Meta support assistant
From AI feature to full AI support system
From 2024 to 2026 I worked as the sole designer for Meta's support assistant for the meta.com Store & Device Help Center, across two generations of the product.
Gen 1 had two rules: Don't redesign the help center, and don't scale beyond three automations.
We shipped it fast so we could validate our direction. But, this meant that the AI felt like an added feature, since you needed to navigate away from any page to use it.
An important part I worked on was AI entry points on articles, adding an big blue sticky footer button, which worked, but also meant you could only type your question after navigating to a new page.
So I put a composer directly in the article, but it tested worse than a sticky blue button. I decided to keep the composer, but added a radial glow entrance animation, based on the assistant's color palette. Engagement came and exceeded the blue button approach.
You could now start interacting with AI on articles, but people come to a help center when they're upset and low on patience. So, I pushed harder by making the page you land on the actual AI, with a twist.
I called this new pattern a job hub.
The page is the AI. There's a composer, and AI works on the same page, but what you land on are more traditional UI affordances: widgets, cards, things arranged by importance. Engage with any of it and the page scrolls down into a chat.
The article becomes an expandible summary card near the top of the page, with seed prompts and a quick way to contact support beneath it.
Selecting any prompt starts a chat on the same page, below the content, with the article acting as context for both the user and the model.
While job hubs are the form AI takes, I also defined how it behaves.
In gen 1 it could only do 3 actions; if something you asked for matched them, it'd answer with a button that launched a flyout with the same linear flow that already existed.
It technically worked, but it was mostly just launching a form, and it would be impossible to scale.
At the same time, the org was adopting an automation platform that would enable us to design AI as front-line support people. Engineering was pushing for a UX of text responses and back-and-forths with widgets one at a time, even if that meant asking 6 sequential questions.
So I designed the pattern system that drove the entire UX, from how widgets behave, grouping sequential user inputs into forms, and handling flow interruptions, like answering ahead, asking a clarifying question, or even changing the subject.
The part I am particularly proud of was the trust tier system. It always asks for permission for consequential actions, and never asks for sensitive data in the conversation. Instead, we open a traditional flyout form, which is better for trust, and reduces risk for the system.
I applied these patterns to all 21 automations, each with their own branching logic and outcomes to cover, and worked to automate my workflow with engineering.
First was a cursor prototype, then a tool the team could use to draft the UX based on the patterns I defined, and finally a Claude Code skill that generates new automations, ensuring the design scales even after I move on to other work.
For device-specific support, your hardware gets its own job hub page; listing everything that could be wrong, and everything you might need to do. Engaging with any of it opens the support flow in-line.
AI takes the first shot, then handsoff to human support on the same page, preserving all of the context. No portal, no ticket number, nobody re-explaining themselves.
Not every page should be the AI, though. Landing pages are doorways, and their job is getting you through to the right destination.
So I redesigned landing pages into full AI launch pads; prominent composer with curated prompt cards below matching the actual reasons people come to help center.
Several aspects of this design have been built, several are still in-flight, to be completed in 2026. Despite that, the result speaks for itself.
95.95%
Case deflection across the automated support UX.
And to me the result that matters just a much is the support my vision garnered; taking a simple AI feature into a full AI-1st support system.