Meta AI support assistant

I was the primary conversation and content designer for three 0-1 AI support assistants, built in collaboration with Product Design, Engineering, Legal, and other cross-functional partners across Meta. All of the AI assistants exceeded the team’s goals and formed the basis for today’s Meta AI support assistant. Today I’m the primary content designer for routines related to enforcements in the Meta AI support assistant, ensuring a scalable design system.

Our team’s goal is to support users who are flagged for breaking the rules and now facing restrictions (e.g. Marketplace seller flagged for selling counterfeit handbag and now at risk of losing Marketplace access). UX Research and data suggest that users struggle to understand what happened and what they can do.

Agentic AI presented a new lever for tackling these pain-points. Our hypothesis was that by offering personalized agentic support that could resolve the user’s issue on the spot, we could improve resolution rates and sentiment.

Intuitive conversation design

• I worked closely with Product Design, strategizing the UX and response framework simultaneously so the conversation could unfold intuitively and help the user understand what happened, why, and what their options are.

• When we began, we didn’t know how best to structure the response framework, so Engineering and I agreed that I would take a stab, then we’d do trial and error to find the sweet spot. Engineers found that my initial response framework was too comprehensive, and would overwhelm the AI. They proposed a more streamlined approach tied to the back-end architecture, and I adjusted the framework to match. This yielded flexible and consistently high quality responses.

• I rooted the response framework in conversational quality guidelines developed for human customer service agents, which I knew was being leveraged for LLM judges and would ensure a consistent experience across assistants.

• During experimentation, I analyzed live conversations and strategized content solutions for recurring issues, which Engineering implemented.

Help Center audit:

• My experience on the Help Center Team made me realize that we were missing a critical lever for improving the conversational output: the accuracy of Help Center articles, funneled through the RAG. I led an audit of the articles relevant to our use-cases, correcting facts, adding flows, updating terminology, and securing stakeholder approvals. 

• This approach is now being replicated by other Meta content designers working on AI support assistants.

Canned prompts:

• I analyzed data from user feedback and found that the same handful of issues kept emerging (e.g. “I didn’t break the rules”). I then wrote short canned prompts reflecting each theme.

• I consulted with my UX Research partner, who advised having a catch-all “I need support” as the first prompt for users who want to put their issues in their own words.

Entrypoint CTA:

• I brainstormed simple, informative options for the primary entrypoint, eventually settling on “Chat with AI about your options.” 

• Engineering later expanded the audience to include users without access issues too. “Chat with AI about your options” no longer made sense, so I used internal AI tools to run an A/B test to see if “Chat with AI about your access” would perform better. The result was a stat-sig increase in CTR for the test content.

• We exceeded the team’s goals for resolution, sentiment, and engagement. 

• Our work was featured as Facebook Foundation Win of the Fortnight. 

• We presented to leadership about our innovative Engineering/Design collaboration, which influenced the product roadmap.

• The support assistants formed the basis of today’s Meta AI support assistant.

• A personally gratifying moment was when my friend casually mentioned to me that she used the AI support assistant for her Marketplace issue and found it helpful, not realizing I worked on it