DEREK MURRAY
Case study · Generative AI, creative infrastructure

From AI
experimentation
to AI production.

A five-layer architecture that turns Generative AI from a collection of tools into creative infrastructure.

Built at · Prezent
Architecture · 5 layers
Principle · Models are swappable
Scroll
The challenge

Generation is easy. Reliable production is hard.

Generative AI dramatically lowered the barrier to creating content, and introduced a new problem for creative organizations. Individual tools could produce impressive images, copy, concepts, and variations, but isolated tools do not make a scalable creative workflow.

The real question was never which model is best. It was these:

How does a business brief become an AI-ready creative brief?
How do strategy and brand standards survive across multiple models?
Which model handles each part of the process?
What happens when a model returns something unusable?
Where does human judgment enter?
The answer was to stop treating Generative AI as a set of tools and start treating it as creative infrastructure.
5
layers connecting intelligence, orchestration, generation, manipulation and human production
1 of 5
models responsible for any single step. No model owns the whole process
∞
models swappable without rebuilding the workflow. The system outlives them
The architecture

Five layers. One job each.

01

Intelligence

Turns the business objective into structured creative direction: objective, audience, positioning, brand rules, channel specs. The output isn't a prompt; it's a creative specification downstream models execute against.

StrategistCreative DirectorCopywriter
02

Orchestration

Decides what happens next and which model performs each task, using model chaining, APIs, agent frameworks, and node-based environments. Separates the workflow from the model, so components stay interchangeable.

WorkflowModel routing
03

Generation

Concepts become assets. The right engine is chosen per brief rather than forcing every project through one: controlled production imagery, concept exploration, conversational editing, typographic work, illustration and icon systems.

FLUXMidjourneyOpenAI / GeminiIdeogramRecraft
04

Manipulation

Stop regenerating; start directing. Inpainting, outpainting, image-to-image, object and background replacement, canvas extension, upscaling. An approved asset becomes a master: 1:1 extends to 16:9 and 9:16 and a full campaign family.

Surgical controlAsset reuse
05

Production

Control returns to human creatives. Typography, grid, identity, color accuracy, compositing, retouching, accessibility, channel specs, creative QA, answering the only question that matters: would a professional team actually ship this?

Human creative finishQA
The principle

No single model owns the process.

No single AI model should be responsible for the entire creative process. Each layer solves a different problem, and because the layers are separate, a better image generator six months from now is a component swap, not a rebuild.

01Intelligence
02Orchestration
03 · SwappableGeneration
04Manipulation
05Production
The workflow stays fixed. The engine inside a layer changes
The models are commodities. The workflow architecture is not.
The feedback loop

Not a waterfall. A learning system.

The five layers aren't a waterfall. Every output is evaluated, the failing layer is diagnosed and corrected, and the pattern is recorded. Successful patterns feed the next brief.

LEARNING SYSTEM GENERATE EVALUATE DIAGNOSE LAYER CORRECT RECORD PATTERN
Successful patterns feed the next brief
Diagnosis

Diagnose the layer, not the prompt.

When something fails, the failure has an address. That distinction is what stops teams from trying to solve every problem by prompting again.

The strategic shift

Most organizations are still in the middle column.

Traditional production
  1. Brief
  2. Team
  3. Production
  4. Revisions
  5. Delivery
Basic AI adoption
  1. Prompt
  2. Generate
  3. Regenerate
  4. Generate again
Most organizations are here
Mature AI operation
  1. Business objective
  2. Intelligence
  3. Orchestration
  4. Generation
  5. Manipulation
  6. Human production
  7. Distribution
  8. Performance data
  9. Learning → better intelligence
The outcome

That final loop is the real opportunity.

As campaign performance, creative decisions, model behavior, brand preferences and human feedback become inputs to future workflows, the production system becomes increasingly specific to the organization using it, and increasingly hard to copy.

The five-layer architecture scales Generative AI while preserving the things that make professional creative work valuable:

StrategyTasteBrand consistencyCreative judgmentProduction discipline

The goal was never to remove creatives from the process. It was to remove production friction so they spend their time on the decisions machines are least qualified to make.

Generative AI creates assets. A production system creates creative leverage.

Want this built for you?

Brand from zero, design systems that hold at volume, and AI production that returns hours to the people who should be thinking.

See the rest of the work → Get in touch