Can a Two-Person Team Ship AI Video Ads Every Week?
Video models keep multiplying and some quietly shut down — build a creative pipeline, not a single clip.
Two opposite things are happening in AI video right now, and small teams should read them together. First, the menu got longer: Google's Veo 3.1 is being distributed through Google Vids, the Gemini app, Google AI Studio, the Gemini API and Vertex AI, while commercial options like Kuaishou's Kling 3.0 Omni, Runway Gen-4, Pika and Seedance sit alongside open models such as Wan, LTX and HunyuanVideo that you can run yourself. Second, the menu also shrinks: OpenAI announced the shutdown of the Sora consumer app and API, the app is already gone, and the API is scheduled to sunset on September 24.
Put together, the lesson is simple. Video generation has moved from "which app makes the prettiest clip" to "how is our creative pipeline built". One beautiful render is not an asset; a repeatable process that survives a model swap is.
What do small teams actually get out of AI video?
The realistic win is not a cinematic brand film. It is the speed of producing many variants of the same product every week. Platforms are moving that way too — Amazon Ads now offers advertisers a generator that turns static product images into video ads — so the advantage shifts from editing skill to deciding what to make and in how many versions.
- Rotating and close-up shots for product pages, without rebooking a shoot
- 9:16 short-form cutdowns and captioned versions of the same asset
- 30-second feature explainers you replace whenever the product changes
- Two or three alternate opening hooks for A/B tests, differing only in the first three seconds
- Short recruiting or company-intro videos that were never worth an agency invoice
A two-person leather bag brand, for example, can skip a new photoshoot: start from one existing product photo and generate a rotation shot, an unboxing scene and an in-use scene, then test three creatives in a single week.
B2B SaaS should tilt the other way. Record the real product UI, because accuracy matters there, and use generation only for the "decorative" parts — the opening three seconds, backgrounds, transitions. Treating AI video as a gap-filler rather than a full-service producer keeps expectations honest.
Is it safe to bet on one model?
No. The Sora API sunset is a plain reminder that a video API you rely on can be announced away in a matter of months. Treat models as swappable parts and keep the creative assets and specs on your side.
- Write a one-page brand spec first: product colors, logo placement, banned depictions, aspect ratios, caption font.
- Version your prompts, reference images and seeds in a repo or a sheet, so a model change still has a starting point.
- Download master files to your own storage instead of leaving them inside a vendor's library.
- If you call an API from code, wrap it in a thin adapter so switching providers is a config change.
- Keep one live alternative at all times — one commercial model and one open model is enough.
- Check the terms for commercial use scope, rights to outputs, and how much notice you get before shutdown or pricing changes.
This is a one-day task, and skipping it means rebuilding six months of creative from scratch the day a deprecation notice lands. If you are currently calling the Sora API, its end date falls this month, so put alternative testing and asset backup on this week's calendar.
What should you do this week?
Seven days is enough for a first pipeline. Do not install a pile of new tools — pick one product and run it through the full loop.
- Days 1-2: choose your best-selling product and write the one-page brand spec.
- Day 3: run the same prompt through two or three different models, four shots each.
- Day 4: score pass/fail on product consistency, hand and logo distortion, and broken on-screen text.
- Day 5: put the two survivors behind a small ad budget or at the top of the product page and watch the response.
- Days 6-7: freeze the winning prompt as a template and reuse it on the next product.
Two cautions before you publish. Avoid scenes that resemble real people's faces or voices and anything carrying another brand's marks, and check each platform's AI-disclosure policy before you run paid media. Video models will keep rising and disappearing. What stays is your team's process for knowing which product, which spec, and how many versions.