Less work, better adswith generative AI.
GenAI adoption across launched surfaces
Meta's messaging ads span Messenger, Instagram, and WhatsApp — driving $18B in messaging ads revenue. The ad creation model hadn't kept pace; advertisers were still struggling to build messaging experiences manually. I redesigned the creation flow so advertisers are responsible for content approval, not content generation.
Role
Lead designer, project owner
Team
4 product designers, 3 engineers, 1 data scientist, 0 PMs
Surfaces
Ads Manager, Messenger, WhatsApp, Instagram
The Thesis
The most unclear part was the goal of getting people to chat with me on Messenger and how this set up enables that. — USABILITY TEST PARTICIPANT, BRAZIL
Before redesign
Usability score
64Task completion
61%The interface where advertisers configure their messaging ads was measurably failing — but getting it fixed meant aligning four groups with fundamentally different priorities:
- A design program that prioritizes usability
- An Ads Manager platform that values opinionated defaults over flat architecture
- Messaging Ads teams that only resource what drives revenue impact
- A company-wide AI initiative that needed a foundation to ship on
Each had a reason to care about this surface, but none were aligned on what “fixing it” meant. My thesis was simple: the creation model was the problem, not the interface. If advertisers stopped writing content and started reviewing what AI generated for them, usability, ad performance, and AI adoption would all improve at once.
I linked the competing priorities into one program: the usability failure justified a North Star redesign, the North Star set the stage for generative AI, and participation in executive-led initiatives gave the whole thing visibility and resourcing from six teams.
After
85
Passing usability score
100%
Task completion on retest
96%
GenAI adoption rate
+74%
Conversations started with GenAI
The Interaction Flip
Manual Creation
The Gap
Ad set up treated every business the same in a flat information architecture. Default prompts like “What services do you offer?” ran regardless of what the advertiser sold or did. Almost 90% never changed any content.
All chat templates were given equal weight with no opinionated guidance, even though we knew what performed best. Advertisers were left to make their best guess or rely on outside opinions.
AI-Assisted Review
The System
The redesigned system uses the ad caption and business' page content as inputs to Llama4, which generates conversation starters relevant to that specific business. The advertiser's new job is to confirm the content is relevant, not to write it.
The new architecture defaults to the highest-performing template and regenerates content whenever the ad creative changes — no manual upkeep.
Design Decisions
Redefine the problem
The UPB study I inherited had already failed an expensive round of user testing. The task — “make it easier for advertisers to discover and edit icebreakers” — was measuring how well someone could navigate what had become the org chart rendered as UI. No interface reskin was going to pass that benchmark.
I made the case to redefine the task, a difficult argument when a prior round of user testing already spent its budget. I proposed incorporating the GenAI work my team was already building. This meant admitting the original task was outdated thinking. We didn't need to make things easier to edit, we needed to make editing and template creation itself entirely optional for improved messaging performance.
The visibility tradeoff
The North Star redesign simplified the primary flow by collapsing automated chat and partner templates behind a button. The teams that owned those templates pushed back. Their goal metrics were directly tied to surface visibility, and we were about to bury them.
I designed system-level defaulting that accounted for 3P tooling advertisers specifically: if an advertiser's integration relied on those templates, the system surfaced them. Everyone else got the simplified flow. Both models coexist without degrading either team's numbers.
Designing for partial AI coverage
Llama4 language support rolls out market by market, but the ad creation flow ships globally. The new Chat Builder card needed to serve advertisers with and without AI coverage.
I designed the system so AI-generated content is the default where available, and the card gracefully falls back to a simplified manual flow in other countries — same architecture, same UI patterns, no visible seam between the two experiences unless you switch languages. Without this, the choice was maintaining the legacy codebase and keeping the poorer information architecture.
What shipped
[ 01:13_Ads Manager | Engagement | Messenger Designation_GenAI ]
GenAI automations outperformed the previous defaults by 74%: a 23.7% CTR vs 13.6%, generated entirely from ad and page content.
More importantly, the competing priorities that blocked this work are no longer in tension. Usability, GenAI adoption, and monetization all ship on the same foundation now — new features like Automated Responses and Greeting both launched without another alignment cycle or redesign.
Language expansion extends the runway further. Every language added to Llama4 opens another market on the architecture we built. Thai already launched at ~$10.9M/year in incremental revenue, with nine more languages in flight.