How to make marketing videos with AI, no camera crew

Making a marketing video with AI is a chain, not a single tool: write the script, film or generate the footage, cut it into clips, caption it, and post it. Done with off-the-shelf tools it’s an afternoon of work, and once it’s a habit you walk away from one sitting with a week of video.

One decision shapes the whole chain, so make it first. Film a real face when you can, and generate only the parts you can’t. A phone in decent light beats a synthetic presenter on the one thing marketing video is for, which is trust, and the numbers behind that are in our AI video tools comparison. That guide is the page for picking a tool and what each costs. This one is the page for actually making the thing, step by step, including the part every other how-to skips: where AI video still looks fake, and how to stop it.

1. Pick the format before you open a tool

Decide which of three videos you’re making first, because the tool and every step after it follow from that choice. There are only three shapes worth your time.

A filmed talking head is you, talking to a phone. It’s the cheapest to make and the strongest on trust, and it’s the one this guide leans towards. An avatar talking head is a synthetic presenter reading your script, for when you won’t go on camera and the content is informational. Generated footage is shots you can’t film, a few seconds of atmosphere or product mood conjured from a prompt. Match the shape to the job, then pick the tool: which avatar or generator to use, and what each costs, is sorted in the video tools comparison, so this guide names tools but won’t re-rank them.

The common mistake, made on day one. Reaching for a generator to fake a video you could have filmed in ten minutes. If you’ll appear on camera at all, that’s almost always the better video and the faster one. Generate only what’s genuinely unfilmable.

Script ChatGPT or Claude

on camera

You Film yourself phone, real face

won't film

Generate avatar or b-roll

Clip OpusClip / Captions

Caption + brand burn in, then check

Post a scheduler

2. Write the script with AI, and write it for the ear

The script is a writing job, so do it in the chat model you may already pay for, not in a video tool’s script button. ChatGPT or Claude will draft a hook, a thirty-second outline or a full script from a one-line brief, and if you’ve taught it your voice the draft comes back sounding like you. Setting that voice up once is worth doing properly, and it’s covered in the AI copywriting tools comparison, so it won’t be repeated here.

What’s different about a script is that it’s heard, not read, so write it for the ear. Three things carry most of the result. Lead with a hook in the first two seconds, because that’s where people bail: OpusClip’s analysis of thousands of Shorts found 50 to 60% of the viewers who drop off are gone inside the first three seconds, and that retention holds best when the whole thing runs 15 to 30 seconds. Keep sentences short and say one thing at a time. And read the draft out loud before you film it, because a line that scans on the page can tie your tongue in knots on camera.

Don’t ask for a “viral script”. You’ll get generic, sing-song filler that sounds like every other AI video. Give the model your actual point, who it’s for, and two examples of videos you like, and ask for your hook three different ways. The specifics are what make it yours.

3. Film yourself: the cheap path that wins on trust

If you’ll go on camera at all, filming yourself is the highest-return footage you can get, and the kit is a phone you already own. The setup is quick. Put a window in front of you for soft, free light, set the phone at eye level rather than looking up your nose, find a quiet room, and talk to the lens like it’s one person, not an audience.

The move that turns this into a system is to record long and batch it. Film three or four scripts back to back in one sitting while the light and your energy are good, because the next step chops each recording into several clips. One filming session becomes a fortnight of posts, which is where the afternoon-to-a-week maths comes from. Don’t over-produce it either. A clear, well-lit phone video with a real person beats a glossy, over-edited one for the kind of trust that makes someone buy.

4. Or generate the footage: avatars and b-roll

If you genuinely won’t film, an avatar reads your script, and if you need a shot you can’t capture, a generator makes it. Both have a place, and both have a ceiling.

For an avatar, HeyGen and Synthesia work the same way: pick an avatar, paste your script, choose a voice, and it renders a presenter delivering your words in minutes. One craft tip from HeyGen’s own studio guidance makes a real difference: keep sentences short, because long ones make the avatar’s delivery drift and stiffen. If you want to see the whole flow before you try it, there’s a clear ten-minute walkthrough in HeyGen AI Tutorial: How to Use AI Avatars by Kevin Stratvert. Note that building a custom avatar of yourself needs a recorded consent step, which the comparison guide explains.

For footage you can’t shoot, Runway and Google Veo generate short clips from a prompt, Veo with sound built in. Two limits decide how you use them: the clips are short, about eight seconds on Veo and up to sixteen on Runway, and the good models are metered by credits, so your money buys a handful of shots rather than a library. The detail on that credit cost is in the video tools comparison. Use generated footage as a few seconds of mood between your real shots, never as the whole video.

Where generation quietly falls apart. Today’s models still struggle with anything that has hands, on-screen text or continuity across shots. Ask for a person waving or a sign with words on it and you’ll get melting fingers and gibberish letters. Keep generated clips to simple motion, landscapes, textures, objects, a slow push past a product, and you sidestep the worst of it.

5. Cut the recording into clips

Whatever you filmed or generated, the marketing value is in the short clips, and AI does the tedious scrubbing for you. A clipping tool like OpusClip or Captions takes your long recording, finds the moments most likely to land, reframes them to vertical, and scores each one for likely reach, so you publish the strong ones first. A 30-minute recording comes back as a stack of captioned clips in under ten minutes.

If you want to watch the whole thing done once before you try, a step-by-step OpusClip walkthrough is embedded below: importing a video, choosing clip length, and styling the result.

The thing to resist is publishing all of them. Lead with the two or three highest-scoring clips, hold the rest, and you’ve got a posting queue instead of a one-off.

6. Caption and brand it

Burn the captions in, because most social video is watched in silence. Around three-quarters of people keep their phone muted while watching, and adding captions lifts comprehension by 56% (Sharethrough research). A clip with no captions is a clip most of your audience can’t follow. OpusClip and Captions both add captions automatically, so the work is in the checking: read them back and fix the misheard names and homophones, because an auto-caption that turns your product name into nonsense undoes the point of having them.

Then make it look like you, not like the tool. Set your fonts, your colours, a small logo in the corner, and one consistent caption style, so every clip is recognisably yours at a glance. The clippers hold a brand kit for this, and Canva is the fallback for any thumbnail or end-card you want to add. If you’re pulling in stock images or generated visuals to dress a clip, the licensing rules on AI-made assets are covered in the image and design tools comparison.

7. Post it on a cadence

The making is done, and posting is a scheduler’s job, not a video tool’s, so don’t go looking for one tool to do both. Drop the clips into a scheduler, queue them across the networks you care about, and let it post on a steady rhythm. Which scheduler, and how to drive one from a chat, is the whole subject of the AI social media tools comparison. Record long, clip many, queue them, and the channel keeps running while you get on with the rest of the business.

Where AI video still looks fake, and how to fix it

AI video gives itself away in a handful of predictable ways, and each one has a fix. This matters because viewers can tell, and around a third say a video they clock as AI lowers their opinion of the brand, with the figures in the video tools comparison. You don’t beat that by avoiding AI. You beat it by knowing the tells.

The avatar tells are a too-smooth gesture on a loop, a flat emotional tone, and eyes that don’t quite land. Keep avatars to short informational clips, break a long script into a few shorter renders so the delivery resets, and don’t ask a synthetic presenter to carry emotion it can’t fake. The voice tell is a text-to-speech cadence that stresses the wrong word. Use your own voice wherever you can, and if you clone one, feed it clean audio and listen back for the odd emphasis. The generated-footage tells are the morphing hands, wrong physics and melting text from step four, fixed by keeping generation to a few seconds of simple b-roll with no hands, words or continuity. The caption tell is mistiming and misheard words, fixed by reading them back. And the editing tell is jump-cut whiplash and a restless, over-zoomed sheen, fixed by using fewer, slower cuts and cutting on a natural pause rather than mid-breath. One more move that helps: if a video is heavily AI-made, a quiet line saying so tends to earn more goodwill than getting caught.

Doing more from one chat: the AI-native rung

Once the chain is familiar, one chat model can handle most of pre-production from a single brief. Instead of writing a script and stopping, ask Claude or ChatGPT for the script, five alternate hooks, a shotlist of what to film, and a rough storyboard, all in the brand voice you saved earlier (the copywriting comparison covers that setup). That turns a blank afternoon into a clear one.

The newer “describe a video and get a finished cut” tools, like InVideo or Veo inside the Gemini app, are real and improving, but be honest about where they are: dependable for short, simple b-roll, still hit-and-miss for anything with a coherent story, a consistent character or on-screen words. Treat a generated cut as a fast first draft you take apart, not a finished video.

Video at volume: the experimental edge

There’s a level past clicking generate, worth knowing exists even though it isn’t a job for a non-technical operator. Every serious tool here has an API, so you can wire one into a flow that mass-produces video from a brief: the same ad in ten variants, one clip localised into five languages, a weekly cut assembled and dropped where you need it. It’s experimental, it burns credits, and it breaks when an API or a price changes, so it needs someone who maintains it. The payoff is real only when you’re producing video in genuine volume, and at that point the sensible answer is to have it built once and handed over rather than to learn it yourself.

Film, generate, or get help

For almost every small business, the answer is to film what you can and let AI do the editing. That’s the cheapest path, the fastest once it’s a habit, and the one that builds the trust a marketing video exists to build. Filming yourself, scripting in a chat model, and clipping with OpusClip or Captions covers it, for the price of one or two cheap subscriptions.

Generate as the second move, not the first. An avatar from HeyGen or Synthesia is a fair stand-in for informational video you won’t film, and a footage generator like Runway or Veo is for the occasional shot you can’t capture. Reach for help only at the volume that justifies an automated pipeline, which is a build, not a setup. The one mistake to avoid is paying for a synthetic face to front your brand when the video that actually wins is the real one you could shoot on a phone before lunch.