A sales deck is not a video brief: build a source gate for AI video

Treat a document-to-video upload as source material, not a brief. Before it enters a generator, turn the document into four controlled outputs: a current source of record, an approved claim set, visual boundaries, and an accountable owner. The model can turn a spreadsheet, PDF, or slide deck into motion. It cannot know whether a number is still valid, whether a screenshot was licensed for public use, or whether a draft promise made it into the final contract.

On August 24, Alibaba Cloud launched Wan3.0 with support for DOC, XLS, PPT, PDF, and Markdown inputs, alongside clips up to 30 seconds.[1] That changes the production entry point. A team can now drop a product deck into a video tool and get an explainer-shaped result without first writing a conventional script.

That shortcut is useful for internal exploration. It is dangerous when the file is treated as publication-ready truth.

A slide deck often contains old pricing, speaker notes, copied competitor language, screenshots from a beta, a diagram that only made sense in a meeting, and a sentence written to provoke discussion rather than make a public promise. A spreadsheet can carry totals without the definition behind them. A PDF can be approved as a leave-behind but not as a public campaign claim. Video gives all of that more confidence than a bullet point ever did.

The production problem is not that document-to-video models will make bad-looking clips. It is that they will make unreviewed material look deliberate.

The document became a production surface

A conventional video workflow forces a few healthy pauses. Someone writes a script. Someone marks the claims that need proof. A producer checks visual references and clears what can be shown. The editor gets a shot list. The client sees a cut.

A document-first workflow can skip straight from an internal file to a coherent visual sequence. The missing pauses do not disappear; they move to the end, where they are more expensive. Teams discover the wrong logo in a generated scene, a discontinued feature in a voiceover, or a country-specific statement in the version going to every market. Then they regenerate, re-edit, re-caption, and re-approve.

We do not solve that by making people write longer prompts. We put a source gate in front of generation.

This is distinct from the web asset contract. An asset contract decides whether a finished image or clip belongs in a page, how it must crop, what it means, and how it loads. The source gate sits earlier. It decides what the generator is allowed to say or show in the first place.

The Mintec source gate: four outputs, not one uploaded file

The gate should take 20 minutes for a simple product video and longer only when the source is genuinely complex. The goal is not to rewrite every document. It is to produce a small, reviewable package that a creative lead, editor, or model can use without guessing.

The output is a brief, but it is a brief derived from a document with its ambiguity removed. That difference matters.

1. Establish the source of record

Start with a single file or a small, explicit bundle. Label it with version, owner, date checked, market, and valid-until date. If the source is a deck plus a pricing sheet plus a product release note, say so. If two documents conflict, do not let the model average them into a confident middle ground. Resolve the conflict before generation.

This rule sounds bureaucratic until a launch moves by a week or a pricing page changes. A document generator does not know that an old slide has been superseded. It sees it as equal context. Giving it a clean source of record is less about model reliability than production accountability: six weeks later, someone must be able to answer what the published video was based on.

For modular sites and headless content models, this creates a useful pattern. Store the public claim, scope, owner, and expiry as structured content where the site can expose and retire it intentionally. Keep the long deck as supporting evidence, not as the one object from which every channel improvises.

2. Turn statements into an approved claim set

A model does not distinguish between a fact, a forecast, an aspiration, and a sales phrase unless you do that work in the input.

Break every candidate statement into one of four states:

• Approved fact: it can be said directly. Include the source section and any required qualifier.
• Approved interpretation: it can be dramatized, but it cannot become a new factual claim. "Cuts handoffs" may support a scene of a shorter workflow; it does not support a promise of a specific percentage.
• Draft or disputed: it stays out of the generation context until an owner resolves it.
• Restricted: it may be true, but it cannot appear publicly, in this market, or in this format.

Write the claim in the language the video will use. "Up to 30% faster" without a baseline, customer profile, or scope is a request for a misleading visual. "In the internal pilot, the review queue fell from 24 hours to 17 hours for the Mexico team during May" has enough shape for an approver to accept, rewrite, or reject.

This is also where teams should stop asking a visual model to imply outcomes that nobody would approve in text. A generated customer celebration, a dramatic before-and-after dashboard, or a scene where a clinician acts on a recommendation can smuggle a promise into the edit without speaking it aloud.

3. Set visual boundaries before the model creates evidence

The source package must say what visuals are authorized, what may be illustrative, and what must not be fabricated.

For a product explainer, that could mean: use the approved UI capture from build 4.2; show generic hands and office environments; do not render customer logos; do not create performance charts; do not depict regulated use cases; leave text-free space for final copy; use an illustrated data-flow scene rather than a fake product screen.

The boundary keeps the team from treating a plausible image as proof. It also gives the editor a better starting point than a naked PDF. Single-take video models make narrative continuity cheaper, but that same continuity can carry one bad implication through 30 seconds of polished motion.

Separate visual direction from evidence. A scene can communicate momentum, friction, or collaboration. It should not silently invent a customer, a clinical result, a dashboard metric, or a product capability that the source does not authorize.

4. Name the release owner and the stop condition

Every source gate ends with a human who can say yes, no, or not yet. The owner is not the person who uploaded the file. It is the person able to approve the particular risk: product for feature accuracy, legal or compliance for a restricted claim, regional lead for a market promise, and brand for public identity.

Also write the stop condition. Examples: "Do not render until the pricing table has a confirmed effective date." "Do not show product UI until the release branch is live." "Do not reuse customer marks outside the approved account list." These are not prompts. They are constraints the workflow must obey.

Provenance helps, but it is not a truth engine

Once the source gate passes, record the model, prompt package, source version, and approved cut alongside the output. This makes the work auditable and makes later replacement possible. Content Credentials can carry signed provenance assertions for an asset and preserve the history of changes across a workflow.[2]

They do not validate the business meaning of the source. The C2PA specification is clear that its systems should not make a value judgment about whether provenance data is "good" or "bad"; they validate that assertions belong to an asset and have not been tampered with.[2] That is exactly why the source gate exists. Provenance can show what generated the clip. It cannot decide that an old sales deck became correct when it was turned into video.

Our provenance-first tool-selection framework covers the file-level side. The operational rule is simpler: validate the document before you record the video made from it.

Run the gate on one real document this week

Pick the PDF or deck your team uses most often. Do not ask for a finished video first. Create the four outputs: current source, approved claims, visual boundaries, and release owner. Give the generator only that package, then compare the result with the clip produced from the raw document.

The difference will not always be more cinematic. It will be easier to defend, easier to localize, and far less likely to require a quiet correction after it has already been published. Document-to-video is valuable because it lowers the cost of turning internal knowledge into media. That only works when the internal knowledge is made fit for public production before the model starts moving it.

Sources

[1] https://technode.com/2026/08/24/alibaba-launches-wan3-0-video-model-with-30-second-generation-and-document-input — TechNode: Alibaba launches Wan3.0 video model with 30-second generation and document input [2] https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html — C2PA Technical Specification 2.2