01

Context

Finance admins at Fyle faced two persistent friction points:

1. Finding answers in support documentation meant navigating help centers and hoping they were looking in the right place

2. Understanding expense analytics required building custom reports or asking support for simple insights like "What's our average expense this month?"

Both felt unnecessarily complex for questions that should have simple answers.

The hypothesis: What if admins could just ask?

02

The Approach: Ship Fast, Learn Faster

Rather than building a perfect conversational interface from scratch, we took a lean approach:

✅ Design a familiar pattern — We modeled the interface after ChatGPT, a pattern users already understood
✅ Focus on capability, not polish — Validate whether users want conversational AI before investing in custom UI
✅ Beta with select users — Launch quickly, gather real usage data, iterate based on behavior

What Copilot Could Do:

• Answer support documentation queries: "How do I set up approval workflows?"

• Provide expense analytics insights: "What was the average expense value this month?"

• Surface relevant information without leaving the workflow

03

Design Decisions

Why we mimicked ChatGPT:

The GPT interface has become the de facto standard for conversational AI. By using familiar patterns, we:

• Reduced learning curve to near-zero

• Focused engineering effort on AI capability, not custom UI

• Shipped in weeks, not months

What we intentionally kept simple:

• Clean chat interface with minimal chrome

• Basic markdown support for structured responses

• Simple suggested prompts to guide discovery

• Clear boundaries: "I can help with analytics and support questions"

What we learned to improve later:

• Contextual triggering based on user behavior

• Richer data visualization for analytics responses

• Tighter integration with the Fyle interface

• Progressive disclosure for complex queries

View Prototype