The familiar video brief—one 30-second hero film plus a few cutdowns—was built for a world in which people decided the media plan and assets stayed relatively fixed.

That brief becomes brittle when the media system can choose formats, move budget, generate missing creative and keep optimizing. A locked film is then the least composable thing a brand can give to the system that is best at recombination.

On September 29, 2026, Amazon Ads introduced Amazon Ads Agent and Full-Funnel Campaigns. Full-Funnel Campaigns coordinates sponsored ads, display, video and streaming TV through one budget and one optimization engine. In the United States, the feature is available in open beta to advertisers that sell on Amazon. Brands provide products, budget and an optional ROAS threshold; Amazon’s AI selects formats, allocates spend across the funnel and continuously optimizes. When more assets are needed, built-in tools can generate images and TV-ready video for advertiser approval.

Two availability claims should not be collapsed. The Full-Funnel Campaigns beta is open to eligible US advertisers that sell on Amazon. The broader conversational experience for planning, diagnostics and action remains a closed beta whose access varies by advertiser eligibility, marketplace, language and campaign type. Amazon’s announcement proves a product direction and a defined launch. It does not prove that every advertiser has the same controls—or that creative judgment has been automated away.

The procurement question is therefore not whether AI can make ads. It is what a brand must own before an agent is allowed to make, combine and optimize them.

A finished film still matters. It is simply no longer the full contract.

Streaming TV, online video, display and sponsored product ads do different jobs. A hero film can build memory. A demonstration shot can prove a feature. A clean product angle can identify the exact SKU. A short hook can earn the first second of attention. Captions, narration and offers have to travel across sound states, markets and languages.

When all of that value is trapped inside one mixed, subtitled and locked master, an automated platform can only crop it, reorder it or generate around it. Cropping can damage the story. Regeneration can amplify whatever is incomplete or outdated on the product detail page.

The stronger the platform’s generation capability becomes, the less useful it is for a production partner to compete on raw variant count. The durable production job is to create what the platform cannot infer safely: real product behavior, accurate materials, claim-bearing evidence, consistent brand identity, cleared rights and rules for how each component may be used.

This extends an earlier ONCE argument about why the brief becomes a core asset when platforms generate campaigns. The brief tells the system what it should know. The next question is what production should deliver so full-funnel optimization can happen without dissolving truth, rights or accountability.

The ONCE Agent-Ready Creative Contract

This is not an Amazon template. It is a six-field procurement method for the agreement between a brand, its production partner and the team operating automated media.

1. Outcome horizon

“Improve ROAS” is not enough. State whether the primary job is new-customer acquisition, immediate sales, branded search, repeat purchase or longer-term sales value. Add the budget ceiling, test window and stop condition.

The same asset can be valuable for discovery and weak for demand capture. If the objective is vague, an optimization engine can efficiently pursue an incomplete answer.

2. Product truth source

Bind each SKU to current specifications, approved claims, packaging version, offer dates, eligible markets, prohibited language and a named owner. The product detail page can be an input, but it should not be the only source of truth. It may lag a factory change or omit the conditions that make an advertising claim accurate.

The contract should answer a mundane but essential question: if the page, the package and the script disagree, which record wins and who updates it?

3. Proof-bearing shots

Separate atmosphere from evidence. Waterproofing, load capacity, dimensions, speed, installation steps, port locations and material behavior should link to live-action capture or an approved 3D asset, with the capture condition and product version retained.

Backgrounds, layouts, decorative motion and non-critical scenarios can be generated more freely. A frame that carries product proof should not become a plausible guess simply because the system wants more variants.

4. Module interfaces

Modular delivery does not mean dumping raw footage into a shared drive. Each component needs a job and an interface: opening hook, product recognition, core demonstration, use context, brand-memory cue, offer or action prompt.

Label a module ID, intended role, safe aspect ratios, audio dependency, text-overlay allowance, retiming allowance, linked SKU and prohibited combinations. Retain clean masters, caption files and audio stems so the next system can recombine approved parts instead of reverse-engineering a finished export.

5. Rights and no-go rules

Talent likeness, voice, music, locations, fonts, third-party marks and permission for synthetic modification must resolve to territory, term, medium and use. The no-go list should also cover generated product changes, nonexistent performance, unauthorized speech and assets cleared for organic use but not paid media.

Amazon offers manual, automated and mixed controls. None of those controls acquires rights on the advertiser’s behalf.

6. Approval and learning return

Advertisers still approve creative in Full-Funnel Campaigns. Internally, the brand should define which changes are mere formatting, which create a new claim, and which require product, legal or brand review.

Every launched version should retain its source modules, generation method, objective, market, format and approver. The post-campaign record should not stop at “this video worked.” It should preserve which proof, opening, audience and placement combination produced the observed result—and what the next shoot needs to add.

That is how platform learning becomes a brand asset instead of remaining a temporary answer inside a platform interface. Amazon Marketing Cloud can support privacy-safe analysis across pseudonymized signals and returns aggregated outputs; the brand still needs its own creative identifiers and decision record to connect those outputs to production choices.

Three buying modes

Platform-first

Use this for standardized SKUs with simple specifications, low claim risk, a complete product page and limited visual differentiation. Existing images and a small amount of video may be enough for the platform to generate and test variants. Product truth, rights and approval remain the advertiser’s responsibility.

Hybrid production

This is the practical default for many cross-border brands. A production partner captures product proof, the core story, brand identity and reusable master modules. The platform handles some adaptation, contextual generation, media orchestration and rapid testing. Budget concentrates on what must be true; automation reduces the cost of variation.

Controlled production

Complex durable goods, regulated categories, high-ticket products, dense technical claims or premium brand storytelling still justify tight human control over script, capture and key versions. The platform can support distribution and analysis without rewriting the proof logic.

A brand without a product-truth owner, rights record or creative approver should not scale generation merely because full-funnel automation is available. Build the input system first.

Price the approved activation unit—not the file count

AI can make variant counts explode, which makes “number of videos delivered” a poor value measure. A more useful scope equation is:

Cost per approved activation unit = (strategy and truth review + live action/3D/AIGC production + post-production + rights + module tagging + quality control + versioning and learning records) ÷ the number of modules approved for at least one named objective–format–market combination.

An activation unit is not any exported file. It has a defined job, valid rights, accurate product truth and traceable approval. One clean five-second demonstration that works safely across horizontal and vertical compositions may be more valuable than ten generated videos that change only the background.

Brands can track two additional signals: how many valid combinations reuse each module, and which modules remain unused after multiple rounds. The first reveals compounding assets. The second tells the team to stop paying for inventory without a job.

What this changes for brands and production companies

Amazon’s system does not eliminate production companies. It compresses the value of mechanical resizing, repetitive versioning and fixed-format delivery.

The emerging production role is to turn product knowledge into recordable proof, turn brand judgment into modules and rules, and return performance learning to the next creative cycle. The finished film remains what a customer sees. The contract increasingly needs to include the truth source, proof footage, module interfaces, rights, approvals and learning record behind it.

When a platform can run more of the funnel, the main risk is not that AI will do too little. It is that the system will do too much with incomplete inputs. The valuable purchase is no longer only a polished file. It is a creative asset system an agent can use without losing the facts and responsibilities the brand must still own.

Sources

This article turns public product documentation into a procurement method. Features, eligibility and regional availability may change; it does not guarantee access or campaign outcomes for any advertiser.

← Don’t Upscale the Whole Archive: Four Decisions Before You Modernize Brand Footage Google AI Max Can Show the Query, Headline and Landing Page. Now Fix the Proof Between Them →