Imagine an AI workflow generating 100 ad variants before lunch. The obvious question is how many more assets the brand can launch. The harder question is who can approve those assets without missing an expired offer, an unsupported product claim, a likeness right, or a market-specific rule.
That is the operational shift. Once generation moves faster than expert review, the scarce resource is no longer the ability to make another version. It is the ability to make a defensible release decision. Brands should still track asset volume, but the more useful procurement unit is an approved creative decision: one version cleared by the right owner, against the right evidence, within an agreed time.
The production layer is scaling faster than the decision layer
Several current primary sources point in the same direction, even though none proves that the whole industry has completed this transition.
Adobe describes Firefly Creative Production as a workflow layer that can generate campaign variants from approved assets, templates and data. The same product documentation retains role-based access, human-in-the-loop approvals, regional thresholds and a log of every run. It also routes personalised work through validation, approval and quality checks before activation. In other words, scaled output does not remove approval; it makes approval part of the system design. Adobe’s production workflow documentation was current when reviewed on 7 October 2026.
Adobe Experience Manager makes the downstream gate even more explicit. Its documentation, updated on 25 September 2026, says only the latest approved version is available for use, while unapproved or superseded assets are held back from general distribution. Approval is not a decorative status field; it determines which asset can move into production or publishing. Read the AEM approval guidance.
Ad platforms preserve advertiser responsibility as well. Google Ads says generated assets are not guaranteed to be approved, and advertisers should review them for accuracy, misleading content, policy compliance and applicable law before publishing. Google’s generated image guidance also states that AI-made advertising is held to the same content standards and enforcement as other ads. Amazon requires advertisers to ensure AI-generated content is accurate, not misleading and properly disclosed, including consent when a real person is mimicked. Amazon’s AI content requirements leave those obligations with the advertiser.
Governance standards add useful evidence but do not make the final release decision. IAB’s September 2026 AI Transparency & Disclosure Framework V2 uses a risk- and materiality-based approach rather than treating every AI contribution identically. C2PA’s July 2026 deployment guidance explains how Content Credentials can provide tamper-evident provenance signals. Those signals can improve a reviewer’s context; they do not determine whether a particular claim, face, price or localisation is safe to publish.
Those are verified capabilities and responsibilities. The review queue below is ONCE’s operating model built from them. It is not an Adobe, Google, Amazon, IAB or C2PA standard, and it is not legal advice.
The queue grows when arrival rate exceeds decision rate
Let λ represent the number of variants entering review each week, and μ represent the number of valid review decisions the team can complete. If λ remains equal to or higher than μ, the backlog grows. Near-zero generation cost does not create near-zero release time. It can produce more waiting, more correction rounds and a larger surface for mistakes.
This matters for teams connected to Chinese supply chains. One master product video may become multiple language, price, retailer, marketplace and social versions. Changing a crop may be mechanical. Changing a translated claim, a promotional price, a person, or a usage scene is not. The output files can look similar while the approval questions are completely different.
We have previously written about setting authority limits for autonomous media operations and turning rights records into reusable operating assets. A review queue connects those two controls: it directs limited human attention to the versions where judgment can actually change the outcome.
ONCE’s four-lane Creative Review Queue
Green: mechanical variants from locked inputs
These change only aspect ratio, crop, safe area, file format or a fixed field from an approved data source. The template, product claim, people and usage rights remain unchanged. A team can automate validation and review a sample from each batch, provided that the batch boundary and escalation rule are explicit. Green does not mean “nobody is accountable.” It means risk has been reduced through locked inputs and repeatable rules.
Amber: creative changes that affect interpretation
New copy, offer combinations, reordered benefits, AI-generated backgrounds, translation and local phrasing belong here. These changes can alter what a customer believes about the product, so a brand or product owner should approve them. One approved master does not automatically clear every language, price and scene.
Red: claims, rights or regulatory exposure
This lane includes health, financial or environmental claims; real people or digital replicas; children; competitor comparisons; copyright or music rights; and market-specific disclosures. Each version needs direct review and a recorded decision from legal, compliance or an accountable local-market owner. A model’s guardrails are useful, but they are not a substitute for that decision.
Black: missing basis, blocked from release
If the product data has no known source, a licence is missing, an expired price remains in the asset, a person’s consent cannot be established, or the version no longer points to a recoverable source, the work should stop before publication. This is not an “urgent approval” lane. It is a hold until the missing basis is restored.
Every review package needs six fields
- Asset lineage: Identify the master, product data, people, music, generation tools and current approved inputs.
- Change type: State whether the batch changes format, copy, price, language, person or scene. “New version” is not enough.
- Decision owner: Name who can approve—brand, product, legal, local market or media—and who is only advising.
- Required evidence: Attach product specifications, rights documents, local guidance, provenance records or the previous approval.
- Rollback and expiry: Record when the offer, right or approval expires, and which version to restore if the release fails.
- Batch boundary: Define the channels, markets, languages, formats and dates covered by the decision. Anything outside the boundary returns to the queue.
The purpose is not another form. It is to let reviewers inspect what changed. Without a change statement, every reviewer starts again from zero. Without a batch boundary, a narrow approval can quietly spread to markets and uses it never covered.
Price the approved decision, not the generation click
A practical comparison metric is:
Cost per approved creative decision = (generation + preparation + review + correction + audit retention) ÷ approved versions with a recorded decision
Preparation includes collecting differences, product data and rights evidence. Review includes specialist time and queue delay. Correction includes regeneration, re-editing and a second review. Audit retention includes the approver, timestamp, version, basis and expiry. This is not an industry benchmark and does not imply a universal price. It is a way to compare suppliers and operating models on the same denominator.
A workflow that produces 500 variants but can clear only 80 before launch may be less useful than one that produces 120 variants with complete change notes and evidence. The unit that reaches market is not a generated file. It is a version approved in time and recoverable later.
When to buy review capacity before more generation
- Automate generation first when inputs are locked, changes are mechanical, failures are reversible and sampling can expose a bad batch quickly.
- Add review capacity first when the existing brand, product or local-market team cannot keep up and queue time is delaying campaigns.
- Improve the review package first when the same questions return repeatedly. The missing resource may be a clear change statement, not another approver.
- Stop release first when product truth, consent, rights or local legal interpretation is absent. More output cannot compensate.
Rewrite the brief around decision time
The next AI creative or multi-market video brief can still specify asset counts, but it should add four acceptance criteria: expected volume in each risk lane, the owner of each decision, target time from submission to decision, and a reason plus rollback record for rejected versions. The production partner should deliver not only files, but reviewable change packages.
For production companies, this changes the margin logic. Mechanical green-lane work becomes template-driven and increasingly automated. Higher-value work lies in organising product truth, brand judgment, rights evidence and local-market context so that fewer decisions are ambiguous. The cheaper generation becomes, the less valuable an unmanaged pile of variants is—and the more valuable a trusted path to release becomes.
A practical next step: take 20 variants from the next multi-market batch. Assign each to a lane, then record who each version is waiting for, how long it waits and why it is returned. Find the point where λ exceeds μ before buying another generation tool. The right intervention may be another reviewer, a locked template, a stronger evidence package, or simply fewer variants with no distinct decision value.
