Copilot Studio: agents you architect, not buy.

Build/buy decisionBuild a custom agent when: • Your workflow uses systems Microsoft's named agents don't touch (line-of-business apps, niche SaaS) • Your process has unusual logic Microsoft doesn't anticipate (industry-specific compliance, multi-step approvals) • You need to embed agent capability in a specific product or experience • Your customer-facing requirements need branding or specific UX Use a named agent (Researcher, Analyst, etc.) when: • The use case is general (research, analysis, meeting notes, sales prep) • Microsoft already covers 80% of what you need • The remaining 20% can be solved via configuration or prompting • You're not ready to maintain a custom agent over time

Agent structureAnatomy of a custom agent: 1. **Trigger** — what starts the conversation? - User invokes by name (chat with 'expense agent') - Inline trigger (typing 'submit expense' anywhere in Copilot) - Programmatic (another system calls it) 2. **Topics** — the conversation flows it handles - Each topic is a flow: triggers, dialog nodes, branching logic - Example: 'Submit an expense' topic handles the whole flow from category to amount to receipt upload 3. **Knowledge sources** — what it knows - SharePoint sites, OneDrive folders, public websites, custom docs - The agent searches these when answering related questions 4. **Actions** — what it can do - Power Automate flows (your existing workflows) - Custom connectors to your APIs - Native Microsoft actions (create a Teams message, file a SharePoint item, etc.)

Knowledge curationKnowledge source best practices: • **Curate, don't dump** — Don't point the agent at 'all of SharePoint.' Point it at 5-15 specific, well-maintained pages or documents. • **One source of truth per topic** — If your expense policy exists in 3 places, the agent will get confused. Designate one as canonical, redirect the others. • **Last-updated dates matter** — agents preferring recent sources is built-in. Documents 2+ years old get used less reliably. • **Structure your docs** — Headings, bullet lists, clear sections. The agent finds info faster. • **Anti-patterns to flag** — outdated wiki pages, drafts marked 'final v2,' SharePoint sites with permissions you can't audit.

Actions designCommon action patterns: • **Lookup actions** (read-only, no confirmation needed) - Look up an employee's PTO balance - Check ticket status in ServiceNow - Find a customer's contract end date in Dynamics • **Create-with-confirmation actions** (always confirm before executing) - File an expense report - Submit a PTO request - Open a support ticket - Send a meeting request • **Notify actions** (low-risk) - Post to a Teams channel - Send a calendar invite - Add a SharePoint item • **Avoid auto-actions** (require human-in-the-loop) - Approving anything financial - Sending external emails - Modifying production data - Any irreversible action

RolloutRollout sequence: 1. **Internal test** — invite 3-5 trusted users to test for 1 week - Have them try to break it: ambiguous prompts, edge cases, off-topic queries - Track every failure or weird interaction 2. **Iteration** — fix the top 10 issues from internal test - Add knowledge sources for what was missing - Refine topic flows for confused conversations - Strengthen guardrails for the off-topic prompts 3. **Pilot** — release to a small department (10-30 users) - Track usage metrics: invocations, completed flows, fallback to human - Survey users at end of pilot - Iterate one more time 4. **General release** — publish to the broader org - Include training material: what the agent does, what it doesn't - Set up feedback channel for bug reports - Monitor monthly metrics; iterate quarterly