What we learned making 85 AI UGC videos for small businesses (our own data)
Numbers from our own pipeline since late September 2026 — video length, render success, and the script problems we fix most before a video is made.

Most articles about AI UGC are opinions. This one is our own numbers: what happened when MRKT DEPT planned, scripted, rendered and posted short-form videos for small businesses on its current production pipeline, from 27 September to 20 October 2026.
The headline numbers
| Measure | Result |
|---|---|
| Videos rendered successfully | 85 of 100 render attempts (85%) |
| Posts published to TikTok | 62 |
| Average video length | 26.0 seconds (shortest 22.0s, longest 29.3s) |
| Average render time | 11.4 minutes per video |
| Spoken words per 3-beat video | about 73 (roughly 24 per beat) |
These are small numbers from a young product, and we will update them as volume grows. They are still useful, because they show what it takes to get a usable AI UGC video — not just what a demo looks like.
1. 20–30 seconds is the sweet spot for one idea
Every video we published runs between 22 and 30 seconds. That is enough for a hook, one useful point and a call to action — about 70 spoken words. Longer scripts mostly added repetition; shorter ones did not have room for the point. If you are writing TikToks for a business, aim for one idea in 25 seconds.
2. The script needs more fixing than the video
Most of the work happens before anything is rendered. In the same period our script checks:
- regenerated a script 91 times because it repeated a post the account had already published,
- rewrote an opening line 201 times to make it more specific,
- added a missing call to action 56 times, and
- dropped a repeated sentence inside a beat 49 times.
The lesson for anyone using AI for video: the model will happily say the same thing twice, reuse last week's angle or end without telling the viewer what to do. Check for those three things before you render or post.
3. One in seven renders still fails — and that's why quality gates matter
15 of 100 render attempts failed. The causes were split between the automated quality gate (audio that ended too early, frozen frames), the render machine itself, and other errors. A failed render is retried or re-queued rather than posted, which is the point: a video that would have embarrassed the business never reaches the feed.
4. Consistency beats volume
62 posts in about three weeks came from a fixed rhythm: plan ahead, render in batches, review, publish on a schedule. The businesses that keep posting are the ones that never have to decide what to post on a Monday morning.
What this means for a small business
- Keep each video to one idea in about 25 seconds.
- Spend your effort on the hook and the script, not the editing.
- Check every AI script for repeats and a missing call to action.
- Post on a schedule you can keep for months, not a burst.
If you want this done for you, MRKT DEPT plans, writes, renders and posts videos like these after you review them — get a free sample video, or see real examples and 80+ TikTok hook examples.
*Method: render, length and publishing figures come from MRKT DEPT's production records for customer accounts (test accounts excluded), 27 September – 20 October 2026; render attempts include retries. Script-check counts come from our pipeline logs for the same period.*
Get your social media done for you
MRKT DEPT plans your posts, makes the videos with an AI spokesperson and posts them to TikTok and Instagram. You review every post. Your first video is free.
Read next
Free tools: TikTok hook generator · caption generator · script generator · hook examples · Ideas for your industry · See real videos