Two years ago you couldn’t get through a single industry conference without someone predicting that generative AI would replace half of video production by 2026.

It’s 2026. Nobody we know has been replaced. But the workflow has changed, in ways some of our clients ask about and most of our peers either oversell or pretend isn’t happening.

This is the honest version: where AI is useful, where it isn’t yet, and the line we won’t cross on client work.

The shift, in one sentence

Pre-production and post got faster. Set didn’t change.

Everything generative AI is good at right now is upstream of the camera (planning, references, treatments) or downstream of the cut (transcripts, alternate takes, color matching, automated logging). On the actual shoot day, AI is invisible — and it should be.

Where we use AI today

Treatment writing

We write the first 60% of every treatment with the help of an LLM. Not “write me a treatment” — that produces vague, formulaic copy. Specifically: a director sketches the scene-by-scene structure on paper, dictates the tone of each beat, and uses an LLM to expand into prose at proper treatment length. The director rewrites every paragraph after.

Time saved: about 6 hours on a typical treatment.

If a director wants to show a client “moments where a camera lingers on a face longer than feels comfortable, in films from the last five years,” that used to be a half-day of curating. Now we describe the moment, an LLM produces a list of 30 candidates, we watch the relevant timecodes, and we pull the eight that fit.

The research isn’t replaced. It’s accelerated. The director still has to watch each clip and exercise taste.

Every interview we shoot gets fed to an AI transcription service. The editor can search “where did they talk about their dad” and get a timecode in three seconds. This used to be an editor scrubbing through six hours of interview tape.

This is one of the largest workflow improvements of the last three years. It’s also boring. Nobody’s talking about it because there’s no demo to dunk on.

Alternate-take generation in post

Specifically: if we shot a brand film and the talent stumbled on one word, we can use a voice-cloning model to generate a clean read of just that word, color-match it, and patch it. The talent always approves; we never use this without explicit consent. But it’s saved several reshoots.

This is more controversial. Some people think it crosses a line. Our position: a brand-side talent who said the word the right way three times in a row but stumbled on take 47 is not being misrepresented if we patch take 47 with their own voice. They are being misrepresented if we generate words they never said. We don’t do the latter, ever.

Color reference matching

The colorist drops a reference still into Resolve, and an AI-assisted match suggests a starting curve. The colorist throws out 40% of those suggestions because they’re wrong, but the 60% that land save real time. The grade still requires a human eye to land. The starting point is just better than starting from neutral.

Subtitle and caption generation

Auto-generate, then human-review for tone and brand voice. Saves an hour per deliverable.

Stock footage replacement (sometimes)

If the script needs a 2-second cutaway of “an empty office at night” and we don’t have it, we can generate it. We’d rather shoot it. But if the choice is “generate the cutaway” or “extend the shoot a half-day for one shot,” we’ll generate it and label it as such in the delivery.

Where we don’t use AI

On set

Generative AI doesn’t write scripts we shoot. It doesn’t choose lighting setups. It doesn’t direct the talent. It doesn’t operate the camera. It doesn’t set the focus pull. The film is what was in front of the lens at the moment the shutter opened.

This is the part of production AI hasn’t moved, and the part where we don’t expect it to. Set is a creative-and-physical exercise that doesn’t have a digital surface to optimize.

To replace cinematography

Generative video models in 2026 (Sora 2, Veo 3, Runway Gen-4) are impressive in the abstract and unusable for paid client work in the specific. They can’t hold a face consistent across a 30-second cut. They can’t render hands without artifacts in close-ups. They can’t match lighting from one generated shot to the next in a way a colorist can rescue. We test each major release. None has produced a frame we’d ship.

To fake a shoot

The line that matters: if the brief says “we shot this,” we shot this. We don’t generate B-roll of locations the brand never visited, employees who don’t exist, or product behavior the product can’t actually do. This is partly principle and partly self-interest — once a client catches a generated frame in the deliverable, the trust is gone, and so is the relationship.

Without disclosure

Every project we deliver includes a one-line note: “Where generative AI was used in production.” If we used it for transcription only, we say so. If we used it for an alternate-take patch with talent consent, we say so. If we generated a 2-second cutaway, we say so. Clients always have approval before any AI-touched frame ships.

Why the disclosure matters

Two reasons.

One, brand-side trust. A CMO who later finds out their hero film contained an AI-generated B-roll shot they weren’t told about doesn’t fire the agency for the AI. They fire the agency for not telling them. Disclosure preserves the relationship.

Two, talent rights. SAG-AFTRA contracts have specific language about voice cloning and synthetic performance, and that language is still evolving. We document every consent. It costs us nothing and protects everyone.

What’s not coming as fast as people say

Full text-to-film for production work. Not in 2026. Maybe by 2028; maybe not. The bottleneck isn’t model quality, it’s controllability. A client who says “make her look slightly more amused in this beat” can be directed by a human; can’t be prompted into a model that holds 47 frames consistent.

AI replacing junior crew positions. A 2nd AC isn’t being replaced by a model. The job is half spatial reasoning, half social — neither is current AI’s strength.

AI replacing editorial. Rough cuts, sure. The director’s cut, no. Knowing where to land an edit is taste, and taste isn’t a feature you can ship.

What this means for clients

If you’re a brand-side team weighing AI-heavy production options against traditional production, three things to know:

1. Cost differences are smaller than they look. A “fully AI-produced” 30-second spot still requires a creative director, a script, sound design, music, color, and approval cycles. The savings come out of camera and crew, which is 30–50% of a typical budget. The other 50–70% is the same.

2. Quality differences are larger than they look. The generated stuff is technically impressive in isolation. In comparison to actual cinematography, on a 60-inch retail screen, in 2026, it’s still uncanny. Audiences clock it.

3. The relationship matters. If your production company doesn’t tell you which parts of your film were AI-touched, find a new one. It’s not a difficult disclosure. The companies that won’t make it have something to hide.

How we’d describe our position

AI is in our pipeline. It’s not in our pictures.

Bookkeeping, search, transcription, post-process touch-up — the camera-adjacent work where the model’s output saves a colorist or editor hours and the audience never sees it. That’s where it earns its place.

The picture itself is still made by a person with a script and a camera. We don’t see that changing soon, and frankly we wouldn’t change it if we could.

If you’re scoping a project and want to know exactly which AI-touched steps would be in your deliverable, send us the brief. We’ll show you the workflow line by line.