Deploy a customer-support response-drafting agent

Step 1: Ground the agent in your actual KB, not general knowledgeCopilot Studio

In Copilot Studio, scope the agent's sources to your maintained product knowledge base and, where permitted, the customer's account and ticket history — never let it draft a response based on general assumptions about your product.

Step 2: Draft, don't auto-sendCopilot Studio

Configure the agent's output as a draft response with attached KB article links, routed to a human support agent's queue for review before anything goes to the customer.

Prompt idea:

Draft a response to this customer's ticket about their export failing with a "quota exceeded" error. Reference their current plan tier from the account record, cite the relevant KB article, and suggest the next step to resolve it.

Step 3: Set clear routing rules for specialist handoffOutlook

Define what the agent should never attempt to draft — billing disputes, anything involving account cancellation, or a ticket where the KB doesn't cover the described issue — and route those straight to a specialist with the ticket context attached.

Step 4: Review draft quality on a sample before wider rolloutOutlook

Pull 15-20 real drafted responses across a range of ticket types and have a support lead rate each against what they'd have written themselves.

Look specifically for topics where the draft is consistently weak, not just an overall pass/fail impression, and fix the KB or agent instructions for that topic before expanding to the full queue.

Check the work

• Confirm every cited KB article actually supports the claim made in the draft.
• Check that account-specific details (plan tier, past issues) in the draft match the real account record.
• Track how often support agents edit vs. send drafts as-is — heavy editing on a topic signals the KB or the agent's prompt needs work.

Source: EPC Group, "Microsoft Copilot Agents 2026: 9 Patterns That Actually Work in Production" (2026)