Automate month-end variance commentary

Step 1: Connect actuals and budget in one placePower BI

Build or confirm a Power BI data model that pulls both the actuals feed and the current approved budget into related tables (not two separate flattened exports) so the agent is always comparing against the live, correct period.

A commentary draft built on last quarter's budget snapshot is worse than no draft at all, because it reads as authoritative.

Step 2: Set the variance threshold and let the agent draftCopilot Studio

Define your materiality threshold explicitly (a percentage, a dollar amount, or both) in the agent's instructions, then have it compare actuals to budget by account using a DAX measure for the variance calculation rather than a static calculated column, so it recalculates correctly as new actuals land.

Prompt idea:

Compare actuals to budget by account for this period. For any variance over [your threshold]% or $[your threshold], draft a one-paragraph explanation of the likely driver based on the account history, and flag it for controllership review before it's finalized.

Step 3: Controllership reviews the exceptionsExcel

Export the flagged accounts with their drafted commentary into a review workbook and have controllership mark each one approved, edited, or rejected before anything moves into the close package.

Nobody signs off on drafted narrative blind, and a reviewer who knows a one-time event (a write-off, a timing shift) can catch a plausible-sounding but wrong explanation the agent couldn't know about.

Step 4: Track cycle timeExcel

Log the close date and total commentary-drafting hours each month in a simple running table, and compare the trend before and after this process went live — that trend line, not a subjective sense that things feel faster, is what justifies keeping the agent running or expanding it to more account groups.

Check the work

• Confirm the budget version used is the current approved one, not a prior draft.
• Read the drafted explanation against what you actually know changed that month before signing off.
• Track whether cycle time actually improved, not just whether the commentary reads well.

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