Elise Zur • Product Manager & HCI Designer

Back

ClearPath

An agentic AI concept that helps airline customer support teams automate routine requests while escalating complex decisions to human agents.

My role

Product Strategy & Design

Timeline

August to December 2024

My team

Caroline Shi
Glenda Tan
Janice Lu
Peace Im
Priscilla Park

🏆

Most Likely to Invest

Voted most investable concept in CMU's Design of AI Products course.

The Brief

How might agentic AI reduce repetitive workplace tasks without removing human judgment where it matters?

Our assignment was to identify a workplace problem where an AI agent could independently complete tasks on behalf of employees. Rather than starting with a specific industry, we began by exploring where agentic AI could provide meaningful value.

The Solution

SkyAssistAI handles routine airline support requests and brings agents in when a decision requires human judgment.

The system is designed around different levels of autonomy. Straightforward requests can be handled by the AI, while complex or high-risk situations are escalated with the relevant context and possible resolution paths already prepared.

Understand

Interpret the customer's request, booking details, and urgency.

Act

Complete common and repetitive requests when the system has enough confidence to proceed.

Escalate

Flag high-risk or complex situations where human judgment is needed.

Risk & Guardrails

We considered how the system could fail and what safeguards would be needed before giving an AI agent control over customer requests.

Our risk analysis considered issues including incorrect actions, overreliance on AI, accessibility, customer frustration, and the need for human overrides. This informed safeguards such as escalation thresholds, logging, monitoring, and clear paths to human support.

Customer Control

Easy access to human support

Human Oversight

Agents can review and override

System Visibility

Actions are logged and monitored

Takeaways

Agentic AI requires designing more boundaries than I expected.

This project pushed me to think beyond what an AI system could automate and consider where autonomy was appropriate. Giving the system more agency also meant thinking more carefully about escalation, oversight, and failure. We also discussed what this meant for human workers who might lose jobs because of this technology. While this project was speculative, it really opened my eyes to how carefully organizations need to think about bringing AI into their spaces.

The Solution

SkyAssistAI handles routine airline support requests and brings agents in when a decision requires human judgment.

Understand

Interpret the customer's request, booking details, and urgency.

Act

Complete common and repetitive requests when the system has enough confidence to proceed.

Escalate

Flag high-risk or complex situations where human judgment is needed.

Our Process

We started broad, then evaluated and narrowed our ideas based on value, feasibility, and risk.

Each of our six team members generated two potential applications for agentic AI, giving us 12 initial concepts to evaluate. We assessed the ideas, narrowed them to our strongest candidates, and used risk-benefit analysis and feedback to determine where to focus.

12 concepts

Each team member generated their top ideas for agentic AI in the workplace.

Narrowed to 2

We evaluated ideas based on technical feasibility, financial viability, and social desirability, then dot-voted to select two finalists.

Risk-benefit analysis

We mapped the twelve concepts based off high-low risk and high-low benefit.

Refined the direction

Feedback highlighted the opportunity to go beyond diagnosing issues and have the agent take action independently on repetitive tasks.

SkyAssistAI

We applied the concept to airline customer support, where there are high volumes of repetitive requests and clear opportunities for an agentic AI tool to create value.

12 concepts

Each team member generated their top ideas for agentic AI in the workplace.

We evaluated 12 concepts to identify where agentic AI could provide the most value.

We compared our ideas across factors including desirability, technical feasibility, financial viability, and impact. Each team member selected their strongest concepts before we narrowed the group further.

Narrowed to 2

We evaluated ideas based on technical feasibility, financial viability, and social desirability, then dot-voted to select two finalists.

Risk-benefit analysis

We mapped the twelve concepts based off high-low risk and high-low benefit.

Refined the direction

Feedback pushed us to think beyond an AI that diagnoses problems and toward one that could take action.

My initial IT troubleshooting concept ranked highly; however, it focused heavily on diagnosing an issue and recommending a solution. Feedback from our professor encouraged us to explore what the agent could actually do on someone's behalf, particularly for repetitive tasks.

SkyAssistAI

We applied the concept to airline customer support, where there are high volumes of repetitive requests and clear opportunities for an agentic AI tool to create value.

Airline support gave us a clear range of tasks to design around: routine requests such as rebooking or rescheduling could potentially be automated, while more consequential situations could still require a customer support agent.

Takeaways

Agentic AI requires designing more boundaries than I expected.

Takeaways

Agentic AI requires designing more boundaries than I expected.