Stop Making Fake AI Ads. Cut Real Demos. (2026)

Stop Making Fake AI Ads. Cut Real Demos.

Google's Omni 1.2 Flash can generate or edit clips from 3 to 15 seconds. At $0.07 per second for 720p, a 15-second clip costs $1.05.

Step 1: Start With Real Product Footage

Use a real screen recording as your source.

Skip generated dashboards and cinematic laptops floating through glass offices. Show the product doing something the buyer cares about.

Google says Omni 1.2 Flash follows references better than Omni 1.1. It can keep packaging, colors, faces, clothing, and typography consistent across a clip. Still, "better" doesn't mean "exact."

Product truth matters more than visual polish.

Build a demo library around use cases:

• Creating a sales report
• Approving an invoice
• Finding a support ticket
• Launching an email sequence
• Updating a CRM record
• Exporting compliance data

Each recording needs five pieces of metadata:

1. Product version 2. Feature name 3. Approved claim 4. Start and end timestamps 5. Expiration date

That last item matters. Your interface will change. Don't send a clip of the 2025 dashboard after the 2026 redesign.

Give every source file an ID. Store the transcript with it. Add approved product claims from your site, help center, or sales enablement docs.

Your AI agent shouldn't decide what the product does. It should choose from approved evidence.

This turns the job into controlled assembly instead of open-ended content generation.

A recent study on AI-assisted marketing video made the same point. Product clips need human checks for shape, color, included items, and real product behavior. The technology can speed up production. It can't take responsibility for factual accuracy.

Step 2: Match One Clip to One Buyer Problem

Most personalized videos aren't personal.

Adding "Hey Sarah" to a generic demo is mail merge with a play button.

Personalization should change what you show.

A CFO might get a 15-second clip about cost reporting. A sales leader might see pipeline alerts. An operations manager might see ticket routing.

The buyer's role determines four things:

• The feature shown
• The approved claim
• The caption
• The call to action

Use CRM data to create a short clip brief. HubSpot, Salesforce, or Pipedrive can trigger the workflow after a call, email reply, or demo request.

A structured brief might look like this:

```json { "contact_id": "crm_28491", "persona": "VP of Sales", "buyer_problem": "slow lead response", "approved_feature": "instant lead routing", "source_asset_id": "demo_lead_routing_v4", "source_timecode": "00:42-00:57", "approved_claim_id": "claim_118", "cta": "Want to see this on your routing rules?" } ```

The AI gets a set of choices. It doesn't get creative freedom.

We use n8n for this kind of workflow. It can pull the CRM record, find the approved source, create a video job, pause for review, and write the result back to the contact.

This idea has been around for years.

Variable-data printing took off in the 1990s because marketers could change names, offers, and images for each recipient. Many teams handled it badly. They printed your name on a postcard and called it personal.

AI video could repeat that mistake.

The best teams will show the right proof to the right person.

Step 3: Cut First, Then Use Conversational Editing

Use predictable tools for the factual cut.

FFmpeg can extract the exact 15 seconds from your source. Descript can cut from a transcript. Adobe Premiere Pro can handle more complex sequences.

Use Gemini Omni 1.2 Flash after the cut.

Google says the model supports multi-turn conversational video editing. You can change the background, lighting, product reference, or aspect ratio without starting over. The edits can cover the full 15-second clip.

That helps with persona and channel versions.

It also creates room for mistakes.

Avoid prompts like "make the dashboard look more modern." That gives the model room to invent interface elements. Tell it to preserve the product area and change only approved parts.

Use a prompt like this:

```text Edit the supplied 15-second source clip.

Preserve every visible product interface element exactly. Don't add buttons, fields, charts, alerts, labels, or features. Don't rewrite product text.

Allowed changes:

• Crop to 9:16
• Add the approved caption below the product frame
• Add a neutral background outside the product frame
• Keep the original product timing

Approved caption: "Route inbound leads before they go cold."

Approved CTA: "Want to test this with your routing rules?"

Return a review draft at 720p. ```

Then use a second conversational edit for one isolated change:

```text Keep the product frame, timing, captions, and CTA unchanged.

Change only the background from light gray to approved brand navy: #10243E ```

That is the right use of conversational video editing. Make small changes with clear boundaries. Don't write a fresh creative brief on every turn.

Google prices 720p output at $0.07 per second. One 15-second draft costs $1.05. A draft plus two full-length edit passes costs $3.15 in model output fees.

Create drafts at low resolution. Approve one. Upscale the approved clip.

Step 4: Put Approval and Versioning in the Workflow

A human review needs more than a Slack message that says, "Looks good."

Tie approvals and version history to the actual asset.

Every job should store:

• Source file ID
• Source file hash
• Source timestamps
• Transcript version
• Approved claim ID
• Prompt version
• Model ID
• Output URL
• Reviewer
• Review time
• Approval status

Google identifies the model as `gemini-omni-1.2-flash`. Save that with each result. Models change, and the same prompt can produce different output after an update.

Your release check should answer five questions:

1. Does the clip show a real feature? 2. Does the caption match the approved claim? 3. Did any interface text change? 4. Did the model add a control or workflow? 5. Is the call to action accurate?

The agent can run automated checks first. OCR can compare interface text. Frame comparison can flag visual changes. A transcript check can catch altered claims.

A person should approve the final clip.

The following is an example flow, not a copy-and-paste Gemini API contract. Google can change field names, upload rules, and endpoints.

```javascript const job = { model: "gemini-omni-1.2-flash", duration_seconds: 15, resolution: "720p", source_video_uri: source.uri, prompt: approvedPrompt, metadata: { contact_id: contact.id, source_asset_id: source.id, source_hash: source.sha256, claim_id: claim.id, prompt_version: "persona-demo-v3" } };

const draft = await submitVideoJob(job);

await saveVersion({ version: 1, status: "PENDING_REVIEW", output_uri: draft.output_uri, ...job.metadata }); ```

Google says Omni outputs include an imperceptible SynthID watermark. It helps track where a clip came from. It doesn't prove your product claims are true.

Watermarks track origin. Approvals protect trust.

Step 5: Send the Clip When Intent Is Fresh

Don't attach these clips to every cold email.

That turns a useful tool into automated noise.

Send them after a signal:

• The prospect replied
• The prospect asked about a feature
• The prospect attended a demo
• The prospect visited a product page
• The prospect opened a proposal
• A deal moved to a new CRM stage

Waymark uses a similar pattern for custom video. Its system creates a video when a lead enters the CRM and alerts sellers when someone watches. Across 412 reported wins from January through July 2026, 79% of sellers said the finished video helped them close faster or save time.

That's vendor-reported data, not a universal benchmark. Still, it supports the workflow.

Creating the clip is only half the job. Timing matters too.

Write the approved clip URL back to HubSpot, Salesforce, Outreach, or Salesloft. Add the clip ID and version to the activity record.

Then track:

• Clips generated
• Clips approved
• Approval time
• Videos watched
• Replies after watching
• Meetings booked
• Opportunities influenced
• Cost per approved clip

Don't claim a reply-rate lift before you measure it.

Start with 100 qualified follow-ups. At one 15-second generation each, the base 720p model cost is $105. If only 60 pass review, the model cost per approved clip is $1.75 before storage, workflow, and labor.

Compare that cost with seller time.

If a rep spends 10 minutes finding, cutting, captioning, uploading, and logging each video, 100 clips take more than 16 hours. The model fee matters less than the selling time you get back.

VCAT reported a larger production example with Naver Shopping Live. AI-assisted production cut the process from 30 days to under three days. VCAT also reported costs near one-tenth of the traditional approach.

Keep video teams focused on work that needs their skill. Let the system handle repeat edits.

The Real Opportunity Isn't Fake Ads

Google's demo reel will show polished scenes, perfect lighting, and camera moves.

Your sales team needs something less exciting.

It needs a truthful clip that shows the exact feature discussed five minutes ago.

StoryPros builds AI agents for sales, marketing, and operations. The systems that work retrieve approved information, take a narrow action, and stop when human judgment is needed.

Gemini Omni Flash 15-second videos are cheap enough for that workflow.

Use the model as an editor. Keep your product as the evidence.

FAQ

How long can Gemini Omni Flash videos be?

Gemini Omni 1.2 Flash can generate and edit clips from 3 to 15 seconds. Google allows duration selection in one-second increments.

How much does Gemini Omni Flash cost per clip?

Google lists Gemini Omni 1.2 Flash at $0.07 per second for 720p video. A 15-second clip costs $1.05 before storage, upscaling, workflow tools, and review labor.

Can Gemini Omni Flash edit videos through prompts?

Yes. Google says Omni 1.2 Flash supports multi-turn conversational video editing across full 15-second clips. Users can request changes to the background, lighting, weather, product reference, or aspect ratio without restarting the job.

How do I call the Gemini Omni Flash API?

Use the paid Gemini API or Google Cloud with the model ID `gemini-omni-1.2-flash`. Upload or reference the source asset, submit the edit prompt and output settings, poll the job, then store the result with its source ID, prompt version, and approval status. Check Google's current API docs before coding because upload limits and request fields can change.

How do I prevent Gemini from inventing product features?

Start with real product footage. Limit edits to approved regions, captions, crops, and backgrounds. Compare output frames and OCR text with the source, then require human approval before the clip enters your CRM sequence.

Related Reading

• Best AI Marketing Agency (2025–2026): Use-Case Shortlist, Pricing, and 30-Minute Decision Framework
• OpenAI's Decisions API Is a Fast Route for Agents (2026)
• 5 AI Pitch Deck Generators Compared: Real Pricing and CRM Limits (2026)
• Stop Personalizing. Start Targeting. (2026 Cold Email Guide)