Role
Sole Product Designer
Timeline
2 month, Redesign for relaunch
Omni Assist Ai Assistant
Introduced MCP connection where Ai can build audiences, campaigns for the user just with a few prompts.
Problem
I was tasked to redesign Omni Assist, the embedded AI chat in Omnicom's marketing intelligence platform. Users had learned to use Omni Assist as a conversational knowledge base—asking questions about audience suggestions, campaign strategy, and creative best practices. But with MCP integration coming online, Omni Assist could now do more than advise; it could act directly—build segmented audiences, design moodboards, test a campaign idea with user personas without requiring users to leave the chat. The challenge was signaling this capability shift to marketing professionals who had developed a fixed mental model of what the AI could do.
Process
I started with a competitive study of popular AI tools like ChatGPT, Copilot, as well as industry specific AI tools like Jasper to understand how leading AI products signal it's tools. The finding: progressive disclosure—show agency without overwhelming. In user interviews we confirmed users saw Assist as a writing assistant, feels not very smart because it have no knowledge of their campaign work.
I also had to juggle tension with our stakeholders, who wanted feature parity with the latest tools like Canvas, AI Presentation design, app vibe coding. But my research and conversations with users showed the opposite need: they were already overwhelmed with day to day work, additional features are nice to have but AI assisting their existing work brings true value.
Working with my director, I conducted a MoSCoW survey with users and stakeholders. This became my anchor in leadership conversations—we're not chasing ChatGPT's feature count; we're prioritizing designing for agentic ai able to build with simple human guidance.
Solution
UI Redesign
The redesigned interface tackled core problems of navigation. The chat screen became resizable, letting users work alongside the main Omni apps instead of being blocked by the interface. Frequently used functions like the agent store and file attachment received higher affordance to reduce friction.
Canvas
We introduced Canvas to extend what users could do with AI outputs. AI-generated text and data files automatically stored in Omni Assets, letting users transfer content directly to other apps without leaving the platform. Users could also attach their own assets for AI to reference and build on.
MCP Connections
We explored surfacing MCP skills—AI capabilities connected to Omni's popular apps like Audience Explorer and Campaign Workflow. The first iteration was a skills dropdown with toggles, but research revealed the real blocker: users struggled with prompting, not feature discovery. We shifted to smart prompt suggestions and automated skill selection, eliminating toggle-clicking and helping users learn MCP patterns faster.
Product Marketing
Designed feature announcement product marketing content to engage users and increase adoption, including a standalone product update website, Chameleon onboarding explainers, and PowerPoint slides for team meetings.
Result
The full UI revamp and technical architecture of Omnicom’s enterprise LLM increase usage by 10% (40K user globally)within the first month. The SEO MCP tool improved client CTR by 25%.
Reflection
Designing for MCP at launch was genuinely hard—there was no UX playbook for how users would interact with agentic tools in a marketing context. But that constraint forced clarity. I couldn't rely on existing patterns; every design decision had to be grounded in user research, not assumption.The bigger learning was around influence. Working with senior PMs taught me that as a designer, my job wasn't to "win" aesthetic arguments—it was to bring evidence to the table. The MoSCoW survey became my credibility. When PMs pushed for feature parity, I could show: "Here's what users actually need, in priority order." That shifted the conversation from opinion to strategy. Looking back, I'd run interactive prototypes with users earlier—not to prove I was "right," but to surface real confusion points faster. That data would've shortened the back-and-forth with PMs and built team confidence earlier. It's a lesson I carry into every project now: evidence isn't defensive; it's collaborative.