Skip to main content
Liorry Herisnor
Runtime state: Running in productionMarketing and Growth

AI Visual Creative Production

Generate broadly against the same approved brief, then select and finish by hand — because the expensive part of visual production was always the discarded attempts.

Runtime state: Running in the business as an operating procedure. Generation is tool-assisted; selection and finishing are manual by design.

Business objective
Cut creative production cost by making exploration effectively free, without shipping work that is recognizably unfinished.
Trigger
An approved brief requires visual assets.
Inputs
  • Approved brief — audience, objective, constraints, tone
  • Brand system — color, type, spacing, treatment rules
  • Format and channel requirements
Outputs
  • A selected, refined and finished set of assets
  • A concept set retained as reference for future briefs
  • Performance data once live

The economics

The expensive part of visual production was never the final piece. It was the ten discarded attempts before it — each one requiring the same setup, the same iteration and the same time as the one that shipped.

Generation makes those attempts effectively free. That is the entire saving, and it is why this workflow generates far more than it uses.

The corollary matters just as much: unfinished generated work is recognizable. There is a distinctive look to output that went straight from generation to delivery. The finishing step is not polish — it is what makes the work ours rather than a tool's.

The workflow

Brief to finished assets

Generate widely, discard most, finish by hand.

  1. Approval gate

    Brief approved

    Decision point: The same brief that drives content production. One artifact, so the visual and the copy cannot drift apart.

  2. Data store

    Brand system loaded

    Color, type, spacing and treatment rules that the finished asset must satisfy.

  3. AI / LLM

    Concept generation

    Multiple directions generated against the brief. Volume is deliberate — most will be discarded.

  4. Human action

    Selection

    Decision point: Which directions serve the brief? Most are discarded here, and that is the intended outcome.

  5. AI / LLM

    Variant generation on selected directions

    Depth on the directions that survived, rather than breadth across all of them.

  6. Human action

    Refine and finish

    Composition, typography, brand application and the corrections that make it not look generated.

  7. Approval gate

    Quality review

    Decision point: Approved, or returned with specific notes. 'Looks generated' is a valid and common rejection.

  8. Automated software

    Export to format requirements

    Channel-specific sizes and formats.

  9. Monitoring / feedback

    Performance observed

    Which directions performed, feeding the next brief.

  • Human action
  • Automated software
  • AI / LLM
  • Data store
  • Approval gate
  • Monitoring / feedback

Decision points

  • Which concepts survive selection? Human judgment against the brief.
  • Depth or more breadth? Generate variants on what survived, or go wider.
  • Is it finished, or does it still read as generated? The most common rejection reason.
  • Approve or return?

Responsibilities

AI / LLM

  • Concept generation against the brief.
  • Variant generation on selected directions.
  • Alternative composition options.

Human and deterministic

  • Brief approval.
  • Selection — which directions are worth pursuing.
  • Refinement, brand application and finishing.
  • The quality gate.
  • Format export, which is a deterministic transformation.
  • Creative direction and strategy.

Human approval points

  1. Brief approval.
  2. Concept selection — the judgment step.
  3. Quality review before delivery, where "still looks generated" is a legitimate rejection.

Failure handling

FailureResponse
All concepts miss the briefThe brief is the problem. Revise it rather than regenerating against the same input.
Output does not satisfy the brand systemReturned for finishing; brand rules are applied by a person, not requested in a prompt.
Finished asset still reads as generatedRejected. This is the failure the finishing step exists to prevent.
Selection fatigue on large concept setsCap the set size. More options past a point makes selection worse, not better.
Generation tool unavailableManual production, at manual cost and pace.

Monitoring

Concepts generated versus concepts selected, revision cycles to approval, time from approved brief to finished asset, and production cost per deliverable. A selection ratio that climbs toward one suggests we are generating too narrowly to get the benefit.

More workflow designs

Back to the library