Generative AI is a good fit for work where the answer is open and a person will shape the result, and a poor fit for work where the answer is already known and has to be reproduced exactly. That sorting rule is more useful to a producer than any list of what a model can currently do, because the list goes out of date and the rule does not.
This article is written for the video producer who has a job board full of real work and a question from above about where “AI” fits. We will be plain about our position. Dataclay makes Templater, a data-driven templating plugin for Adobe After Effects, which is not a generative model. Our own pages describe generative AI at a high level and do not test or rank generative video tools, so we are not going to hand you a scoreboard. What we can do is give you a way to sort your tasks that holds up as the tools change, and to say which parts of it come from our documentation and which are our judgment.
Sort by one question: is there a right answer already?
Every task on a production job board falls somewhere on a line. At one end the output is undetermined. You do not yet know what the piece should look like or say, and the work is to find out. At the other end the output is fully determined. You know exactly what the piece is, and the work is to produce it correctly, many times.
Our post on generative AI and media automation describes the two sides in these terms. Generative AI, it says, uses machine learning to create new content such as text, images, audio, and video. Media automation uses software to automate repetitive tasks in media production, such as editing, encoding, and distribution. One makes new things. The other does known things over and over.
That is the first sort, and it already answers a surprising amount of a job board. Open-ended tasks suit a tool that makes new things, and closed, repeated tasks suit a tool that does the same thing reliably. The mistake we see most often is applying the creative tool to a repeated task because it is new and exciting, and then spending the savings on checking.
The rest of this article walks through task types in that order, from the most open to the most closed. Where something comes from our documentation, we say so. Where it is our judgment as people who work with production teams, we say that too.
The open end: where generation earns its place
The clearest sourced example is text. Our post describes a generative model that writes unique product descriptions for an ecommerce platform, with those descriptions then fed into Templater so that the descriptions become thousands of product videos. In that picture, the generative step does something a spreadsheet cannot do on its own, which is write a sentence about each product. The templating step does something a model should not be asked to do, which is put that sentence in exactly the same place, in the same type, on every video. The producer’s job is to approve the sentences and the template, and neither is skipped.
Beyond text, here is our judgment, offered as general practice and not as measured fact. Generation tends to earn its place in early, exploratory work, where being wrong is cheap and a person is going to look at the result anyway. Producing a range of directions for a treatment, drafting alternate headlines, producing quick visual references to show a client what a mood might be, and making rough placeholders to hold a timeline while real assets are produced are all tasks where a draft with flaws is still useful. In these tasks the output is not the deliverable. It is an input to a person’s decision.
That distinction, between output as input and output as deliverable, is the most helpful test we know. If a person will look at it and decide, a rough result is fine. If it goes straight to an audience, the standard is different.
There is a second category on the open end that is easy to miss, which is help for the person doing the work. Drafting a first pass at a shot list, a call sheet, a description for a platform, or alternative wording for a title, all of these are text tasks around the video and not the video itself. They fit the same rule. A person reads them, fixes them, and owns them.
We would add one caution about the open end, and it applies to every tool. Whatever generation produces has an author problem. If it is going out under your name or your client’s, someone has to be able to say who checked it and against what. Open-ended work does not remove the need for that sign-off. It only changes what is being signed.
The closed end: where generation is the wrong tool
On the other end of the line sit tasks where the correct output is known. A lower third that must show this person’s name, in this font, in this position, with this animation. A price that must read exactly the number in the system. A sports score graphic that has to show the score as it is. A regional version of an ad with a legal line that must match local rules word for word. A set of two hundred personalized videos where each has to carry the right name.
These are the jobs our own product exists for, so we are not neutral, and you should weigh that. The reasoning does not rest on our product, though. The reasoning is that a task with a known correct answer has no use for a tool that makes new details each time. The best case is that it reproduces the right answer, which a template does by construction. The worst case is that it produces something plausible and different, which you then have to find. Finding a small wrong detail in a large amount of otherwise good video is the most expensive kind of review there is.
Here is what the template side contributes, from our documentation. Templater reads each row of a data source as a separate render job, customizes the template for it, and renders it. Layout rules adjust dynamic layers relative to the composition and the layers around them. Time Sculpting adjusts the timing of other elements when source footage changes length. Spot checks produce still images or clips of chosen frames from the data-merged compositions so you can see how the data landed before you render everything. Templater Bot can watch a data source and process new rows in the background. None of that makes anything new. It makes the same thing correctly, at volume, with a way to check it.
Be careful with the middle of the line as well, because that is where producers get caught. Many tasks sound creative and are in fact closed. Captions are a good example. They feel like writing, but the words are fixed by the audio, and the work is to put them on screen accurately and on time. Our Dynamic Subtitles rigging utility automates the addition of subtitles, closed captions, and translations, driven from data. The risky part is not the layout, it is the words, and that is a human check whether the first draft of the words came from a person or a machine.
A producer’s week, sorted
An example helps here, so picture a producer at a mid-size brand on a Monday with a board of eight jobs. We will sort them using the rule above, and we will mark where our documentation supports the call and where it is our judgment.
The first job is a set of social cutdowns of a finished spot into several aspect ratios. The content is known and the task is repeated, so it sits at the closed end. A template and a render queue are the tool, and generation has no role beyond maybe suggesting where to crop, which a person would approve.
The second is a batch of product videos from a catalog of a few hundred items, with a name, an image, and a short line each. This is closed for the video and open for the line. The line, as in our documented example, can be drafted by a generative model and reviewed by a person. The video is a template fed by a sheet. Both tools appear, in sequence, with the human review between them.
The third is a pitch for a new campaign that needs three visual directions by Thursday. This is the open end. Generating references and rough motion studies is a reasonable use because a person will look at them and choose. Nothing generated is going to air. The risk is low and the benefit is speed.
The fourth is a legal-reviewed disclaimer that has to appear on every regional version of a promotion. This is closed, and with a high cost of error. It belongs in a template as fixed text, reviewed once.
The fifth is a tribute video for an internal event that is made once, has a strong emotional point, and has no repetition. This is a job for people. A template has no payoff for a single piece, and generation produces something that sounds generic exactly where the piece needs to feel specific.
The sixth is translation of an existing explainer into several languages. The text is open, since translation involves judgment, and the video is closed, since it is the same layout with different words. A draft of the words can come from a tool, a fluent person has to review it, and the versions come out of a template. This is the same shape as the second job, with the review more important.
The seventh is choosing music for a series of automated news clips. We include this one because the study in our news video post makes a point about it. In a study covered in our post, audiences rated music selection as less suited to the story in highly automated videos than in human-made ones, where the software required an operator to choose from the tracks the interface offered. The finding is about an automated platform and not about generative models, and we do not want to stretch it. The lesson that carries over is that choices that look like details, such as which image or which track, are where audiences noticed the difference, and they are the choices that deserve a person.
The eighth is a weekly sports results graphic from a data feed. Closed, repeated, and time-pressured, with correctness mattering more than novelty. A template, a feed, and automation are the tool.
Look at that board again. Three of eight were fully closed, two were open on one side and closed on the other, one was open, one was for people alone, and one was a hybrid with review at its center. The pattern on a real board looks like that, and almost none of it is a single tool used alone.
How to test any generative tool on your own work
Because we do not benchmark generative tools, and because they change, the most reliable thing we can offer is a way to test one yourself, which is general practice and not something we measured.
Use your own material, not the vendor’s. Take a real job that has already shipped, so that you know what correct looks like, and give the tool the same brief. Compare the result with the shipped version for the things that matter to that job. For a brand piece, check the logo, the type, the colors, and every number. For a news piece, check whether the images relate to the story, which the study we cite found to be a weak point for automated video. For anything with people in it, look at faces, hands, and text inside the frame, and look more than once.
Then run it again with the same input, and look at what changed. For an open task a difference is a feature, since you may want options. For a closed task a difference is a defect. How much it varies on repeat tells you more about where it belongs on your board than any demonstration.
Finally, time the whole path and not the generation. Include prompting, retries, selection, correction, and approval. The number that matters to a producer is hours from brief to approved delivery, and a tool that is fast at the first step and slow at the last has not changed that number.
The bottom line
Ask of each task whether a right answer already exists. If it does, use the tool that reproduces it and lets you check it, which for repeated video is a template fed by data. If it does not, generation can help you find one faster, as long as a person decides. Most real jobs are mixed, and the skill is in cutting them at the right seam, generating the open part, reviewing it, and templating the closed part.
We have deliberately not told you which products are good at what. We do not have a basis for that, and a claim about a specific tool would be out of date before you finished reading it. The sorting question will still be useful after the tools change.
For more, read our post on generative AI and media automation, our piece on automated news video and the human touch, and our guide to making hundreds of personalized videos from data. The Templater page describes the editions, and our post on dynamic subtitles covers the captions example. For quality checks on batches, see spot checks.
These are the questions producers ask most when they start sorting their own boards.
Frequently Asked Questions
What kinds of video tasks suit generative AI?
Open-ended tasks where a person will review and decide: exploring directions, drafting copy, producing rough references, and writing text such as product descriptions. Our own post describes a generative model writing product descriptions that then feed Templater. The test we suggest is whether the output is an input to someone’s decision or a deliverable that goes straight to an audience.
Which tasks should not be generated?
Tasks with a known correct answer that must be reproduced exactly: names, prices, scores, legal lines, logos, and personalized versions at volume. A tool that produces new details each time can give you something plausible and different, and finding a small error in a large amount of good video is expensive. For these jobs, a template fed by data with spot checks is a better fit.
Does Dataclay rank or benchmark generative video tools?
No. Our pages describe generative AI at a general level and do not test specific products, so we make no claims about what any particular tool does well or badly today. We suggest testing a tool on a job you have already shipped, comparing it against the final version, repeating the same input, and timing the whole path from brief to approved delivery.
Is automated video as good as human-made video?
A study covered in our post on news video found that audiences liked human-made and partly automated videos about the same, and liked highly automated ones significantly less. The partly automated videos had been edited by humans. The study looked at automated news video from a platform, not generative models in general, so it should not be read as a verdict on every tool.
Can I use both generation and templating on one job?
Yes, and it is the common shape of real work. Let a model draft the open part, such as the text, have a person review it, and merge it into a template so the closed part, layout, type, timing, and branding, is reproduced exactly. Templater reads each row of your data source as a separate render job, so the reviewed text becomes the input.
How do I decide if a task is worth building a template for?
Ask whether the task repeats and whether the right answer is known. A template takes upfront design and rigging, so it pays back with volume or frequent change. For one-off, emotional, or highly specific pieces, a person making the video by hand is usually the better choice. For a weekly graphic or a catalog of versions, a template and a data source are worth the effort.
