FEI HUANG
I build the infrastructure that designers and engineers
rely on.
PRODUCT DESIGNER I SYSTEMS • ENTERPRISE WORKFLOWS
From design infrastructure to AI interaction architecture
PREVIOUSLY BUILT SYSTEMS AT
FEI HUANG
Collaborators
Platform Engineering and Advantage+ product teams
My focus
Defined the design system's first shared model for AI behavior — moving from independently designed AI interactions to a reusable system governing how AI proposes, communicates, and returns control across Meta's business products.
Designed the human control ladder as the core interaction model, defining five levels of AI involvement and user commitment — from suggestion through confirmation, editing, override, and recovery. The system gave design and engineering teams a shared language for AI behavior: a single model reusable across surfaces, without reinventing interaction logic each time.
Overview
In early 2025, as Advantage+ shifted from setup-driven flows to goal-based AI execution, the interaction model had not kept pace. AI was making decisions users had not explicitly configured, but the design system lacked a shared framework for how AI should behave, communicate those decisions, or return control when needed. At the time, AI in product was still operating as a co-pilot, suggesting, not acting, which defined both what the system needed to solve and what it could reasonably govern.
The work moved beyond component design into system behavior, establishing the shared interaction model that made AI behavior more consistent, predictable, and governable across Meta's business products. The questions this work answered in 2025 have only become more urgent as AI systems have grown more capable and autonomous since.
Challenges
The shift from setup-driven to AI-driven workflows exposed a structural gap. Product teams were building AI interactions independently — each surface inventing its own model for how AI proposes decisions, how users confirm them, and what happens when AI gets it wrong. The result was fragmented trust signals, inconsistent control patterns, and no shared recovery model when AI made the wrong call at the campaign scale.
The problem was not UI. The design system already had mature components. What it lacked was a shared model for AI behavior: when AI acts, how it communicates, and where user control begins and ends. Without that model, every new AI feature required reinventing the same interaction logic from scratch.
Diagnosis & Strategy
The core reframe was architectural, not visual. The previous approach treated AI interactions as isolated UI patterns designed for individual product contexts. What the system needed was a higher level of abstraction: a shared model that defines how AI behaves across surfaces, not how a single AI feature appears on a single screen.
The strategy was to define the full interaction arc as a reusable system: from AI proposing a decision through confirmation, editing, override, and recovery. Rather than designing more components, the work focused on establishing the behavioral structure any AI interaction could operate within.
System Evolution
The centerpiece was the human control ladder, with five levels that defined different degrees of AI involvement, user commitment, and recovery responsibility across the interaction lifecycle. Suggest, Confirm, Edit, Override, and Recover established how AI proposes actions, how users intervene, and how control returns when AI gets it wrong.
​​
- Human Control Ladder-
The architectural move was to define shared rules for AI behavior rather than isolated UI patterns. A UI pattern specifies how something looks. A shared behavior rule defines how AI behaves, how users respond, and what remains under user control, regardless of the product surface or workflow context.
​
Six shared behavior specifications structured the interaction model across the full lifecycle: system states, trust signals, human control, platform guardrails, loading states, and autopilot mode.​
- AI Control Architecture -
The 2025 model also surfaced its own limitations. Control assumed the user was always present, trust signals remained binary while AI confidence was probabilistic, and recovery was not yet designed for users returning to find decisions already made on their behalf. Naming those boundaries explicitly made the architecture more credible — defining what the system could support rather than overstating AI capability.
​
The shared behavior model was defined within the design system and aligned with the Advantage+ product team — establishing a consistent framework for how AI interactions should be built across surfaces before each team developed independent solutions. The architecture gave design and engineering a shared reference for trust signals, control points, system states, and autopilot behavior: one model covering the full interaction arc rather than per-surface reinvention.
​
Engineering constraints directly shaped several design decisions. Real-time confidence scores proved unreliable because model output fluctuated per inference call, creating false precision. Outcome framing became the more honest interaction model, communicating direction rather than unstable percentages.
- AI System States -
Impact
The work established the design system's first shared model for AI behavior — defined and aligned with the Advantage+ product team before independent per-surface solutions could fragment the interaction model. Where no shared framework had existed, teams now had a consistent reference for how AI should propose, communicate, and return control across Meta's business products.
Within the design system team
-
Defined the first shared behavior model for AI interaction — giving design and engineering a consistent reference for system states, trust signals, control, and recovery
-
Aligned design and engineering on the core concepts governing AI behavior: when AI acts, how it communicates, and where user control begins and ends
-
Engineering constraints shaped more honest design decisions — outcome framing replacing unstable confidence scores, constraint clarifying the design rather than compromising it
Near-term across product teams
-
Reduced the risk of fragmented AI interactions as more teams began building Advantage+ features — one model covering the full interaction arc rather than per-surface reinvention
-
Established a shared reference for how AI interactions should be structured before independent solutions could fragment the interaction model
Foundational impact
-
Framework established the architecture for governing AI behavior as AI scales from co-pilot suggestion toward goal-based execution
-
Architecture transferable beyond Meta, any product where users delegate decisions to AI requires explicit models governing how AI proposes, acts, and returns control