PART 08

AI × Design & Creative

Generative tools can help a creative team explore drafts, formats, and localized variations. They cannot prove performance before a test, establish rights, or replace brand, factual, cultural, and accessibility review.

Generate drafts; verify rights, claims, provenance, and accessibility

Draft Variants, Product Visuals & Video

Treat each generated asset as an unapproved draft with a recorded source, model, license context, transformation history, and accountable reviewer.

BASELINE RISK

Uncontrolled asset generation

Catalog data, customer images, fonts, music, likenesses, and reference work enter tools without a rights record → false product details or inaccessible text reach review late → provenance and approvals disappear across exports.

CONTROLLED PILOT

Human-directed draft workflow

Approve inputs and permitted uses → generate a bounded set of drafts → attach provenance → check product facts, rights, disclosure, brand, culture, and accessibility → human approves an exact asset → run a pre-registered test → stop or roll back on harm.

Evidence to collect

PROVENANCE
Trace every published asset
Record approved source assets, model and version, prompt or transformation, reviewer, disclosure decision, license context, and final destination.
Artifact: asset lineage record
QUALITY
Test the failure modes
Score product fidelity, unsupported claims, trademark or likeness issues, cultural harm, text accuracy, contrast, captions, alt text, and localization.
Artifact: representative review set
EXPERIMENT
Measure, do not predict
Pre-register audience, placement, spend cap, primary metric, guardrails, attribution window, stopping rule, and comparison before delivery.
Artifact: dated experiment record
ACCESS
Include equivalent experiences
Verify captions, audio descriptions where needed, readable text, contrast, motion controls, alt text, and an equivalent non-visual path.
Artifact: accessibility review
Vendor performance and production-speed claims were removed because they do not predict your audience, placement, brand, costs, or review burden. Generated work still needs rights clearance and task-specific measurement.

Reference architecture: draft-only creative assistant

Use tenant-isolated projects, least-privilege read access to an approved catalog and asset set, and no customer-data inputs. Validate a structured asset manifest, retain provenance, and route every draft through factual, rights, disclosure, brand, cultural, and accessibility review. Publishing and campaign connectors stay disabled until a human approves the exact asset and destination; logs, stop controls, and rollback live outside the model.

Approved catalog viewRights-cleared asset libraryHuman approval queue
A successful pilot demonstrates safe review and traceability first; creative or commercial impact must be measured separately.

Pilot evidence package

  1. Rights and permitted-use register for every source image, font, voice, likeness, music track, trademark, and generated component.
  2. Versioned evaluation set covering factual fidelity, disclosure, brand, cultural context, localization, and accessibility.
  3. Exact human approval, publication record, experiment plan, complaints, stop actions, and rollback evidence.