Build a collections agent

Step 1: Define the prioritization signalExcel

In your AR aging export — converted to an Excel Table so the columns stay named and addressable — decide which fields actually feed the priority score: balance size, days past due, payment-history reliability, and any prior broken payment promises.

Write the weighting down explicitly (even a rough "days past due matters twice as much as balance size") and hand that rubric to the agent rather than letting it infer its own scoring logic from the raw columns.

Step 2: Build the prioritization and drafting agentCopilot Studio

In Copilot Studio, connect the agent to your AR aging data and have it score and rank accounts, then draft an outreach message per account matched to that account's history (a first reminder reads differently than a fourth).

Prompt idea:

Rank these 40 past-due accounts by collection priority using balance size, days past due, and payment history. For the top 10, draft a short outreach email appropriate to how many times we've already contacted them.

Step 3: Route drafts to collectors, not customers directlyOutlook

Save each drafted message as an Outlook draft assigned to the owning collector, or post them into a shared review queue — never configure the flow to send outreach directly.

The agent prioritizes and drafts; a person reads the specific account's history and decides whether the tone is right before anything reaches a customer's inbox.

Step 4: Track outcomes back into the priority modelExcel

Add an Outcome column to the Table (paid, promised, no response, disputed) and update it after each outreach round, then periodically check whether accounts with similar scores are actually converting at similar rates.

If a particular signal (say, a specific payment-history pattern) isn't predicting outcomes the way the rubric assumed, that's your signal to adjust the weighting in step 1, not just the agent's prompt.

Check the work

• Verify the aging and payment-history numbers behind the top-ranked accounts against the AR system directly.
• Confirm drafted messages don't reference a payment promise or prior conversation that didn't actually happen.
• Review the bottom of the ranked list occasionally — make sure smaller or newer past-due accounts aren't being silently ignored.

Source: Microsoft Copilot Scenario Library — Finance (2026)