Generative AI Platform

Copilot turns complex analytical work into a conversation. A user picks a specialized skill, such as DocInsight, points it at their documents or Moody's datasets, and asks.

A research report that once took weeks now takes minutes. Trust meant the accuracy and readability of the reports Copilot generates, so we designed in reliability with fact checking and source citations. The Copilot Kit was already underway, and the trust and user-centered patterns we discovered, observed, and recorded went into it, so other teams could reuse them.

Copilot was used for real work, and a conversational workflow needs memory: people had to be able to pick up where they left off. Chat History was a must-have. Each person's conversations are stored for them, the many AI skills are consolidated into one continuous chat stream, and people can filter by skill or by time to find the one they need. It became part of why 80% of the company adopted Copilot, and it lifted internal productivity and user satisfaction.

As skills multiplied, the prompt box became the bottleneck: too much capability in too little interface. Behind it, language models can draw on a person's included datasets in different ways, known as RAG and GraphRAG, and each suits different tasks. We did not ask people to understand the difference.

We designed the box around the task they were trying to get done: when a skill needs a file or a dataset, it says so and walks them through attaching a PDF or choosing datasets.