Where AI actually helps in video production, and where it still does not

Two claims are made about AI video, and both are wrong.

The first is that generated video will replace production. The second is that it is a toy and can be ignored. In practice it is now a real part of how a film gets made — it just is not the part most people assume, and the useful applications are far less exciting than the demos.

This is a stage-by-stage account of where it genuinely earns its place in our work, where it does not, and the one question that has nothing to do with quality at all.

Where it genuinely helps

Storyboards and pre-visualisation

The clearest win on the list, and the one nobody makes videos about.

A client who cannot read a script can look at eight frames. Being able to generate those frames in an afternoon — and then regenerate them after the feedback — moves the argument about what the film is to before a crew is booked, which is the cheapest possible place to have it.

Nobody sees these images. They are a thinking tool, and they are excellent at it.

Transcription and logging

Quietly the biggest time saving in the whole process.

Six hours of interview rushes, transcribed and timecoded, turns a day of scrubbing into a text search. An editor looking for the sentence where the MD explains the expansion finds it in seconds instead of watching until they hit it.

This is not glamorous and it has changed our edit turnaround more than anything else.

Cleanup that used to need a specialist

Noise removal from audio recorded in a room that was not quiet. Upscaling old footage. Reconstructing speech under an air conditioner. Subtitles in English and Odia, generated then corrected by a person.

All of it is reliable now, all of it is cheap, and all of it used to be either a specialist line item or something a client was told could not be fixed.

First-pass assembly

Cutting silences, removing false starts, roughing a long interview down to something an editor can start work on. It produces a starting point, not a cut — but a starting point on a two-hour interview is a genuine hour saved.

Where it does not help

The shoot itself

This is the one that matters, and it is not close.

No model has footage of your plant, your convocation, your chief guest, on the day it happened, with the people who were actually there. That footage exists because somebody stood in the room and recorded it. A film about a real company is mostly made of things that really happened, and generation has nothing to offer that.

Anything that has to hold for more than a few seconds

Coherence in generated video degrades past roughly ten seconds in most current models. It drifts — a face changes slightly, a hand does something wrong, a logo becomes an approximation of itself.

For a three-second texture insert, fine. For a shot a viewer looks at, the drift is exactly what the eye catches, and the audience does not know why it looks wrong, only that it does.

Consistency across a campaign

A single striking clip is easy. Fifteen clips that look like they came from the same film is the hard problem, and it is still hard — style references and locked prompts help, and they do not fully solve it.

This is precisely backwards from what a brand needs. Brands need the boring consistency, not the striking one-off.

Judgement in the edit

Which take carries the meaning. When to hold on a face two seconds longer than is comfortable. Which of two good sentences to cut so the other one lands.

A model will assemble something competent. The decisions that make a film move are still made by a person who understood what the film was for, and that has not shifted.

The question that is not about quality

Before any of the above matters, there is a consent problem, and it arrives earlier than people expect.

Do not synthesise a real person. Not your MD’s voice for a line they did not say, not a client’s face reading a testimonial they did not give, not an employee generated into a scene they were not in. Even where it is technically convincing and legally arguable, it is a representation of a real person doing something they did not do, published under your brand.

The same applies to generated footage passed off as a record of a real event. A film about your factory that contains a shot of a factory that does not exist is not a stylistic choice; it is a false claim about your own operation, sitting on your own website.

Our line is simple: generated material can be used for things that are obviously illustrative — abstract texture, a concept frame, a graphic element. It is never used to depict a real person, a real place or a real event that we did not film.

What this means for a budget

Realistically, in 2026:

• AI has reduced our post-production time, particularly on long-form interview and podcast work, and that shows up in delivery timelines more than in the quote.
• It has not reduced shoot costs, because the shoot is the part it cannot do.
• It has added a pre-production step — visual references before the shoot — that clients find genuinely useful and that costs almost nothing.

Anyone offering a corporate film substantially cheaper on the basis that AI makes it is either producing something you would not want on your website, or is doing the same work and describing it differently.

The short version

Use it for the thinking, the logging, the cleanup and the first pass. Shoot the film. Keep a person on the decisions. Never let it depict someone real.

That is roughly where the line sits today, and it moves — the sensible response is to check it every few months rather than to settle the argument once in either direction.