The trailer that knows what you watched.
How Punch built AI-avatar email campaigns for Netflix — a personalized video for every lapsed subscriber, cut from the shows they loved and the ones they haven't met yet.
NETFLIX · MEDIA ML · EMAIL · NETFLIX.COM ↗
One actor. Every emotion. Infinite reads.
The avatars started as people: green-screen capture sessions where presenters read scripts in distinct emotional registers. Each performance was trained against a labeled emotion class, so a generated face carries the right affect for the moment — a win-back nudge gets the warm read; a "your shows are leaving" teaser gets urgency. Same avatar, different register, chosen per script.
A model that learns how you talk, not just what you say.
The capture footage trains a custom Punch model on the presenter's lip and head movements, mannerisms, and gestures — paired with semantic analysis that maps each emotional state to script context. That mapping is what gives the delivery its dynamism: the avatar doesn't recite a line, it performs it in the register the sentence calls for.
With movement and emotion learned, audio and likeness become fully synthesizable — every personalized win-back video rendered from a script the avatar has never read, sounding and moving like the person who trained it.
No actor. No reshoot. Every emotion on cue.
The script doesn't just say the words. It scores them.
The copy itself is generated per subscriber from their taste and history — but generation doesn't stop at words. Every sentence is tagged with an emotional register, and that tag is the instruction the avatar performs to. A win-back line lands warm; a "your shows are leaving" line lands urgent. Click a line to see how it's delivered.
Relaxed, low-stakes open — the avatar eases in with soft eyes and an unhurried cadence. No pitch yet.
A win-back reel built for an audience of one.
Every signal — what they watched, what they loved, where they dropped off — flows through a single automated pipeline that writes the script, performs it as the avatar, renders it into the scene, and lands it in the inbox. One system, end to end, producing a win-back reel no other subscriber will ever see.
One play button. One reason to come back.
A reel built from their taste lands in the inbox. They press play — and a page opens that greets them by name and makes the ask, personally. Press play below to walk it.
"Come back" means nothing. "Come back — season two of the show you binged just dropped" works.
Win-back emails are the weakest message in the funnel: generic, ignorable, sent at scale. Netflix's viewing data made something better possible — a video rendered for each lapsed subscriber: trailers and teasers matched to what they actually watched, fronted by an AI avatar making the ask to renew, personally.
Punch built the pipeline: taste signals in, a per-subscriber script and cut out — avatar narration, title cards, and the teaser reel rendered automatically for every recipient.
Case study PDF ↓IN THE NEWS
ABOUT.NETFLIX.COM
The Netflix newsroom. ↗
NETFLIX TECHBLOG
How Netflix builds — the tech blog. ↗
SEND PREFERRED ARTICLES · [CONFIRM PICKS]
