Sole designer across two generations of AI support at Meta. Gen-1 proved AI's value without touching the Help Center. Gen-2 redesigned it as AI-native.
Your Meta Quest headset stops working. You look in settings, then you search, and you probably land up on Reddit before you end up on meta.com/help/quest. Eventually you find a support article that may help, but when it doesn't, you give up on finding your own answers and go looking for a person instead.
Our research pointed to the problem: people rated conversations with live support highly. It was everything leading up to it that was miserable, including a confusing support intake flow, with notoriously low NPS. And most people would rather have a self-service option, when it covered their problem and they could find it. Usually it didn't, and they couldn't. The org also had incentives for finding ways of reducing support costs, and AI was providing a path forward.
So the thing to fix wasn't support, but the journey of getting to it, while providing more self service opportunities with AI.
Users typical fell into 3 categories: Someone waiting on an order, someone whose device stopped working and someone with an account problem. The ending is the same for all three: an article that often falls short of totally helping, a list of categories that didn't include their problem, and then an email follow-up that feels unresponsive. This was research my team inherited, but it became core to how we understood the user's pain points.
Gen-1 was minimal by design
Gen-1 : Start chat from home page composer
Gen-1 : New chat start screen & menu
We had two intentional rules: don't redesign the Help Center, don't scale past three flows.
Small enough to prove value before earning the right to transform support with AI. But the assistant lived apart from the Help Center, so asking a question meant navigating away and losing context.
The AI handled three actions, refund, order status, device return, matching intent to a button that opened a flyout with the existing deterministic flow. It made these actions more discoverable, but a mid-flow question lost your place, and it couldn't scale to real investigation.
Gen-1 : Game refund (desktop)
Gen-1 : Order status (mobile)
Gen-1 : Device return (mobile)
Getting to the AI from articles with a sticky blue button worked, but only because of brute-force noticeability by forcing users to spot it, decide, and wait for a new page to load. When I replaced it with an in-line composer, engagement dropped. The team saw it as proof of a failed concept, but I knew the test simply conflated visual prominence with actual value. The composer wasn't a bad idea; it was just invisible.
I kept the design and fixed the real issue of discoverability by adding a radial glow entrance to catch the eye. Engagement bounced back to baseline or better, and got us one step closer to transforming the page into the assistant itself.
Further user testing led us to adding an automation prompt chip above the composer to kickstart support actions with AI.
Before
A saturated button you can't miss.
After
I kept the composer and solved the real problem: a radial glow entrance, resolving on arrival and settling out of the way.
Instead of articles users land on action-oriented pages
Gen-2 : Article job hub
Articles compress into a summary card with seed prompts; the article becomes context, not a destination.
Devices got their own hub, with handoff to a person happening on the same page, session intact, no ticket number, no re-explaining yourself.
Device support became a troubleshooting center
Gen-2 : Device support page
Building trust in AI workflows
Gen-2 : AI always asks for consent with a CTA
Gen-2 : AI will never capture personal info in chat
The assistant asks consent before each consequential action, and never takes sensitive data into the conversation, opening a flyout form instead.
Results
95.95%
Support case deflection rate, end to end, after Gen-2 shipped
3→21
Self service automations, Gen-1 to Gen-2
↓17%
Global support case volume
$6.5+ Million
Projected annual savings
Merged case and chat history
Support history: merging chat and case history
Standard chat history mislabels intent by indexing only the first prompt. By analyzing the entire exchange, we can display threads as actionable task cards, letting users resume workflows rather than restart them.
This eliminates duplicate tickets caused by missing case management. Beyond support, any generative AI product needs this persistent state model to help users return to jobs in progress.
Scaling from the typing and tapping interaction patterns, when in full voice mode widgets stay visible, but their purpose changes.
Speak, and the interface becomes a visual aid showing your current state, options you can choose from, and imagery to better convey concepts.
Full voice conversations, building on the patterns
Voice mode with visual aid widgets
The best troubleshooting is an experience you never have to navigate to. Instead of forcing users to a generic web portal, this architecture uses a unified agent layer powered by device telemetry and account state to solve problems exactly where they happen. The UI surface is simply where the agent reaches you. I pitched this localized approach to the Ray-Ban Display team to establish early interest in moving support directly to the point of impact.
