Autonomous Invoice Intake Agent with Copilot Studio and Computer Use

Build a production-ready invoice intake agent that ingests PDF invoices, extracts structured data, updates legacy billing systems through Computer Use, and posts Microsoft Teams notifications, all inside Copilot Studio.

Autonomous intake agents benefit from persistent memory of prior invoice patterns and edge cases. memnode is the MCP-shaped memory layer for this workflow.

Architecture at a glance

• Ingress: chat uploads (single PDF) and Outlook inbox trigger for emails with attachments.
• Processing: AI Builder prompt extracts all invoice fields into one structured payload; Copilot Studio orchestrates.
• Actions: Computer Use fills the legacy invoicing UI; Teams tool posts an adaptive card.
• Knowledge: internal invoice policy in Dataverse with a clear description to enforce compliance.
• Models: GPT-4o primary with automatic GPT-5 routing for deep reasoning and multimodal parsing.

Build steps

1. Create the agent (Copilot Full), name it “Invoice Autonomous Intake Agent,” and expand instructions after creation.
2. Author responsibilities: create invoices, retrieve status, flag unpaid, convert currencies to base, and notify stakeholders. Disable web search; enable file uploads.
3. Seed knowledge: upload the invoice policy, describe when to use it so the agent treats it as authoritative.
4. Design notifications: paste a proven adaptive card schema from the Adaptive Card Designer and reference it in instructions.
5. Build the email trigger: monitor an O365 inbox for attachments, save them to OneDrive, and loop per attachment.
6. Add an AI Builder prompt: multimodal GPT-5 reads the PDF, extracts fields, and emits a single structured payload (test it first).
7. Wire tools:
   • Teams postcard: mostly unconfigured so the agent fills recipients/content dynamically using the adaptive card example.
   • Computer Use: short instructions, open the invoicing system, create a new invoice with the payload, handle popups, and save.
8. Create a chat Topic for uploads: trigger on “upload/attach/submit invoice,” enforce PDF-only and single-attachment rules, and surface the AI Builder output as a variable (case-sensitive tweaks in the code editor).
9. Publish and test: upload a sample PDF, watch Computer Use screenshots as it logs in, clicks “Add Invoice,” fills fields, saves, and posts a Teams adaptive card confirming success.

Reliability tactics

• Provide the adaptive card JSON example to minimize hallucination and guarantee schema compliance.
• Keep Computer Use instructions short; the model handles navigation and popups.
• Use a personal inbox in demos to avoid shared-mailbox attachment quirks; swap to shared later.
• Loop attachments so multiple invoices per email are processed independently.
• Treat the policy document as authoritative for edge cases (currency conversion, approval thresholds).

QA checklist

• Trigger fires only when attachments exist; non-PDFs are rejected in the chat Topic.
• AI Builder output is captured in a single variable with consistent casing.
• Computer Use run completes without timeouts; popups are handled.
• Teams adaptive card shows invoice id, total, vendor, due date, and status.
• Knowledge source is referenced in responses (policy-backed language visible).
• GPT-5 routing kicks in for complex or multimodal steps.

Next optimizations

• Move to a shared mailbox with role-based access and auditing in Dataverse.
• Add duplicate-invoice detection before writing to the legacy system.
• Introduce real-time FX conversion instead of static tables.
• Record Computer Use runs and AI Builder outputs for traceability and SOC2 evidence.