When one system can move budget across sponsored ads, display, online video and streaming TV—and generate new assets from product information—the most important creative change is not “make more sizes.” It is redefining the unit of production procurement.
Instead of commissioning a disconnected film for each channel, brands should build verified, licensed evidence modules that can be approved for different funnel jobs. The practical question is no longer whether AI video is cheaper than a shoot. It is which product truths can safely be generated, which need to be filmed, and how many real deployments each production investment can support.
What Amazon has actually confirmed
On September 29, 2026, Amazon Ads announced that Full-Funnel Campaigns had entered open beta for US advertisers that sell on Amazon. An advertiser provides products, a budget and an optional ROAS threshold. Amazon’s system can then coordinate sponsored ads, display, video and streaming TV through one campaign, selecting formats and allocating budget across the funnel.
Amazon also says built-in tools can generate images and TV-ready video when more creative is needed, with advertiser approval before launch. Its broader Amazon Ads Agent announcement describes shared signals across formats and unified reporting that includes impressions, clicks, sales, new-to-brand and long-term sales metrics.
Those facts establish the product direction and US availability. They do not prove that every category or advertiser will reproduce Amazon’s beta results.
Long-Term Sales needs another boundary. Amazon defines it as an ad-attributed estimate of immediate sales plus the projected value of new-to-brand customer engagement over the next 12 months. The projection uses historical outcomes after actions such as detail-page views, branded searches, add-to-carts and purchases. It is a useful planning metric; it is not cash already realized in a brand’s P&L.
The creative-to-media interface is now the operating problem
In a coordinated campaign, streaming TV is not isolated “awareness” and sponsored ads are not a separate last-click machine. A clear idea introduced at the top of the funnel may help lower-funnel assets work; product-page accuracy and conversion signals can influence what the system chooses next.
Traditional channel-by-channel delivery breaks in predictable ways. The TV film communicates a brand mood while the retail ad suddenly becomes a list of specifications. The product page shows a new pack while the video contains an old one. A platform can generate variants, but the brand has never recorded which claims, people, music, markets or edits are permitted.
The answer is not unlimited content volume. It is a creative supply system that both software and humans can understand.
The ONCE Full-Funnel Evidence Map
Organize one shoot or source library into six asset classes. Every module should carry its relevant ASIN, market, language, rights window and approval state.
1. Recognition modules
Show product identity, brand cues, current packaging, category and a use result that remains legible in a small frame. Recognition gives every format a stable identity; it does not need to carry every benefit.
2. Situation modules
Show why the product matters now: noise during a commute, limited counter space, a power failure during an outdoor trip. The situation should be understandable as an image sequence, not only through voice-over.
3. Proof modules
Turn a claim into an observable action. If setup is faster, preserve the installation sequence and completed state. If two tasks can run together, keep both visible under comparable conditions. Film the condition under which the claim holds, not only a polished result.
4. Objection modules
Answer the last practical question before purchase: Will it fit? Is it difficult to clean? Which accessories are included? Is the interface understandable? These modules often belong closer to consideration or product-detail environments than inside a single brand narrative.
5. Retail-continuity modules
Record the correct ASIN, pack, color, bundle, accessories, hero image and destination page. Leave volatile price, promotion and inventory information to live retail components where possible instead of burning it into a long-life master.
6. Rights and metadata modules
For every shot, record provenance, talent releases, music rights, permitted markets, expiry, prohibited claims, crop tolerance and whether generative alteration is allowed. Technical editability does not equal commercial permission.
Use three production routes—not an AI-versus-live-action argument
| Product and risk condition | Recommended route | Human approval focus |
|---|---|---|
| Accurate product page, simple benefit, low claim risk, urgent format gap | Generate first and use it as low-cost exploration | Product geometry, copy, pack, voice and destination continuity |
| Function needs demonstration, shoppers have clear objections, or footage must serve several markets | Film an evidence master, then generate or crop controlled variants | Action continuity, claim conditions and version metadata |
| Premium launch, strong brand story, regulated or safety-related category | Human-authored master with claims and scenes locked against unapproved alteration | Legal review, performance validation, talent/music rights and market limits |
Generation is not automatically good or bad for a brand. If the product page is inaccurate or the claim lacks evidence, automation reproduces the problem faster. If product structure is stable, the message is simple and every output can be checked, generation can reduce the cost of filling format gaps.
Amazon’s own AI creative solutions reflect this spectrum: advertisers can create from an ASIN, work with a conversational creative partner or exercise more granular control. A partner perspective published by Global Overview in June 2026 makes a useful adjacent point: AI video generation is a tool, not a strategy. The production decision still depends on the business question and the available source evidence.
Recalculate production economics by approved deployments
Comparing the price of one finished film makes generation look permanently cheaper. Counting export files creates a different error: ten variations of the same unverified claim look like ten assets.
ONCE recommends an internal measure:
Fully loaded cost per approved deployment = (strategy + capture/generation + product validation + rights + modular post + localization + QA + rework) ÷ approved funnel-stage × format × market × product combinations.
The denominator is not the number of files rendered. A combination counts only when product representation is accurate, the claim is supported, rights are valid and the destination matches. A master that can safely support US streaming TV, online video and two product-detail variants may be more valuable than ten cheap outputs whose sources cannot be explained.
This model also creates a cleaner production quote. Separate fixed evidence-creation cost from the incremental validation and adaptation required for each additional market, format or product. The client can see what creates new proof and what safely deploys existing proof.
Move approval upstream of the launch deadline
An operational workflow begins with inputs. The brand freezes the relevant product page, claim list, target markets and prohibited changes. Production delivers modules, parent-child version relationships and rights metadata. Media then uses platform generation or assembly. Product, brand and compliance owners review against the same approval sheet.
At minimum, use three controls:
- Input approval: Is the ASIN, pack, claim and product page current?
- Assembly approval: Did generation alter structure, usage conditions or brand voice?
- Activation approval: Does the exact ad match its destination, market, language and inventory state?
For teams building broader Amazon video systems, ONCE’s guide to creating Amazon DSP video assets for CTV, online video and retargeting provides the adjacent format-planning context. The present framework adds the approval and evidence layer required when one optimizer can connect those formats.
When one finished film is still enough
A single, purpose-built film can still be rational when the budget is small, there is one product, one market and a short campaign window, and the asset will not enter automated recombination or a long-term library. Modularity has an operating cost; not every project needs it.
But once a campaign spans several funnel stages, formats or markets, buying separate finished films by channel creates repeat shoots, claim drift and rights blind spots. The first priority should then be the evidence map and approval status—not a larger edit count.
Amazon has confirmed a platform change: media orchestration, creative generation and long-term value estimation are moving into one environment. ONCE’s inference is that procurement must change with it. The most valuable production partner will not be the one that delivers the most files. It will be the one that makes every product claim filmable, reviewable, recombinable and easy to retire when its conditions change.
Sources and limits
Core sources checked for this article include Amazon Ads, “Grow total sales with Full-Funnel Campaigns” (September 29, 2026); Amazon Ads, “Amazon Ads introduces Full-Funnel Campaigns, DVA+, and enhanced chat experience in AI-powered ad platform, Amazon Ads Agent” (September 29, 2026); Amazon Ads, “Understand the value of new-to-brand shoppers beyond immediate sales”; and Amazon Ads’ AI Creative Solutions page. Global Overview’s June 11, 2026 Partner Perspective was used to compare a peer view of AI-video production economics, not to establish adoption or guaranteed performance. The Full-Funnel Evidence Map, route table and cost model are ONCE analysis.
