How AI Commercials Are Made: A Director's Workflow (2026)
TL;DR: A good AI commercial is not one prompt. It is a production pipeline — references, locked characters and products, image generation, video generation, sound, and a real edit. This is the workflow Mohamed Abdelwahed, an Egypt-based AI video director, uses on commercials for brands like BYD, LikeCard, Haseem and Domty.
What is an AI commercial, really?
An AI commercial is an ad where most or all of the footage is generated by AI models instead of being shot on set. The craft is the same as traditional advertising — an idea, a story, a product truth, a strong edit. What changes is how the pictures are made: instead of cameras and crews, a director builds shots with image and video models, then finishes them in the same tools any editor uses.
The difference between an "AI demo" and a commercial a brand can air is discipline. Here is the pipeline, step by step.
Step 1: Why start with references, not prompts?
Every project starts with a reference phase, before any concept or prompt. Three kinds of reference:
• Content references — similar ads and films: how others told this kind of story.
• Visual style references — colour, light, texture, framing.
• Motion references — how the camera moves, how fast the cuts are, how people act.
References also come for characters and locations — films, Pinterest, real photos. These references are fed into the image models as real inputs, not just described in words. A text-only prompt gives you the model's average; a reference gives you a direction.
Step 2: How do you lock the product, the characters and the locations?
Before a single video take, lock the three things that must never change:
• The product. For a car, the official studio renders are the law: grille, lamps, wheels, paint. On the BYD Ti7 Iraq film, every frame was matched to BYD's 45° renders and ten interior renders. For packaged food, the real pack photo — not a description — defines the shape and size, and the pack must sit at a believable size in a hand.
• The characters. A character sheet (front, side, back, close-ups, outfits) for every person or mascot. LikeCard's Monster has a strict design — four fingers, short legs, card-shard fur — and every take is checked against it.
• The locations. Build each location as an empty plate first, then place the product into it, so light and perspective come from the place.
Step 3: How are the key frames generated?
Stills come first, because a still is cheap to iterate and a video model follows a good first frame. Typical tools in 2026: Nano Banana (Google Flow), GPT Image and Seedream, run in Google Flow or Higgsfield.
Lessons from real projects:
• Use prompt chains, not single prompts. Each pass changes one thing: pull the camera back, swap the sky, move the sun, remove a hand. The Ti7 storage shot went roof closed → boot loaded → four sky swaps.
• Generate at the final resolution. Generating natively at 4K is often cheaper and cleaner than generating at 2K and upscaling.
• Every edit pass costs quality. Re-editing a finished image again and again softens it. Go back to the clean source instead of stacking edits.
• Compare side by side. After every generation, put it next to the reference — the product and the characters — and only move on when they match.
Step 4: Which AI video models do you use, and for what?
In 2026 the main video models are Seedance, Kling, Higgsfield (which hosts several models) and Google Flow (Veo). Each has strengths; a real project mixes them. The BYD Egypt World Cup film used 17 Seedance, 2 Kling and 28 Higgsfield takes; Haseem's National Day film mixed Higgsfield, Seedance and Flow.
What matters more than the model:
• Generate many takes, choose few. A 30-second LikeCard spot used 64 video takes; a 69-second BYD film used 63 Higgsfield runs and 84 video files. Most takes never make the cut.
• Judge a take on its first and last frame. If the start and end don't match the plan, the middle won't save it.
• Direct the camera like a camera operator. For realism, ask for handheld shake, quick push-ins and pull-outs, natural whip pans — like someone filming on a phone.
• Direct the acting. Faces need intention: embarrassment, puzzlement, relief. Write the emotion, not just the action.
• Decide cuts vs. one continuous shot on purpose. Some ideas need one long move; most ads need separate shots with real cuts.
Step 5: How do you handle sound and voice-over?
Sound is half the film. Every world gets its own layer: ambience, foley, music bed, sting. On some films the right choice is no music at all — only natural, real sound.
For Arabic voice-over, dialect matters: an Egyptian ad needs an Egyptian read, a Saudi ad a Saudi one. AI voice tools (ElevenLabs, Google's Gemini TTS) can do it, but the director must push for the right accent, speed and energy, and check every line.
Step 6: How is an AI commercial finished?
The finish is classic post-production:
• Edit in Premiere Pro or After Effects, with real pacing.
• Composite — keys, masks, type and logos — in After Effects.
• Grade — colour comes last, often from a defined cinematic palette, once the cut is locked.
• Deliver every format the brand needs: 16:9, 9:16 verticals, squares, digital billboards.
What are the biggest mistakes in AI commercials?
1. Letting the model redesign the product. If the car, the pack or the mascot drifts, regenerate — never ship it.
2. Skipping references. Without them, everything looks like generic "AI".
3. Inventing text. Use only the copy the client approved; express extra ideas with visuals.
4. Over-editing images. Quality drops with every pass.
5. Ugly-perfect people. Characters should look real and fit the brand — neither plastic-perfect nor careless.
How long does an AI commercial take?
It depends on the idea, but real examples: BYD Egypt's 50-second World Cup film was made in 4 days (10–13 June 2026); Haseem's 70-second National Day film over 16 days (3–18 September); LikeCard's Ramadan spot had all 64 takes generated in a single day.
Who is Mohamed Abdelwahed?
Mohamed Abdelwahed is an Egypt-based AI video director and senior art director. He creates AI commercials and campaigns for brands across Egypt and Saudi Arabia, including BYD, LikeCard, SAB, Haseem, Domty and Wafeq, and teaches AI video generation. More than ten art directors came out of the teams he led. Portfolio: abdelwahed.art · Contact: one@abdelwahed.art
Mohamed Abdelwahed — AI Video Director & Senior Art Director · one@abdelwahed.art