Put review gates where the money is
Approval at the end is approval after the spend. A four-gate structure — concept, key frame, motion, export — and exactly what locks at each one.
Most review processes put one gate at the end, because that is where the deliverable is. It is the wrong place. By the time there is something to look at, every decision that could have been changed cheaply has already been paid for, and the review is no longer an approval — it is a post-mortem with an invoice attached.
The rule that fixes it is short: a gate belongs immediately before each irreversible spend, and nowhere else. Everything else is a status update, which is useful and should not be called a review.
What "irreversible" means in a generative pipeline
Irreversible does not mean expensive. It means undoing it requires generating again rather than editing what you have.
Rearranging clips, retiming a cut, changing a caption, swapping a music bed, adjusting a grade — all reversible. The editor is one timeline and it re-renders; preview: true gives a 480p pass at no credit cost with a short per-user cooldown, and the export charge lands once on the version you confirm, however many clips it contains. How previews and the final export are billed has the mechanics.
Changing what is inside a frame is the other category. A different presenter, a different set, a different camera move, a different product angle: none of those are edits. They are new generations, and everything built on the old ones — the grade, the timing, the lip-sync, the captions cut to the old pacing — goes with them.
So the spend boundaries are: before you generate stills, before you animate them, and before you export. Add the brief itself, which is the cheapest boundary of all and the one most often skipped. That gives four gates.
The four gates
| Gate | What the client sees | What locks | What is still cheap after it |
|---|---|---|---|
| 1. Concept | Script, reference board, shot list, format spec | Message, direction, look, duration, aspect ratio | Everything downstream |
| 2. Key frame | One still per scene, plus the character and product references | Casting, set, palette, composition, product accuracy | Motion, pacing, audio, edit |
| 3. Motion | Raw generated clips — ungraded, unedited, no music | Performance, camera move, physics, shot-level pacing | Assembly, grade, captions, music |
| 4. Export | 480p preview of the assembled cut | Timing, captions, music, spec compliance, provenance | Nothing — this is the last stop |
Each gate needs four things or it is decoration: one artifact, one named decider, a response clock, and a stated default if the clock runs out. The default is what stops a gate becoming a stall. "No consolidated response by Thursday means we proceed on the recommended option" is not aggressive when it is agreed in advance; it is the only thing that makes a gate compatible with a deadline. Content approval workflows that don't stall goes deeper on the queue-time problem, which is a different failure from the one this piece is about.
Gate 2 is where the savings actually are
If you only add one gate, add the key-frame gate. Stills are the cheap rehearsal for video, and the categories of failure they catch are exactly the ones that are most expensive to fix after animation: wrong face, wrong set, wrong palette, wrong product, wrong composition.
The workflow is unglamorous and it works. Generate a still for each scene, get those approved, then animate the approved frames rather than generating video from text and hoping. Storyboarding brand videos with AI images first is the practice, and image-to-video is the step that consumes the approved frame. For scenes that run long enough to drift, keyframe chaining is how the approved look survives to the end of the shot.
It is worth being honest about the numbers underneath this, because the ones circulating are not measured. Community-reported figures put usable output somewhere around one clip in four, or three to five attempts per usable clip, with harder cases — specific character action, complex motion — running worse. Those are aggregated anecdote rather than published research, and no credible measured hit-rate study exists. The operating principle survives the weak data, though: the metric that matters is attempts per usable clip, and a gate that stops you animating a frame nobody liked removes an entire block of those attempts. Reroll rates and the shots you throw away is where to start instrumenting your own number instead of borrowing one.
Show raw at gate 3, finished at gate 4
The most common way a motion gate fails is that it is presented too well. Music, captions and a grade on a set of clips that have not been approved as clips means the client is now reviewing the wrong variable. They will respond to the edit — the thing you have not asked about, and the thing you can still change late — and the shots will pass unexamined because the package felt finished.
Send gate 3 raw and say so: "These are the shots on their own, no music or captions yet. I'm looking for a yes on the performance and the camera before I cut them together." That sentence does most of the work. It tells the reviewer what to look at and, more importantly, what not to comment on yet.
Gate 4 inverts it. The assembled cut goes out as a preview, with the question narrowed to timing, text and spec.
Making a gate cheap to hold
A gate is only sustainable if passing through it is fast for the client and holding it is cheap for you. Three things do most of that work.
Estimate before you dispatch. Knowing what the next block of generation will consume is what turns a gate from a ritual into a decision with stakes. Estimating credit cost before you dispatch a batch covers the arithmetic, and it is worth checking the balance too when a batch is large enough that running dry halfway would be worse than waiting a day.
Make review a tap. The single biggest cause of a stalled gate is an account wall. Asking the agent to share a generation with a link mints a link with a rich preview card that opens without a login, which is the difference between a reviewer answering on their phone in a queue and answering next Tuesday.
Write down what locked. A gate that does not produce a record produces the same conversation twice. The output of every gate is a line in a decision log: what was approved, what was rejected, who decided, and the link to the artifact they were looking at.
Where gate structures go wrong
Gates on the calendar instead of on the spend. A Tuesday check-in is a status meeting. If nothing irreversible happens after it, it does not need an approval.
Gates with no decision right. If the person at the gate cannot say yes on their own, the gate is a queue with a nicer name. Name the decider by role at the start of the project.
Too many gates. Four is enough for a single deliverable. Adding a fifth typically means splitting the key-frame gate per scene, which converts one decision into six and multiplies the waiting rather than the quality.
A gate that reopens a locked decision. If gate 3 notes ask for a different presenter, that is not a note — it is a direction change, and it belongs in a commercial conversation rather than a production one. The whole point of writing down what locks at each gate is that this classification is settled before anyone is annoyed.
FAQ
Does this work for a single 15-second ad, or is it only for bigger jobs?
It scales down cleanly. On a short deliverable, gates 1 and 2 often collapse into one conversation — script plus a couple of stills in the same message — and gates 3 and 4 stay separate, because the raw-versus-finished distinction is what protects the shots. What you should not do is collapse 3 and 4, which is the same as having no motion gate.
What if the client wants to see everything as it happens?
Give them visibility without giving them a vote at every step. A shared project folder or a running link list satisfies the impulse; the gates stay as the four moments where a decision is actually requested. Confusing visibility with approval is what produces feedback on unfinished work.
How long should the response clock be at each gate?
Shorter early, longer late. One or two business days at concept and key frame, where the client is reading a page and looking at stills; three to five at export, where legal or a brand team may need to look. State the clock and the default in the same sentence you send the artifact.
Who should the named decider be if the client has three stakeholders?
One of them, chosen by the client, with the other two feeding into that person. If they will not choose, consolidate the notes yourself and send back a single list for confirmation before you act on it. That written confirmation is the artifact that makes the gate real.