Start with a task. Then build an agent.

Make the job smaller

“Help our operations team” is an ambition. “Find the status of a purchase order and explain the next action” is a task you can design and test. Start with a single user, a recurring question and the evidence needed to answer it. A useful first agent is deliberately narrow.

Separate knowing from doing

Retrieving a policy and summarising it has a different impact from changing an order. Write down the information the agent can read, the actions it can propose and the actions that need approval. For an initial lab, use sample records and keep business-system changes behind a human review step.

Give it an honest exit

An agent should be able to say that a record is missing or a question is outside its scope. Design the handoff before polishing the greeting. The user needs to know what happened, what was not completed and who can help.

Test the difficult cases

Create a small set of normal requests, ambiguous requests, missing records and unauthorised actions. Compare the answer with the underlying evidence. Record whether the right tool was used, whether the result was correct and whether a reviewer had to repair it. A convincing demonstration is a starting point; repeatable behaviour is the goal.

Microsoft sources and further reading

• Microsoft Learn: Copilot Studio overview
• Microsoft Learn: security and governance