What a year of testing AI video tools taught us about the job we already had
We have spent much of the last year testing AI video tools to understand where they create real value and where they don't. Like many people working in production, we were skeptical of the idea that a single prompt could replace the craft behind effective communication. What surprised us wasn't that AI replaced production, but how quickly it became part of it.
The tools getting genuinely useful aren't the ones that generate a better single clip. They're the ones that support an actual production: storyboarding a sequence before committing to it, defining a set of characters and environments once and reusing them across shots, reorganising scenes after seeing how they play together, and only then rendering. That's not a prompt box. That's a small production pipeline, and it looks a lot more like the process we already had than the “type a sentence, get a video” story everyone was telling a year ago.
References beat prompts, every time
The clearest finding from a year of testing: when we describe what we want in words alone, AI fills the gaps with its best guess, and the guesses drift: a character's face changes slightly between shots, a logo gets reinterpreted, a colour palette wanders. When we give it something to look at instead: reference footage, style frames, a locked set of visual assets, the drift mostly disappears. The model isn't inventing an answer anymore. It's matching one we already approved.
That's a small shift in how you work with the tool, but it's the difference between output you can put in front of a client and output you're quietly regenerating for the sixth time.
What a year of testing AI video tools taught us about the job we already had
We have spent much of the last year testing AI video tools to understand where they create real value and where they don't. Like many people working in production, we were skeptical of the idea that a single prompt could replace the craft behind effective communication. What surprised us wasn't that AI replaced production, but how quickly it became part of it.
The tools getting genuinely useful aren't the ones that generate a better single clip. They're the ones that support an actual production: storyboarding a sequence before committing to it, defining a set of characters and environments once and reusing them across shots, reorganising scenes after seeing how they play together, and only then rendering. That's not a prompt box. That's a small production pipeline, and it looks a lot more like the process we already had than the “type a sentence, get a video” story everyone was telling a year ago.
Less prompting, more directing
Put those two things together and the job looks less like typing instructions and more like the job a director already does: define the vision, set the visual rules, decide what's allowed to vary and what isn't, then let the tool fill in the rest inside those boundaries. The creative decisions didn't move to the AI. They moved earlier: into the brief and the reference set, before a single frame gets generated. In practice, that splits pretty cleanly in two:
Where AI earns its place
- Fast to script and refine: a scene that isn't working gets caught before a minute of render time is spent, not after.
- Makes video possible for things that were never filmable: hypothetical products, future scenarios, situations too costly or risky to recreate in traditional production.
- AI voiceover means updating a script; it doesn't mean booking a studio again.
- Genuinely strong for early drafts and concept testing, before anyone commits budget to a direction.
Where it still needs a person
- Output quality tracks the reference material, not just the prompt: thin references still produce thin results.
- Highly specific or detailed content is still hard to pin down exactly.
- Every generation costs tokens or credits, so iteration has a real, visible price, unlike a supervised render.
- Someone technical or creative enough to judge the output against reality is still required. That part hasn't gone anywhere.
Want to stay in the loop?
Get occasional newsletters from Cadpeople.
