Let anyone put a human presenter in a video, without filming one.
Enable users to add AI-generated avatars to their scenes easily, creating engaging, human-like presenters without needing to film or hire talent — while keeping full creative control over how the presenter appears.
Stock presenters felt generic. Real ones cost too much.
The problem was never generating a face. It was making a generated face feel like it belonged in this video.
Discovery 01 · User interviews
“That's amazing. If I could get an AI picture of myself and just me talking, and not having to do it — that would be such a lifesaver.”
Marketer · user interview
Discovery 02 · Competitive review
Four avatar platforms reviewed across 2023–24. All offered stock libraries with limited customization; the advanced tools pushed personalized avatars and cloned voices. The shared weaknesses were poor lip-sync and inflexible in-scene editing — which is exactly where an editor-first product could win.
Six steps from an empty scene to an animated presenter. The last two — reuse across scenes, and animated preview — are the ones research asked for loudest and the ones that shipped behind a beta flag.
Built inside SundaySky's system, not beside it.
Every new component is a thing somebody has to maintain. Two additions were justified; a third would not have been.
Decision 01 · Add & replace
An avatar library in the left panel, matching the media library, with three tabs: custom avatars with credit-based upload, stock avatars, and used in video for quick reuse.
Consistency with the media workflow made avatars discoverable without teaching a new pattern — and answered the reuse constraint directly.
Decision 02 · Preview
The avatar preview matches existing media previews and thumbnails. Users play a predefined text sample to check lip-sync and appearance, then adjust the voice later from the voice library.
Familiar across every asset type, and it lets someone test a presenter before spending a credit animating one.
Decision 03 · Position
Move, resize and align an avatar inside the scene using the same handles and toolbar the canvas already uses for every other asset.
Precise control to fit a presenter into any layout — and the single most-used interaction after launch. 73% of users adjusted the avatar immediately.
Decision 04 · Framing
Two distinct view modes rather than an open crop tool, so the avatar can be adjusted to fit a layout or a storytelling need.
Two modes cover the layouts people actually build, and keep framing predictable when the same presenter appears across several scenes.
Decision 05 · Voice
A voice dropdown inside the narration panel, available whenever a video contains an avatar — including scenes where no avatar is visible.
One presenter across the whole video. The voice is what the feature actually sells, so it could not be trapped inside the avatar's own settings.
Measure 02 · Path analytics, Pendo
Adding and positioning an avatar felt simple — nearly three quarters adjusted it straight away, which says the entry flow and the scene controls landed.
Preview and apply were underused, and frequent replacement revealed quality and trust issues further down the flow.
Measure 03 · Interaction, Pendo
Of 185 avatars added, 68 were replaced — a replacement rate close to the preview rate. Users were sampling rather than committing.
Low preview use also points to a discoverability problem: the step that would have built confidence is the one people skipped.
Measure 04 · Session replay, FullStory
Across 23 sessions, most users moved through the flow smoothly. The friction was concentrated in one place.
In one recording a user clicked the avatar thumbnail expecting it to play, then pressed play and saw no animation. The flow was intuitive; the preview's behaviour did not match what people expected of it.
This is the same gap the click data showed from the other direction. Two methods, one finding — which is when I stopped treating it as a hunch.
← → to move · F for full screen · Read the case study