How to Scope and Create Custom GPTs

• Focus on a single, narrow use case (like drafting proposals or answering FAQs) for your custom agent. 

• Be rigorous about data. Only upload essential, non-sensitive documents. 

• Commit to maintenance. Treat your agent as a "living product." 

Join Maya Kuppermann and Ruthwick Pathireddy for a practical deep dive into creating and scoping custom AI agents (or GPTs). They define these as highly-tailored AI assistants designed for a specific role or workflow, which operate using a contained, curated set of data and unique instructions, making them more focused and reliable than general chatbots. The session walks attendees through the essential steps for development, emphasizing the importance of defining the agent’s purpose, curating its specific knowledge base, and using an iterative approach that involves rigorous testing and refinement.

The presenters illustrate the concepts with tactical examples, including a custom agent built in Microsoft Copilot for adhering to an AI policy and an internal support GPT created with ChatGPT for managing frequently asked questions. A core takeaway is that these agents are most valuable for repetitive, manual tasks or when internal documentation is unclear, aiming for approximately 80% accuracy with a crucial "human in the loop" to handle the final necessary refinements. They also address real-world challenges, such as dealing with disorganized data sources and the need for continuous maintenance to prevent the agent from using outdated or incorrect information

​