INTRO
Copilot for Merchants is an AI-driven interface for a B2B agentic commerce platform, designed for a 2027 vision of e-commerce operations. Rather than replacing human judgment, the copilot surfaces the right decision at the right moment — flagging what needs attention and letting merchants act, delegate or override.
ROLE
UX/UI Design (University Collaboration with a company)
TOOLS
Figma, FigmaJam
PROCESS
Research: mapped two core use cases where AI-assisted decisions create the most operational leverage — catalog content quality and revenue anomalies
Defined two personas per use case (Ellen, E-Commerce Manager; Caroline, Content Editor) to separate strategic oversight from execution-level review
Built a Detect → Decide → Act → Learn framework to structure how the copilot surfaces, resolves and learns from each flow
Designed user flows resolving key logic ambiguities: persona handoff, confidence-tier routing, reject behavior, post-pause branching
Iterated wireframes across a prioritized punch list, with Caroline's four-action decision UI as the highest-stakes screen to get right
DESIGN DECISIONS
Confidence-tier routing — Not every AI suggestion deserves the same scrutiny. High-confidence catalog fixes auto-apply; low-confidence ones route to a human for review.
Four-action decision pattern — Caroline's review screen always offers the same four moves (e.g., Approve / Edit / Reject / Pause), so the copilot's logic stays predictable across dozens of content types.
Persona handoff clarity — When a flagged item moves from Ellen's strategic view to Caroline's execution queue, the interface names why it moved — trust in an agentic system depends on visible reasoning, not just output.
Detect–Decide–Act–Learn as UI skeleton — The framework isn't just internal logic; it's legible in the interface itself, so merchants understand what stage of the loop they're looking at.
REFLECTION
This project pushed me to design trust into an interface, not just usability. When an AI system makes recommendations, the real UX challenge isn't the individual screen — it's making the reasoning behind each suggestion visible enough that a human is comfortable acting on it.
