Introduction
An agent that answers employees' questions can give a wrong answer with confidence, share information with the wrong person, or take an action that's hard to undo. Responsible AI for an agent comes from a set of design decisions you make about the agent's principles, content moderation, model, and guardrails, before launch and after.
Suppose your organization is building an HR policy agent that answers employees' benefits and leave questions and might later submit leave requests for them. Employment is a sensitive domain, where a wrong answer or an unreviewed action can affect someone's pay or job. The higher the stakes in an agent's domain, the more each of those design decisions matters to the people who rely on it.
In this module, you explore:
- How to apply responsible AI principles to agent design
- How to configure content moderation for an agent
- How to evaluate whether a model is ready for production use
- How to design guardrails that keep an agent within its scope
- How to design transparency practices for an agent's users
Note
This module covers agents on the standard and GitHub Copilot harnesses in Copilot Studio and points out where their settings differ.
By the end of this module, you'll be able to design a responsible AI strategy that covers the principles, content moderation, model selection, guardrails, and transparency of an agent in Copilot Studio.