FourKites AI Product
Design Direction
AI/ML
SaaS
Supply Chain
Design Systems
Data Visualzation
Overview
Supply chain operators needed a faster way to investigate shipment disruptions without sacrificing trust in AI-generated recommendations. I led the interaction design for FourKites' first conversational AI experience, defining when AI could act autonomously and when human review was required.
My belief was simple: for high stakes decisions, an AI system earns trust by asking permission, not by acting first and explaining later.
Challenge
FourKites operators were navigating dense, high-stakes shipment data across multiple systems, often under time pressure when something was already going wrong. A purely automated system risked acting on bad data or false confidence. A purely manual one couldn't scale. We needed an interaction model where AI could trigger real actions, like flagging exceptions or surfacing risk, while keeping a human checkpoint for anything an operator needed to trust before it shipped.
I defined the human-in-the-loop interaction model that determined which AI-generated actions could execute automatically and which required operator approval. Working alongside product, engineering, and data science, I established review patterns, confidence thresholds, and conversational workflows that allowed operators to approve, reject, or investigate recommendations without losing context. Rather than designing a generic chatbot, we built an AI experience grounded in the language, workflows, and decision models operators already trusted.
We considered fully automating shipment exception handling, but rejected it because operators would not trust silent AI decisions on high stakes freight. We chose a checkpoint model instead, where the AI recommends and the operator approves, trading some speed for trust we could not afford to lose.
70% weekly adoption after 90 days among operators using the AI workflow.
