How do you build your first agent?
Created with AI.
A first agent should answer a few questions well rather than many poorly. This guide walks you through the setup and the choices that determine whether customers get the help they need.
- 6 steps
- Total time: 2 h
- Preparation: 1 h
- Result: An agent that answers the five most common questions well
Required tools
- Copilot Studio — With permission to create and publish agents.
- Updated knowledge articles — The agent bases its answers on these. If they are outdated, the answers will be too.
Select scope
1.Find the five most frequently asked questions
Get them from the case statistics rather than from a brainstorming session. The customer service categorization gives you the answer in minutes, and these are the actual questions, not the ones you think will come up.
2.Decide what it should not answer
Write down the topics that should always be referred to a human: complaints, cancellations, and payment issues. This is just as important as what it should be able to handle.
Build and test
3.Create topics for the precise answers
For questions where the answer must be the same every time, create a topic with the exact answer. Free-form responses are for answers that vary.
4.Connect the knowledge source
Direct the agent to the articles intended for the target audience—not to everything you have. A customer-facing agent should not be able to reproduce internal procedures.
5.Set up handover before publishing
The agent must be able to hand off the conversation with the context intact. This must be in place from day one, not added after the first complaints.
6.Try to get it to give an incorrect answer
Ask ambiguous questions, out-of-scope questions, and questions designed to elicit internal information. Better that you find the weaknesses than a customer does.
Discussion
Ask a question or share what helped you.
