What AI Actually Does in Film Production — And What It Doesn't

Updated August 2026. Two department-level follow-ups have since been published: what AI has actually taken from the VFX pipeline, where the exposure is highest and most geographically concentrated, and where generative video actually fits in a real pipeline, written after OpenAI announced the Sora API sunset.

The useful framing

The argument about AI in film is usually conducted between people claiming it will replace everyone and people claiming it does nothing. Neither is a useful planning assumption.

The accurate version, as of 2026, is narrower and more boring: AI is good at the mechanical, high-volume, low-judgement work that sits between the creative decisions. That work absorbs a genuinely large share of a production budget, so compressing it matters. It is also not the work most people got into filmmaking to do.

Where it genuinely works now

Script breakdown first passes. Reading a screenplay and tagging every proper noun, recurring prop, location change and named character is mechanical. Automated first passes are reliable at that and unreliable at judgement calls — which is why the useful implementation presents suggestions for an AD to correct rather than applying them. On a 90-page feature this is often the difference between three days of tagging and one. It is why our own breakdown tool does the first pass and stops there.

Schedule drafts. Given a completed breakdown, sorting scenes by location, cast overlap and time of day is an optimisation problem. Generating a draft stripboard for a 1st AD to refine beats starting from an empty board. It does not replace the judgement about which trade-offs matter.

VFX cleanup. Rotoscoping, roto-assist and colour matching have absorbed real AI gains, with reported reductions in manual labour on large-scale projects in the 30 to 50 percent range. This is the most mature application in the pipeline.

Localisation. AI-assisted dubbing and subtitling has compressed global rollout packages from roughly twelve weeks to around ten days on some 2026 releases. For anything with international distribution this is a substantial schedule change.

Storyboards and concept visuals. Fast, cheap previsualisation for pitching and blocking. Good enough to communicate intent; not a substitute for a storyboard artist on complex action.

Where it does not work

Judgement about what a scene needs. An automated breakdown will tag the gun. It will not notice that the scene implies the character has been holding it for the whole conversation, which is a wardrobe, blocking and continuity problem.

Performance. Nothing close.

On-set decisions. The value of an experienced 1st AD is the call made at 4pm when the day is behind and something has to be cut. That is a judgement about people, money and story simultaneously.

Anything where being wrong is expensive and invisible. Turnaround compliance, continuity, safety. These are areas where a plausible-looking wrong answer is worse than no answer.

The backlash is also real

Worth planning for: audience and industry resistance to AI-generated work has hardened rather than faded. A visible counter-trend has emerged — productions shooting on film, handheld and minimally processed, specifically to signal human authorship. We wrote about that in radical authenticity.

The practical implication is that AI in the pipeline and AI in the image are being judged very differently. Using automated breakdown to save your AD three days carries no reputational cost. Generating your footage does, at least for now.

What this means for crew

The honest read is that AI is compressing roles rather than deleting them, and the compression is uneven.

Most exposed: entry-level work whose value was volume rather than judgement — basic roto, first-pass logging, routine localisation, template-driven graphics.

Least exposed: anything requiring physical presence, real-time judgement, or accountability. On-set roles broadly, plus anything where somebody must be answerable for a decision.

Changing rather than shrinking: post roles are moving up the stack. Less time on mechanical cleanup, more on the decisions that were previously squeezed out by it.

The uncomfortable part is that entry-level mechanical work is also how people traditionally learned. If the industry compresses that away without replacing the training path, it creates a pipeline problem that nobody currently owns.

A reasonable position

Use AI where being wrong is cheap and checkable. Do not use it where being wrong is expensive and invisible. Keep a human accountable for every output that reaches the floor.

That is not a compromise position — it is just what the technology is currently good at.