An AI-powered platform that unified fragmented marketing workflows into intelligent orchestration.
I led the experience design of an AI orchestration layer that unifies Google's fragmented marketing tools into one intelligent system, carrying the strategic thread so teams move from planning to launch without losing context.
Mandala threads planning, audiences, creative, and activation into a single orchestrated workspace, so marketers move from intent to launch without losing context between tools.
Google Marketing lived across hundreds of tools, dashboards, and spreadsheets. Planners spent more time coordinating than deciding. Every new product added another seam, and every campaign meant re-stitching context by hand. The real opportunity wasn't another marketing app, it was rethinking how marketing work itself moves through the org.
Most proposals on the table tried to optimize individual tools. I argued for the opposite. Stop optimizing the parts and start orchestrating the whole. That single reframe is what made Mandala possible, and what made leadership pay attention. AI wouldn't replace the marketer, it would carry the thread between tools so people could stay in the strategic work instead of the plumbing.
Before converging on Mandala, I led a body of exploratory work: heuristic evaluations of the tools teams used every day, archetype activities to ground the work in real roles, FigJam design jams across the org, and sketch-level explorations of what an AI-native workspace could feel like. Each artifact became a vehicle for discussion, helping product, engineering, and leadership see what was possible.
Mandala is less a single product and more a connected set of surfaces. Each one shares the same orchestration model, so a decision in planning ripples into briefs, approvals, audiences, concepts, and assets, without losing context along the way.
Mandala connected Google's fragmented marketing ecosystem into a single intelligent platform. Instead of searching individual tools, AI retrieved the most relevant information across documents, campaigns, analytics, research, and planning systems, assembling the right context before generating a response. Grounded in that knowledge layer, Mandala functioned less like a chatbot and more like an AI thought partner, combining retrieval, reasoning, summarization, and generation to help marketers explore ideas, answer complex questions, uncover insights, and accelerate decisions. My role wasn't designing prompts, it was designing how humans and multiple AI capabilities collaborate to solve real marketing problems.
Mandala was the north-star, but it sat inside a broader body of work. I designed AI capabilities across Google's marketing ecosystem, from audience building to creative workflows to measurement. Each capability followed the same orchestration model: intelligence inside the work, grounded in the team's real material, so adopting one made the next feel familiar.
Theme · 01
Marketing Garage: the connected workspace where briefs, plans, and approvals move together.
Theme · 02
Intelligence for understanding who you're talking to, and how the work is landing in market.
Theme · 03
AI that helps marketers move from idea to asset without leaving the work.
Theme · 04
The longer-horizon concept work that set direction for where the ecosystem could go next.
Every AI proposal in Mandala carries its own defense: the inputs it used, the reasoning in one plain sentence, and the trade-offs it made. Click any input to jump to the source. Reject one and the proposal re-computes in place. V1 shipped a confidence score, "92% match", and marketers ignored it. A number without a story can't be defended upward. Provenance in, reasoning out, trade-offs on the surface became the pattern the rest of the product inherits.
Consumer AI assumes a forgiving user. Enterprise marketing does not. Four failure classes shaped Mandala's interaction model, each with a rule I held the line on and a mechanism that made the rule visible in the product.
Failure · 01
AI never invents a number. It retrieves one, or asks for a query.
Every reach, forecast, spend, or lift figure is retrieved from a system of record and stamped with its source on the surface. If Mandala can't cite it, it doesn't say it.
Failure · 02
Guardrails check the work before it renders, not after.
Concepts, headlines, and assets are validated against the brand model (voice, taxonomy, restricted claims) pre-render. Blocked outputs explain what and why, so marketers trust the guardrails they can't see.
Failure · 03
AI drafts, compares, recommends. Humans commit.
Mandala could go brief-to-launch on its own. It doesn't. Every phase has an intentional human checkpoint, because a fully-automated pipeline is one bad prompt from a brand incident.
Failure · 04
Stream the decision first, the rationale after.
Retrieval across Google's marketing data is slow. Proposals stream progressively: headline decision, then rationale, then sources. Marketers can commit before the rationale finishes if they already trust the call.
When not to use AI
Anything with legal, comms, or regulatory implications routes out of Mandala into the humans who own that risk. AI hands the work off with context; it does not try to be the reviewer. Knowing which decisions to leave alone is as much of the interaction model as the ones AI touches.
Beyond Mandala, the project produced a set of Human-AI creativity principles that other teams in the org continue to build against. The guiding philosophy: foster a collaborative partnership between marketers and GenAI, so AI amplifies expertise rather than replacing it. For AI to truly empower marketers, it must be a trusted partner, not a black box.
The vision earned SVP support the moment leaders could watch the work behave. Two high-fidelity Mandala prototypes, and the broader ecosystem work behind them, turned an abstract argument into something you could point at. Campaign speed climbed 60%, satisfaction moved from 34% to 89%, and leadership greenlit a dedicated team and multi-year budget to deliver the roadmap. More lasting than any single screen: a shared language, orchestration, not optimization, that reframed how the org thinks about marketing tooling.
Four years shipping AI inside one of the most political, most brand-sensitive orgs on earth gave me a stance, not a list of tips. This is the POV I bring into every enterprise-AI conversation now.
The instinct inside big companies is to make each tool smarter. The leverage is in the seams between the tools, not the tools themselves. A workflow that flows is worth more than any single feature that ships. Every enterprise-AI project I take on now, I start by mapping the handoffs, not the surfaces.
Marketers, planners, and analysts don't just use the output, they have to defend it upward. If the AI can't explain itself in one sentence a VP will accept, it doesn't matter how good the answer is; it won't ship. Design the rationale before the answer. Provenance, trade-offs, and sources are not a settings panel, they're the primary surface.
AI can draft, compare, and recommend. Humans commit. In a domain where a bad output can become a Twitter incident by lunchtime, the interaction model has to hold the line on human decision-making at every phase. This isn't caution, it's craft, the constraint is what makes the tool safe enough to actually get used.
The moment intelligence needs a separate tab, it becomes another tool to manage. The most-used AI features I shipped were the ones marketers didn't call AI. Contextual, in-flow, and quiet until asked, that's the shape of enterprise AI that gets adopted instead of demo'd.
Orchestration, not optimization gave eight partner orgs the same vocabulary. That vocabulary reshapes roadmap conversations long after the specific surfaces are redesigned. The most durable staff-level design work is the frame the organization keeps using after you've moved on.
SVP support unlocked the moment leaders could watch the work behave, not read about it. A working artifact stops debate about the concept and starts debate about the details, which is the only debate worth having. This is now the first thing I build on any new enterprise engagement.