How Revisions and Approvals Work on an AI Video Project

Where the approval gates sit in a generative workflow, why they differ from conventional production, and what a revision can and cannot change.

Approval on a generative project follows a different rhythm from conventional production, and clients who apply their usual expectations encounter friction that neither party anticipated. The gates sit in different places, some changes are cheaper than in filming and others are considerably more expensive, and knowing which is which makes the process predictable.

There are four gates. The script, which fixes the message and the runtime. The visual language, approved as still images, which fixes the world. The storyboard, which fixes every shot with its duration and production method. And the offline edit, which fixes the structure before finishing begins. Each is a written approval, and each one closes a category of change.

The visual language gate is the one without a conventional equivalent and the one that matters most. In filming, the look emerges from a location, a camera and a lighting setup that everyone can see. In generative work it exists only as a specification and a set of approved frames, and if it is not approved explicitly it will be argued about at the edit. Approving stills is the cheapest decision point in the entire project.

What a revision can change easily is different from conventional production in useful ways. Reordering shots, adjusting timing, changing on screen text, altering the grade, swapping the music and adjusting the mix are all straightforward. Regenerating a single shot with a different action or angle is feasible. Extending a sequence with additional shots in the same world is comparatively cheap, because the world already exists.

What a revision cannot change cheaply is the visual language. A client who approved a cool, restrained, industrial look and then asks for something warm and organic at the edit stage is not requesting a revision but a new film, because every shot was generated within the approved world. In filming, a grade change might partially achieve this; in generative work the world itself is baked into each shot.

The other expensive change is adding environments. Because cost scales with the number of distinct worlds rather than with runtime, a request to add a scene in a new setting is a substantial addition even if it occupies four seconds. Clients accustomed to thinking in runtime consistently misjudge this, and explaining the cost structure at the brief prevents the surprise.

Iteration inside a stage is not a revision and should not be counted as one. A studio may generate a shot many times before it serves the storyboard, and that is production rather than revision. Clients occasionally see this activity and assume the film is being reworked; it is the normal cost of hitting a target rather than approximating it. Nour (2026), comparing prompt engineering with model selection, found these to be distinct levers with different effects, and part of that iteration is determining which lever a given shot needs.

The number of rounds should be defined the same way as in any project: one round is one consolidated set of feedback from one named approver, delivered at one time. Mirzaei et al. (2025) argue that project methodologies should be customised to the specific project rather than applied uniformly, and generative work benefits from tighter gating precisely because so much of the film's character is settled before anything moves.

There is a client education point worth making at the outset. Because the stills look finished, clients sometimes treat the visual language stage as a deliverable and comment on details that will not survive into the film. Framing that stage explicitly as intent rather than output, and stating what is being asked, whether the world is right rather than whether the reflection is correct, keeps the review productive.

The practical instruction for a client is to spend attention early and hold the line late. Review the stills carefully and say what is wrong while it is free. Approve the storyboard knowing that it defines exactly what will be produced. Then treat the edit as a refinement stage rather than a reopening, because in this workflow the decisions that shape the film were all made before generation began.

References

Nour, R. R. (2026). Prompt engineering versus model selection for cognitive accessibility in large language models: An empirical study. IEEE Access, 14, 44740–44754. https://doi.org/10.1109/ACCESS.2026.3667133

Mirzaei, M., Mabin, V. J., & Zwikael, O. (2025). Customising hybrid project management methodologies. Production Planning & Control, 36(9), 1188–1205. https://doi.org/10.1080/09537287.2024.2349231