An overseas brand may want one film for readers, another for home workers and a third for design-conscious shoppers. If the product comes from the same Chinese supplier, however, the three stories still need to describe the same real item.

Before buying a batch of personalized AI video ads, approve the permitted changes—not just the number of exports. ONCE recommends including cross-version review and an ordinary, non-personalized fallback in the pilot scope.

A research signal at the script layer

The MAPS preprint, submitted October 19, 2026, proposes a multi-agent framework connecting profiles, scripts, video synthesis and feedback. Its script module receives user preferences and product attributes; the training objectives cover structure, personalization and product alignment.

This is a researcher-reported system, not evidence that every brand can buy the capability today. Its performance figures are not adopted here, and it is not an ONCE implementation. The procurement implication is our analysis: once the script itself varies, creative freedom and permission to make product claims need separate treatment.

Change the reason to care, not the underlying facts

Consider a fictional desk-lamp brand. A verified adjustment feature could be introduced through reading, desk organization or interior styling. The entry point, characters and rhythm can change. None of those choices authorizes a new health benefit, a different power arrangement or the disappearance of a relevant installation condition.

Separate information that must remain accurate across the set from information a particular story may emphasize. The first category need not use identical words, but its meaning and qualifications must survive. The second can respond to reliable audience information that the project is authorized to use.

If the product information is uncertain, resolve it before expanding the batch. More audience labels will not establish a missing fact. Fluent, audience-appropriate language cannot validate an unsupported promise.

ONCE’s same-question review

Choose a consequential product question: how is it powered, what installation is needed, or under which conditions does a demonstrated result apply? Review the answer across the entire batch, rather than approving each film in isolation.

For each question, retain four items:

  1. The approved fact and its source.
  2. The meaning that cannot be lost across versions.
  3. The expression allowed to vary.
  4. Its location in the script and actual delivered film.

This is an original review method, not the paper’s experiment or a platform template. It does not require every ad to include the same speech. It requires the team to explain why a changed story still communicates accurate information appropriate to its intended use. A fleeting disclaimer cannot repair a main visual conveying the opposite message.

Review words and images together. A script can accurately mention angle adjustment while a generated shot invents a telescoping mechanism. If a tool cannot reliably preserve a necessary action, consider approved live-action footage, product photography or a different presentation. An all-AI production label is not a reason to keep a false demonstration.

Define the version for an unknown viewer

Idomoo Academy’s current personalized-video script guide advises checking data availability, reliability and permitted use, and accounting for iterations that contain different information. Its engagement figures are not used here.

ONCE’s proposed addition is an approved ordinary version for missing, conflicting or unsuitable preference information. Do not disguise a default value as knowledge about a viewer. This is a suggested delivery requirement, not a claim that MAPS already provides this feature.

Use clearly simulated inputs in a pilot: first omit the preferences, then provide conflicting ones. Ask the supplier to identify the uncertainty and follow the agreed fallback while preserving product facts. Do not upload personal viewing histories or customer records to an unapproved tool for this exercise. Actual data handling requires its own authorization and arrangements.

Price the meaningful changes

Start with one well-documented product, a few genuinely different stories and an ordinary version. Scope information preparation, narrative design, generation or filming, cross-version review and delivery records separately. Specify what needs reconfirmation when language or market changes.

Changing a salutation alone does not demonstrate meaningful creative adaptation. Conversely, new characters, actions and locations may involve different production work even when runtime stays the same. No assumed hours, industry prices or performance guarantee are offered here.

The potentially durable asset is the permitted relationship between product facts, variable expression and approved versions—not a collection of personal profiles acquired without permission. For a continuing service, its maintenance work and economics still need to be tested in real projects.

The next brief should answer three questions before specifying output volume: which changes matter, which meanings must remain, and what should play when preferences are unknown?

Sources checked October 26, 2026. Peer-review status was not established for the MAPS preprint; a precise publication date was not established for the current Idomoo guide. Examples and review methods are proposals, not client results. The original AI cover is conceptual.