AI tools can make a video clip in minutes. Getting a finished video a client will sign off on is a different job. This is how a client’s product video came together, from storyboard to final cut, and what we changed along the way.
An AI video model can turn a still image into a moving clip in about two minutes. That does not mean you have a video.
Recently I led the production of a 60-second explainer for a fintech client launching an app in Canada. We used AI for almost all of it: the images, the motion and the voice. It took more than 20 rounds of revisions to reach the final cut. Most of what slowed us down had nothing to do with how fast the tools were.
The tools, and how I connected them
Instead of clicking through a web app for every shot, I connected the tools and ran most of the work through Claude Code.
- Higgsfield made the images and the video. Its Soul model made the stills of our main character, and Seedance turned them into short clips.
- ElevenLabs voices, used through Higgsfield, did the voiceover.
- ffmpeg cut the shots together, timed the voice and mixed the audio.
- Python scripts checked the output for frozen frames, camera shake and colours that should not be there.
With this setup I could rebuild the whole video in minutes. That helped. It also meant I could make the wrong thing very quickly.
Plan the story before you generate anything
My first cuts came from the script. I wrote the lines, then made shots to fill them. The client pushed back because the video did not follow the storyboard we had already agreed on. We started again from the client’s own list of scenes, and each scene got an approved image before any animation.
Images from the client caused a different problem. When the client sent an image for the opening, we treated it as direction and made our own version. They meant for us to use it as it was. That one misunderstanding cost several rounds.
Late in the project, the client also pointed out that the opening showed our character leaving Canada while the voiceover said she had just landed. The picture and the script were each fine on their own. Together, they told the wrong story.
What we do now: The storyboard is signed off before any animation. Every image a client sends is marked as a guide or as exact. The script and the storyboard are reviewed together, not separately.
Keep the character and the world consistent
AI models do not remember what they made last time. In one shot our character wore trousers. In the next she wore a skirt and pulled a different suitcase. The fix was to make every shot of her from one approved image, and to decide which hand held what: suitcase in the right, phone in the left, bag on the arm. After that, props stopped appearing and disappearing.
The models also added things we did not ask for. I asked for an airport with plain aircraft and got British Airways and Air Canada tails, image after image. The client could not show airline logos for legal reasons, so we removed them after generation. Text was worse. The models can’t spell, so shop signs came out as words that do not exist.
What we do now: One approved image per character, with fixed props. Every generated clip gets checked for logos. Any text that must be readable gets added after generation, not left to the model.
Motion is where AI video shows its limits
Our early cuts had clips that ran too long and motion that felt unnatural. In one shot her suitcase rolled behind her with no hand on it. We also had 71 seconds of script but only about 48 seconds of distinct footage, so the edit reused shots to fill the gap. The client noticed right away. We cut the script to about 46 seconds and made a rule that each shot is used once. Our edit script stops if a shot appears twice.
Generated clips often start with the person standing still. If you cut to the first frame, they look like they wake up on screen. We measured when the movement started in each clip and cut in after that point.
The hardest shot was a Canadian flag in the opening. Indoors, it hung still and looked like a photo. Outside on a pole, the video model redrew the pole so it went through a parked plane. That taught us something important: the model redraws every frame, so anything that must stay put, like a pole, a logo or a line of text, can drift.
What we do now: Each shot is used once, every clip starts on movement, and anything that must stay fixed in the frame gets extra checks.
Show the product with real screens
The client needed the app on screen. We first placed the app screens onto the phone in her hand. They looked slanted because the phones were tilted in the footage, and her thumb went under the screen instead of on it. Clean, full-screen shots of the app, cut between the live shots, looked better and took less time.
What we do now: App screens come from the real designs and appear as full-screen shots, not placed onto a phone in the scene.
Voice and music
We first tried to clone a real person’s voice from three recordings. The result sounded like a phone call. The recordings were made on a phone, and a clone can’t sound better than its source. We moved to a stock ElevenLabs voice. We also removed the footstep and suitcase sounds the model added, at the client’s request.
The product name took three rounds. The voice said it wrong, then said it as two words. We tested several spellings and measured the gap in the middle of the word to pick the one that sounded like a single word.
Music was the one thing the platform could not do. Higgsfield has music and sound models, but only for its game tools. For a marketing video you need a licensed track from a stock library.
What we do now: The voice gets picked early, the pronunciation of the product name is agreed before recording, and music is budgeted from the start.
Review in motion, and know when to stop fixing
At first I reviewed the output at thumbnail size and missed things. We moved to full-size reviews against the approved images.
We also learned to protect the approved cut. After the client approved it, a later round drifted, and one change I thought I had made never went in. We went back to the approved cut, added only the new requests, and checked that each change actually landed.
Then I made the worst mistake of the project. The final opening clip had a slow camera move, so I added a step to hold the camera still. I checked it frame by frame and it looked fine. In motion, it shook like an earthquake. The background changed direction 60 times in four seconds. The untouched clip changed direction zero times. We shipped the untouched clip.
What we do now: Every cut gets watched as it plays, at full size. We measure camera shake on every shot. Every change builds on the last approved cut. And we stop fixing when the fix is not needed, because sometimes the best version is the one you did not touch.
What I will agree on before the next project
Most of our revisions came from decisions made too late. Next time I want these agreed in writing before production starts:
- The storyboard and scene order. No animation until it is signed off.
- Guide or exact. For each client image, is it a guide, or must we use it as it is?
- Formats. A vertical version for social is not a quick crop. When we checked, 12 of our 18 shots would lose about 60% of the frame.
- What a revision round is. One list of feedback, from one person, for each stage.
- Text and brands. Which signs, logos and screens must be readable, and who clears them.
- Voice and music. The voice, how the product name is said, and the music licence.
None of these are AI problems. A normal video shoot needs the same decisions. AI just makes them easy to skip, because making another version feels free.
Conclusion
AI cut the time it takes to make a shot from days to minutes. It did not cut the time it takes to decide what the shot should be, or to check it properly. That is still where the real work is.
If you are planning a product video and want to know where AI fits and where it doesn’t, contact 247 Labs and we will help you work it out.
