Stand up a self-service agent for routine finance questions

Step 1: Collect the questions that actually repeatSharePoint

Export three months of finance team tickets or inbox threads into a spreadsheet and sort by subject line or category to find the handful of questions driving most of the volume.

Build the agent around that short list first (expense policy limits, reimbursement timing, PO thresholds), not an attempt to anticipate every question anyone could ask.

Step 2: Ground the agent in approved policy documentsCopilot Studio

In Copilot Studio's Knowledge settings, connect the agent only to the specific, currently-approved policy and procedure documents — not a shared drive with old drafts mixed in.

An agent that can see a superseded policy version will cite it with just as much confidence as the current one, and employees have no way to tell the difference.

Step 3: Define what it escalates instead of answeringCopilot Studio

Write explicit escalation rules into the agent's instructions — any question involving a named dollar threshold, a documented exception, or missing paperwork routes to a person rather than getting a best-guess answer.

Test this by asking it a question you know should escalate; if it answers instead of handing off, tighten the rule before anyone else uses it.

Step 4: Pilot it with a small group before rolling outCopilot Studio

Enable it for one team only, and keep a log of every question it answered incorrectly or should have escalated but didn't, for a few weeks.

Fix those specific gaps in its grounding or escalation rules before extending access company-wide — a mistake that reaches one team is a bug fix, the same mistake at full rollout is a trust problem.

Check the work

• The agent's grounding documents are the current, approved versions.
• Escalation rules correctly catch edge cases instead of letting the agent guess.
• A sample of real pilot-period answers were checked against actual policy.

Source: Microsoft Copilot Scenario Library — Automate financial query management (2026)