What an AI video pipeline actually removes from the edit

The rough cut went first, and it went completely

On the line I ran, a topic entering the system is scored against a set of format templates and routed to one of them. The template carries the shot order, the pacing of the copy and the packaging structure. Footage is gathered by format rather than by project, and gaps — transitions, atmosphere, establishing shots — are filled by a generation step. Nobody lays down a first timeline, because the first timeline is what the template is.

This is the same shape the verified record shows at a much larger scale: a streaming platform told shareholders that generative workflows were used across roughly 300 titles in a year, and that the work concentrated in post-production. Post is where the specifiable hours live.

Worth saying plainly what this costs. Those hours were how junior people used to get paid while they learned the judgement needed for everything else in this list. On my line that entry ramp does not exist, and I do not have a replacement for it.

Captions, formats and packaging became modules, not work

Subtitle cards, map segments, transparent overlays, title and end cards, cover-frame extraction, loudness normalisation and export are each a component any format can pull in. Producing a second aspect ratio or a second language is a parameter, not a pass.

This is the least interesting of the five tasks and the one that disappeared most completely, and those two facts are the same fact. A task that can be fully specified is a task that can be handed over. The ones below resisted precisely because I could not write them down.

What did not go away is knowing what good looks like — and it got more expensive

We ended up with ten hard rules for the line. Every one of them was earned by having work sent back: the footage must serve the subject, the numbers must be honest and consistent in their definition, no shot may repeat across adjacent episodes in a series, the cover is the first frame. None of them was foreseen. Each was written after something shipped that should not have.

The thing that actually changed the output was not writing them down. It was translating each into an assertion in the QA script. For the weeks they lived in a handbook, the same mistakes kept happening. The rule I now hold is that a rule which cannot be machine-verified has not entered the process yet — it is still prose.

This is why the pipeline makes judgement more valuable rather than less. It executes whatever standard you encode, at volume. A team that cannot tell good from bad now produces the bad version several times faster than before, and with better packaging on it.

The new job is directing generated footage — and deciding what it may not touch

The single most useful boundary we drew: generation is responsible for texture, the program is responsible for fact. Models generate scenes, materials and controlled motion. Titles, dates, model names, routes, prices, arrows and captions are composited by code onto a locked foreground layer. Readable text is never handed to a video model.

That boundary is not an aesthetic preference. It is the difference between an error that looks wrong and an error that reads as authoritative and is wrong. The second kind is the expensive one, and a generative pipeline produces it willingly unless something structurally prevents it.

The verified record has already put a name on this work from the other side. In the 2024 IATSE Basic Agreement, work done by prompting an AI system or supervising its operation remains covered work. Whatever else is unsettled, the act of directing the machine was recognised as the job, not as the absence of one.

The client-facing task was not automated — it was deleted

This is the one I did not expect. Marketing and sales on my line do not file requests and do not wait for a slot. The line runs on its own, finished pieces land in a library, and they open it and take what they want. When they use something they mark it used, and that signal feeds back into what gets made next.

Reading and managing a client is a task that presupposes a client who briefs you. Remove the brief and the task has nothing to attach to. It was not that an AI got better at handling stakeholders; the interaction that the skill existed to handle stopped occurring.

Worth being precise about what this is and is not. It is one team's internal workflow, not a client relationship with money and a contract in it. But it is a concrete case of a task ending for a structural reason rather than a capability reason, and those are easy to miss when you are only counting what the machine can do.

Everything cited here

Each one opens the full record: its source and source tier, the dates, the scope it applies to, who verified it and what it does not establish.

1. Logging footage and assembling a rough cut — task on Video editor
2. Netflix told shareholders GenAI workflows had been used in roughly 300 of its 2026 titles, with the largest concentration of the work in post-production — Netflix (Q2 2026 shareholder letter)
3. Versions, formats and captions — task on Video editor
4. Finding the story and the rhythm — task on Video editor
5. Coca-Cola's 2024 holiday campaign reimagined its 1995 'Holidays Are Coming' spot with films made by three AI production studios using generative video tools — The Coca-Cola Company — media centre, 2024 holiday campaign
6. Directing generated footage — task on Video editor
7. IATSE's 2024 Basic Agreement added Article XLIX, keeping work done by prompting or overseeing an AI system inside covered union work — IATSE — 2024 Basic Agreement MOA, Article XLIX (fully executed; copy hosted by IATSE Local 728)
8. Reading and managing the client — task on Video editor