An AI commercial is usually far cheaper and faster than a shot one: a traditional social video ad typically costs $3,000 to $15,000 and takes 2 to 6 weeks per batch, while AI-made ads are documented in the tens to low thousands of dollars and days of work. Traditional production still wins when you need real people performing, real product texture, or legal certainty about who owns what. For most product launches, the best answer is a hybrid.
AI commercial vs traditional production at a glance
| AI-generated commercial | Traditional production | |
|---|---|---|
| Cost per ad | $30 to $530 in invideo's documented runs; $500 to $1,200 for a fixed-price studio package | $3,000 to $15,000 per video (NoHo Arts) |
| Timeline | Days (the Kalshi ad took two) | 2 to 6 weeks per batch |
| Iterations | Cheap: regenerate a shot | Expensive: reshoot or live with it |
| Product fidelity | Labels, logos and proportions drift | Exact, because it is the real thing |
| People | Faces and hands are the weak spot | Real performance, real emotion |
| Physics | Liquids, fabric and hands often fail | Real, or controlled CGI |
| Rights | Uncertain for purely generated output | Clear chain through contracts and releases |
| Disclosure | Realistic output may need an AI label in the EU | None for real footage |
Line-by-line pricing: what AI video ads cost.
Cost and speed: where AI clearly wins
The NPR story on the Kalshi ad is the cleanest public data point. P.J. Accetturo made it with Google's Veo 3 in two days, and Kalshi said the cost of prompting "was under $2,000," not counting his fee (WUNC/NPR). Accetturo told NPR that the spots it aired alongside likely cost hundreds of thousands of dollars and took months to produce.
The bigger win is iteration. NoHo Arts points out that even winning ads fatigue after 2 to 4 weeks (NoHo Arts). If you need a fresh set of hooks every month, a shoot per refresh is hard to justify.
Control and quality: where it gets complicated
On set, a director says "slower, hold on the cap." With an image-to-video model you describe the move and choose from what comes back: you direct by selection. That suits camera moves, light and abstract worlds, and fails when one detail must be exact, like the embossing on a bottle.
Our fix: keep real packshot pixels wherever the product is legible, and let AI handle the world. Most of the "AI look" comes from asking models to do what they are worst at; see why most AI commercials look bad.
Talent and people
A real actor gives you a performance. A generated person gives you a likeness that may slip between shots, and under the Commission's final Article 50 guidelines, fully AI-generated actors are likely to count as deepfakes that need disclosure (Mishcon de Reya). If your story depends on a human face, shoot the human.
Rights: the least discussed difference
With a shoot, you own the footage through contracts with crew, talent and location. With AI, it is murkier. The US Copyright Office concluded in January 2025 that "prompts alone do not provide sufficient human control" for the output to be copyrightable, though human selection, arrangement and expressive inputs can be protected (Sidley summary). So your edit, titles, packshot and script are yours; a single generated clip may not be. Read each tool's terms too.
Where traditional still wins
- Dialogue and testimony. A founder or customer speaking on camera.
- Food, cosmetics textures and fabric drape where viewers judge the real thing.
- Regulated categories (alcohol, health, finance) where any "is this real?" doubt is a risk.
- Hero brand films for TV with long shelf life and big media spend behind them.
- Products with mechanisms that must move exactly as engineered. That is a job for CGI, not generative video, as we explain in CGI product ads.
The hybrid approach we would recommend
- Capture the product true. A half-day packshot session, or a 3D model if you already have CAD.
- Generate the world. Use an image model for keyframes of the environment and light, with the real product composited in.
- Animate between keyframes with start/end-frame morphs in an image-to-video model, so each clip ends on the frame the next one starts on.
- Shoot people for real if the film needs them, and keep AI away from their faces.
- Edit traditionally. Titles, pacing and sound in the edit, cut for 9:16 first.
- Label what is realistic and generated, before the platform does it for you.
You keep the cost and speed advantage on most of the frames and spend real money only where the viewer would notice: the product and the people.
