Each of the 11 organizations my team supported was doing exactly what good product teams do, heads down, focused on improving the surfaces they owned. The Compliance org was investing in making their policy pages clearer. Account Health was iterating on their alert UI. Every team was making rational improvements to the problems they could see from where they sat.

My team had a different vantage point. Because we were embedded across all of them simultaneously, we could step back and see what no individual org could: the same resolution gap playing out everywhere at once. Sellers were being told clearly what was wrong. Across every surface, in every domain, they had nowhere to fix it from there. The individual teams weren't missing anything, they just couldn't see the cross-org pattern the way we could.

In usability sessions and seller roundtables spanning multiple business lines, the signal was consistent: sellers understood the problem. The friction was execution, leaving the surface, navigating somewhere else, figuring out what to change, coming back and hoping they got it right. I brought that synthesis to the org leaders we worked with. Not as a criticism of what they were building, but as a different lens on what was driving the seller struggle they were each trying to solve.

When Seller Assistant launched, using it as a smarter help system was a reasonable first move. Ask a question, get a policy summary, follow the link to the relevant page. It was better than searching help documentation manually, and it fit naturally with how each org was already thinking about the problem. Every team was focused on making their piece of the experience clearer.

But from the cross-org view my team had, a different picture was forming. A seller who got a policy summary still had to figure out which field in their listing was the issue, navigate to the right editor, make the change, and resubmit, hoping they got it right. The assistant answered "what's the rule?" It never answered "what exactly is wrong with yours, and here's how to fix it right now." That gap was where sellers kept falling, and it was showing up in the data across every business line we supported.

That pattern is what I brought to the conversations with org leaders. Not "your approach is wrong", but "we're seeing something from across the business that might reframe the problem you're trying to solve."

The Compliance org was where we went first. They had a well-defined problem, sellers getting listings suppressed for policy violations they didn't know how to fix, and a team willing to explore a new direction with us. I worked with their leadership to scope a hero use case, and directed my team to prove out the concept: instead of explaining the policy, diagnose the listing. Look at the seller's actual content, identify the specific violation, and propose the exact change with a before/after diff.

The trust problem was real and we had to solve it together. Sellers weren't going to act on AI recommendations they couldn't verify, especially for something as consequential as a compliance submission. The answer my team landed on was what we called "showing the receipts": every recommendation links directly to the policy that produced it. Sellers see the evidence before they do anything. That principle, earned through the Compliance work, became the non-negotiable foundation I brought to every other org as the framework expanded.

Stage 2 still required sellers to navigate to the listing editor and apply the fix themselves. But it answered the question every org leader needed answered before they'd take a bigger bet: would sellers actually trust a system that diagnosed their specific problem and showed its reasoning? The Compliance data said yes. That opened the door to Stage 3.

With the trust model proven in Compliance, Stage 3 was about closing the remaining gap, moving the action itself into the conversation. The seller reviews the proposed change, approves it, and it applies inline. No navigation, no context-switching, no coming back to check if it worked. Start to finish, one thread.

The organizational work was as important as the design work. The teams I was partnering with had been building toward the same seller outcomes through a different path, dedicated pages, owned workflows, surfaces they could control and iterate on independently. Asking them to route those workflows through a shared platform was a real ask. It meant trading direct ownership for consistency. My job was to make that tradeoff feel worth it, not by arguing them out of their concerns, but by addressing those concerns in the framework itself, vertical-specific guardrails, use-case-specific logic, clear definitions of what lived in Seller Assistant versus what stayed on a page they owned.

I worked through each org's specific edge cases with their leads. What happened when the AI was wrong? What did the approval flow look like for a high-stakes compliance submission versus a low-stakes listing edit? How did their team stay in control of the experience without owning the surface? Those conversations shaped the framework as much as the design work did. The Compliance launch gave us the proof point. The relationships we had built gave us the credibility to bring the other orgs along.

The same pattern holds across every use case. A seller asking why their sales declined gets a full trend analysis with a recommended ad campaign they can launch with one tap. A seller asking about a compliance document gets shown the exact document needed, the field to upload it to, and can complete the submission inline. The AI doesn't just know what to do, it does it, with the seller in control the whole way.

As more orgs came on board, the design challenge shifted from proving the concept to maintaining coherence across a system that no single team owned. Each org had different data, different risk tolerances, different action types. The natural tendency, and the reasonable one, from each team's perspective, was to optimize for their specific use case. My team's job was to protect the patterns that made the whole thing work.

I worked with my team to define four interaction principles and then held them consistently across every org, every design review, every edge case that came up during build. These weren't imposed top-down, they emerged from the Compliance work and the seller research, and we made the reasoning transparent so each org could see why the principles mattered, not just what they were. The argument that kept resonating: a seller who learns to trust the system in one context will bring that trust to others. Inconsistency doesn't just confuse, it erodes the trust that every org had worked to build.

~90% acceptance is a meaningful number when you consider what it represents: sellers approving an AI agent to modify their listings without independently verifying the change first. That level of trust is earned, not designed in. It came from a consistent, evidence-first pattern that my team defined and the org leaders helped shape, and then held together across every use case we expanded into.

The operational results bore out the original case. Compliance resubmission rates dropped 87% in covered categories. Median resolution time fell from hours to 47 seconds. CS contacts tied to policy misinterpretation dropped measurably. Those were the metrics I had used to build the business case at the start, and bringing them back to the org leaders who had taken a bet on the framework was one of the better moments of the project.

The prototype covers the three core use cases, inventory analysis, sales growth, and policy fix, on desktop and mobile. Each one walks the full flow from question through analysis to inline action.