Automate the stages where the objective is unambiguous and the output is checkable in seconds: footage logging and tagging, script drafting, shot selection, voiceover, captions, ratio exports, file naming and trafficking prep. Keep the offer, the claim guardrails, and the approval decision with a person, because each of those is a commitment your company is making rather than a task being completed. Everything else is negotiable, and the table below gives our verdict on each stage with the reason attached.
The production chain, stage by stage
This is the chain as most teams actually run it, from a request landing in a channel to a decision about whether the ad worked. "Agent" means a model decides within constraints you set, which is different from automation running a sequence you defined.
| Stage | Verdict | Why |
|---|---|---|
| Brief intake | Automate the form, not the thinking | A structured template beats a Slack thread. The fields still need a human to fill in with real answers. |
| Defining the offer | Keep human | Discount depth, bundle structure and shipping thresholds are margin decisions. No tool has your unit economics. |
| Claim and compliance guardrails | Keep human, then reuse | Write them once as a constraint block. After that they are an input, not a review step. |
| Footage logging, transcription, tagging | Automate, no exceptions | The most boring, most valuable stage. This is where the hours actually go, and machines are better at it than tired humans. |
| Concept generation | Agent, with a human picking | Generating twenty angles is cheap. Choosing which three deserve spend is judgement. |
| Script drafting | Agent | Drafting is fast and revision is faster than writing from nothing. Native-language drafting beats translating a finished English script. |
| Shot selection | Agent | Only works if tagging happened properly. Bad tags produce confident, wrong shot choices. |
| Voiceover | Agent, with a fixed voice | Synthetic voice is fine for most performance work. Lock one voice per brand or your batch sounds like six different companies. |
| Music selection | Automate from a shortlist | Let the tool pick from a licensed set you approved. Do not let it pick from the open internet. |
| Captions and subtitles | Automate, then proofread | Near-perfect on clean audio, unreliable on product names, brand names and numbers. |
| Ratio exports | Automate | 9:16, 4:5, 1:1 and 16:9 from one assembly. There is nothing to think about here. |
| Safe-zone checking | Automate | Text sitting under the TikTok UI is a mechanical error and should be caught mechanically. |
| Brand overlays and end cards | Automate from a locked asset | Template work. Determinism is the point. |
| File naming and versioning | Automate | Do this early. Your future self reading a results table will thank you. |
| Quality review | Keep human, with a checklist | Covered below. This is the stage teams try to automate and regret. |
| Approval | Keep human, named individual | Somebody signs off. It is a liability decision, not a workflow step. |
| Upload and trafficking | Automate the prep, keep the button | Bulk sheets and naming, yes. Auto-publishing unreviewed creative, no. |
| Reading results | Automate the reporting, keep the call | Dashboards should compute. Deciding whether a 12 percent lift on 400 clicks is real stays with you. |
| Kill and scale decisions | Keep human, informed by rules | Frequency above about 3 usually means fatigue is arriving. Acting on it is a choice. |
| Library upkeep | Automate the mechanics, schedule the shoots | Deciding to film twenty seconds of unglamorous product-in-use coverage is a human call that improves every future variation. |
Why the offer and the guardrails stay human
These two get lumped together because they share a property: both are statements about your business that a model has no access to.
The offer is arithmetic your tool cannot do. Whether 20 percent off beats free shipping depends on your basket size, your margin, your return rate and what a discounted cohort does to lifetime value. We have watched teams let a tool generate offer language, and the failure is not that it writes badly, it is that it writes a promise nobody costed.
Guardrails are the same shape. What you may claim, what needs a disclaimer, which comparative statements your legal team has blocked, which country needs which wording. A model does not know these and will not invent them correctly.
The move that saves the most time is to write both of these down once, in a reusable block, and treat them as brief inputs rather than review notes. A constraint in the brief is enforced on every variation. A note in review is enforced on the one variation somebody happened to notice. That distinction is most of the difference between a creative process that scales and one that produces a steady trickle of near-misses. Our agency creative workflow writeup has the version of this that survives multiple clients with different rules.
Why approval stays with a named person
Approval is not a stage in a workflow, it is somebody accepting responsibility for what a company said in public with money behind it. Automating it removes the only step whose entire function is accountability.
There is a practical argument too. The stages before approval are now fast enough that a batch of 20 variations can exist within an hour of the brief. Speed at that end of the chain is worthless if the wrong thing gets out, because a pulled campaign, a platform-level policy strike or a regulator letter costs more than a week of production ever did. Keep the gate, and name the person who owns it.
Where you can compress is the shape of the gate. Reviewing 20 variations one at a time in a player is slow and error-prone. Reviewing them against a fixed checklist is 30 to 45 minutes and catches more.
What automating review actually costs you
Three specific things, in order of how badly they bite.
Errors that carry spend. Wrong price, expired promotion, discontinued packaging, a competitor visible in the background, a claim your market cannot legally make. Every one of these is invisible to a generator and obvious to a person who knows the account.
The rejection signal. The variations you throw away are the most useful data in the process. If nobody looks at them, nobody notices that four of six batches drifted toward the same hook, or that the tool keeps reaching for the same three clips because the rest of the library is badly tagged. Reviewed rejections turn into better briefs. Unreviewed rejections turn into nothing.
Calibration. After a month of reviewing output you know what this tool does well and where it fails, so your briefs get sharper and your review gets faster. Skip review and you never build that, which means you are permanently guessing about your own pipeline.
The one part of review you can safely automate is the mechanical layer: spellcheck, safe zones, duration limits, loudness, aspect ratios, presence of required disclaimers. Automate all of that, and spend the human attention on claims and accuracy. More on how we split this in our notes on running an ad agent inside an automation stack.
Where Genyad sits in this chain
Disclosure: Genyad is ours. It covers the middle of the table. You upload footage once, it transcribes and tags every clip, and each variation is a fresh script, shot selection, voiceover, caption set and export drawn from that library rather than a recut of the same timeline. Scripts are written natively in English, German, French, Spanish, Italian or Hindi. Exports are 9:16, 4:5, 1:1 and 16:9 at 1080p with no watermark on any plan, and editing, re-exporting and uploading footage cost nothing.
It does not cover the ends. There is no product-URL import and no product-feed or CSV template rendering, so it cannot pull your offer from a page. There is no direct publishing to Meta or TikTok, so you export and upload, which also means the approval gate stays where we think it belongs. There are no AI avatars, no static banners and no predicted performance scores. If you want the definition we are working from, our glossary covers what an AI agent for video ads does.
Frequently asked questions
What is the highest-value stage to automate first?
Footage logging, transcription and tagging. It is pure overhead, it is where teams lose entire days, and every downstream stage gets better when it is done properly. Automating script writing before your library is tagged produces confident scripts referencing shots you do not have.
Can I automate creative approval if I trust the tool?
We would not, regardless of the tool. Approval exists so that a named person is accountable for a public claim backed by spend, and no model can hold that. Compress the gate with a checklist instead of removing it.
How long should reviewing a batch of AI-generated ads take?
Around 30 to 45 minutes for 20 variations if you work from a checklist covering claims, prices, dates, product accuracy, on-screen text and concept distinctness. If it is taking materially longer, the usual cause is missing constraints in the brief, so the same corrections keep reappearing.
Does automating production mean I need fewer people?
In our experience it means the same people spend their time differently. Production hours fall sharply, review and judgement hours rise, and the constraint moves to how fast someone can decide what is worth running. Teams that cut headcount to match the production saving usually end up unable to approve what they can now produce.