AI video for marketing: what works, what it costs, what’s allowed

Three kinds of AI video, and what each is good for

“AI video” covers very different tools. Mix them up and you pick the wrong one and end up disappointed. The distinction that matters for marketing:

For most small teams, image to video is the best place to start: the starting image gives you control over person, light and style, and the model only adds motion. Avatars pay off when you regularly need to turn text into an on-camera video. More on avatars and AI actors in AI avatars and UGC ads.

Models and tools compared

The market moves fast. The table shows what vendors’ pricing pages said on November 10, 2026. Prices exclude tax, mostly in US dollars.

As of Nov 10, 2026. Google’s plan page describes Flow as offering “expanded access to Google’s creative models” without per-plan quotas.

When choosing, check three things: how billing works (credits, seconds, subscription); whether commercial use is included in your plan (Pika’s free plan, for example, has no commercial licence); and whether credits expire or roll over. Runway’s Standard and Pro credits, for instance, don’t roll over.

The workflow: from script to published video

A storyboard, not one long scene

Video models are strongest with short shots and one clear action. The longer a clip, the more likely faces, hands and backgrounds drift. So plan several short shots and assemble them in the edit, which also gives you control over pacing.

Plan for several takes

Expect that not every take will be usable. Budget several attempts per finished shot and compare the cost per usable clip, not per generation. With per-second pricing, a short, precise storyboard is your biggest saving.

Prompt anatomy: how to describe a video scene

A good video prompt describes not just what’s in frame but what happens and how the camera behaves. Six building blocks work well:

Add a negative list: no text in frame, no logo, no extra people, no distorted hands. And never describe a real, well-known person or a resemblance to one. That’s legally risky and against most vendors’ terms.

Same character, same style: consistency and faces

• Start from an image, not text. If you need a character across clips, generate a reference image first and animate it. Pure text produces a different face every run.
• Reusable characters. Some vendors offer features for recurring characters or elements; use them instead of re-describing every time.
• Lock the voice. If a character speaks, the voice should stay the same. Some models invent a new voice each run; cloned or fixed voices solve this but need the person’s consent.
• Check faces. If an AI character resembles a well-known person, regenerate it. That’s a fixed rule for us.
• A style skeleton. Light, colour, camera and framing as a fixed text block in every prompt.

Music, voice and rights

Most legal mistakes happen in the audio. Three rules:

1. Music only with a licence. Trending sounds from a platform’s library are often fine in organic posts but not automatically in ads. Use licensed music or commercially cleared libraries for advertising.
2. Voices only with consent. Cloning someone else’s voice without their explicit permission is banned by reputable vendors and can violate personality and publicity rights.
3. Label AI voices that sound real. In the EU, realistic AI audio that could pass as authentic falls under the AI Act’s disclosure duty.

Labelling and the rules

Advertising law applies too: an AI character must never pose as a real customer, and ads must be recognisable as ads; in the US the FTC’s endorsement guides set the bar. More in the rules section of our AI marketing guide.

How to tell whether AI video pays off

AI video cuts production cost, not the demands on the idea and the opening. If viewers are gone after two seconds, it’s almost never the model’s fault.

Checklist before you publish

An AI clip renders fast, but a mistake goes public just as fast. This check takes a few minutes and saves you deletions, rejections and awkward comments.

My rule of thumb: a person watches every video once, start to finish, with sound, before it’s scheduled. Automated checks help but don’t replace that. If you adopt one thing, adopt this.

When not to use AI for a video

• Real experiences: testimonials, before-and-after, product tests. Here, being real is the point.
• Sensitive topics: health, finance, politics. A realistic AI person quickly reads as deception.
• Your own face as the brand: if people follow you, they want to see you. Use AI for B-roll, not for yourself.

12 prompts for AI video

The first prompts go to a chat assistant (script, storyboard); the scene prompts go straight to the video model. Replace everything in square brackets.

Write a 30-second script for a vertical video about [topic] for [audience]. Hook at 0–2 s, punchline or key point by second 5, then three short scenes, ending with a question. Voiceover the way people talk. No invented numbers.

Break this script into 4–6 shots of 3–8 seconds: [script]. For each: what’s in frame, the action, camera, light, voiceover on top, transition to the next.

Write an image prompt for this shot’s starting frame: [shot]. Person [age, look, clothing], location, light, 35mm look, vertical 9:16. Negative list: no text, no logo, no resemblance to well-known people.

[Subject] [one action]. Camera: [move, height]. Light: [mood]. Style: documentary, natural skin tones. Duration: 5 seconds. No extra people, no text in frame, natural hands.

For this video about [topic]: list 10 B-roll shots (3–5 seconds) that need no people and fit the voiceover. Give a short scene prompt for each in [style].

Describe the character [name/role] as a fixed text block for every prompt: age, face, hair, clothing, typical posture, voice. Max 50 words, specific enough that every starting image looks the same.

Turn this voiceover into captions for a vertical video: max 32 characters per line, max 2 lines per card, breaks at natural pauses, the punchline word on its own card: [text].

I’m planning [number] videos of [length] seconds with [number] shots. The model costs [price per second/credits]. Assume [number] takes per shot. What’s the cost per video and per month? Show your working.

Describe the right music for this video [description]: tempo, mood, instruments, length, where it needs to drop. Note: I only use commercially licensed music.

I’m planning this video: [description incl. AI use, realistic people yes/no, ad yes/no]. What labelling do I need under the EU AI Act, on Instagram/Facebook and on TikTok, organic and as an ad? A first pass, not legal advice.

This take has these problems: [e.g. distorted hands, face changes, camera jumps]. Rewrite the prompt to make these less likely without changing the scene: [prompt].

Adapt this video concept [description] for Instagram Reels, TikTok and YouTube Shorts: hook variant, length, caption, labelling and call to action for each.

Tools around AI video

“Try” links are affiliate links (advertising): I earn a commission at no extra cost to you.

More tools with pricing: best AI marketing tools.

The most common mistakes

1. One long scene instead of several short ones. Long clips lose faces, hands and logic.
2. Costing per generation, not per usable clip. Failed takes belong in the budget.
3. Not checking commercial use. Some free plans don’t allow advertising.
4. Recreating well-known faces or voices. Legally risky and banned by platforms.
5. Posting without a label. Platforms reject or restrict ads; in the EU the AI Act applies.
6. Optimising the model instead of the idea. A better model won’t fix a weak opening.