an ever-increasing number of touchpoints.
Videos in particular quickly reveal where simple tools reach their limits: consistency across scenes, recurring poses or camera movements, brand-typical look, clean transitions.

The difference between "trying out AI" and "AI delivering value" is:

Control (brand safety, rights, approvals)
Consistency (look and feel across different variants, scenes, and time periods)
Production capability (reproducibility and quality management)

Depicting products in detail: Where generative AI alone is not enough

A particularly critical issue in AI-supported asset production is the realistic and brand-compliant representation of products. Standard generators often deliver "product-like" visuals, but not reliable reproductions. Logos are distorted, labels are illegible, proportions are incorrect. Or details vary from version to version. For genuine brand work, e-commerce, launch communications, or POS materials, this is a showstopper.
‍
This is precisely where a combination of AI and a powerful CGI setup comes into its own. Precise 3D models are used to create high-quality, realistic reference images: correct materials, exact geometries, and defined perspectives. These provide a controlled basis for AI-supported creation of variants, rather than serving as a vague starting point for prompts.
‍
Another advantage is that labels, claims, product markings, and logos can be integrated in a targeted, correct, and CI-compliant manner. Whether mandatory information in the DACH region, language-specific variants, or different packaging statuses—all of this can be set up in a reproducible manner without the AI misinterpreting or hallucinating, which could ultimately lead to errors.

For companies in the industrial, consumer goods, fashion, medtech, and e-commerce sectors in particular, this means:
AI not only accelerates creation, but also enables the scaling of existing product implementations without compromising their authenticity.
In combination with locally operated models and CGI, this creates an asset setup that is both legally compliant and suitable for production – from hero visualizations to series implementation for campaigns, marketplaces, and international rollouts.

The three biggest AI stumbling blocks for marketing assets

1. Content policies often affect the wrong people

An underwear manufacturer wants to create subtle, high-quality motifs, but is blocked from the platform for showing too much naked skin, even though everything remains within the bounds of regular product communication. A hospital is planning an image film and needs striking medical scenes—also blocked. Platforms such as Google or OpenAI must prevent abuse, but standard tools do not provide satisfactory context differentiation.

2. Data protection and confidentiality conflict with cloud workflows

Especially with prototypes (for example, new machines or upcoming product lines in the electrical appliance segment), uploads to external services are often taboo: IP risk, NDAs, and internal governance are the reasons for this.

3. Reproducibility: Prompts alone are not a process

In prototyping, a single prompt often delivers exactly the desired result. In production, however, this result is difficult to reproduce or scale. Consistent quality, consistent imagery, and clear parameters: not a chance. Instead of reliability, there are deviations, iteration loops, and inconsistent assets. Without defined workflows, AI remains a fluke rather than a robust production tool. Proliferation of tools and dependence on external roadmaps exacerbate the problem.

Models change, features disappear, cost structures shift. Anyone embedding AI into ongoing processes needs predictability: versioning, documented setups, and controlled updates.

Local AI models as enablers – with clear governance

Local AI models provide a remedy here: they offer more control over data, setups, and availability. Typical "hidden champions" in the stack include Stable Diffusion, Flux AI, Qwen Image, and Z-Image. The model name is less important than the orchestration: This includes targeted selection, finely tuned workflows, and operation in controlled environments, such as in your own or rented data center or on a local computer, isolated from the internet.
‍
Important: "Local" does not mean "unregulated." Only with your own guidelines (rights, consent, brand safety) does AI become scalable and responsible.

The four building blocks for production-ready AI asset pipelines

Use case cut & guardrails

We define the intersection of creative goals, risk, and feasibility, including dos and don'ts, data classification, and approvals.

Controlled model setup

Models run in the appropriate environment. Setups are versioned and "preserved" as needed to keep processes stable.

Pipeline instead of prompting

Reusable workflows, parameter templates, batch processing, logging, automated checks (format, resolution, CI characteristics), plus a clear quality assurance/review process.

Creative finishing

AI accelerates basic material and variants. Compositing, retouching, typography, color grading, and editing make it ready for branding. Consulting, creation, and finishing from a single source—with our agency family.

This is our typical approach

Recording → Workshop → Concept → Implementation → Iterative optimization

Whether your team manages the pipeline itself later on or we provide ongoing support: both options are possible.
‍
In the workshop, we focus specifically on decision-making rather than tool discussions: you will receive clear use case prioritization, a recommendation for an operating model and the appropriate environment, as well as a concrete roadmap for how to get from individual outputs to variant production – including quality assurance, roles, and process points.

Especially in localized brand communication, "one size fits all" is outdated—today, it's more like "one fits none." Those who set up asset diversity cleanly gain speed, ensure consistency, and create scalability.

Our conclusion

AI delivers the greatest benefit in marketing when it goes beyond demos and enables production-ready asset creation—even under content restrictions, strict data protection requirements, and fragmented tool landscapes. Local open-source setups are a powerful lever for this.

Patrick Fürderer-Tosolini and Erim Kansoy,
Data Scientists

Interested?

In this workshop, we will show you how to turn AI experiments into reliable asset production.

Arrange appointment