Designing & Training a Customer Support AI
Timeline: Q2 2023 - Q2 2024
Role: UX Writer / Instructional Designer
I contributed to the design and training of an AI-powered customer support assistant used by both customers and internal support agents. The goal was to ensure the AI delivered accurate, empathetic, and policy-aligned responses while supporting agents with consistent, trustworthy information.
**Due to confidentiality and legal constraints, this case study focuses on outcomes, decision-making, and final deliverables rather than early-stage drafts or exploratory artifacts.
The Problem
Customer support teams were spending a significant amount of time handling high-volume, repetitive inquiries related to orders, returns, and seller questions, resulting in longer wait times and less attention for complex or sensitive cases.
Existing documentation was not always utilized as intended, and inconsistent responses increased the risk of confusion or escalations.
The challenge was to train an AI system that could respond accurately to customers while also serving as a reliable tool for support agents while maintaining Grailed's brand voice.
Constraints & Considerations
This project required careful consideration of several constraints:
-
Responses needed to remain fully aligned with platform policies
-
The AI had to handle edge cases and ambiguity gracefully
-
Incorrect or overly confident answers posed trust and safety risks
-
Internal tooling needed to support agents without replacing human judgment
-
Training data and early iterations were reduced to small test groups
As a result, this case study focuses on content strategy, decision logic, and outcomes rather than proprietary training data.
Goals & Success Criteria
The primary goals of this project were to:
-
Provide customers with clear, empathetic, and accurate responses to their most asked questions
-
Reduce repetitive support inquiries through consistent information delivery
-
Support agents with fast, reliable guidance during active cases
-
Establish content rules that scaled across new questions and scenarios
Success was evaluated through internal QA review, agent feedback, and most importantly the AI’s ability to handle common and edge-case inquiries without escalation.
Process & Content Strategy
I approached AI training as a content design and instructional problem rather than a technical one. My process included:
-
Analyzing frequent customer inquiries and failure points
-
Defining response principles for tone, confidence, and escalation
-
Writing and refining example responses to model desired behavior
-
Identifying scenarios where the AI should defer to a human agent
-
Iterating on responses based on observed gaps or ambiguity
Special attention was given to ensuring accuracy of information and successful adoption of brand voice.
Results
Responses designed with human empathy in mind while maintaining accuracy
Language intentionally avoids definitive claims in edge cases
Clear handoff to human agent when presented with issues outside of their domain
This project highlighted the importance of treating AI responses as a user experience that requires the same care, structure, and ethical consideration as human-written content. Thoughtful training helped reduce ambiguity, supported internal teams, and reinforced trust with customers.
If expanding this system further, I would explore deeper collaboration with agents to continuously refine response logic and identify emerging patterns in customer needs.