The "Sleeping" Agent That Solves Your AP Nightmares: Zero-Touch Invoice Reconciliation
In the world of manufacturing and supply chain, the physical goods often move faster than the paperwork.
You have trucks arriving at the dock, but the payment to the vendor is stalled because of the "Three-Way Match" problem.
The Purchase Order (PO) in your ERP says one thing (e.g., "100 units at $5.00").
The Receiving Log says another (e.g., "Received 98 units").
The Vendor Invoice is a PDF attachment in an email that says something else entirely (e.g., "100 units at $5.50 + Freight").
Bridging these gaps usually falls to a human Accounts Payable (AP) clerk. They spend their days "swivel-chairing" between Outlook, a PDF viewer, and SAP/Oracle, manually keying in data and chasing down approvals for the $50 discrepancy. It is tedious, morale-crushing work that delays vendor payments and hurts credit terms.
At Evonence, we are solving this with a new kind of worker: the Asynchronous Reconciliation Agent.
The Solution: An Agent That "Sleeps" Until It Is Needed
Unlike a chatbot that waits for a user to type, this agent is event-driven. It lives in the background, completely dormant (and costing you nearly nothing) until a specific trigger occurs—like a new invoice arriving in a Cloud Storage bucket or an email inbox.
The Workflow:
Trigger: A vendor emails an invoice. The file lands in Google Cloud Storage.
Wake Up: Eventarc detects the file and "wakes" the agent.
Think & Extract: The agent extracts the data (Invoice #, Line Items, Total).
Reconcile: The agent queries BigQuery or your ERP to find the corresponding PO. It compares the two.
Match? It posts the payment to the ERP automatically.
Mismatch? It applies logic: "Is the variance <$10? If yes, auto-approve. If no, draft an email to the vendor asking for correction."
Sleep: The agent spins down.
Under the Hood: Eventarc + Vertex AI
This architecture is a breakthrough because it is serverless and stateful.
1. The Trigger: Eventarc
Traditional automation often requires "polling"—a server constantly checking "Is there a file yet? Is there a file yet?" This wastes compute and money. Eventarc is standardized event delivery. It acts like a doorbell. The agent doesn't run until the doorbell rings.
2. The Brain: Vertex AI Agent Engine
Most OCR tools just dump text. Our agent uses Vertex AI to reason about the discrepancy.
Scenario: The invoice lists "Part A-123" but your PO lists "Part A-123-RevB".
The AI: Recognizes these are likely the same part based on description and price history, links them, and notes the alias for next time. A rigid script would have just failed.
The Migration Advantage: Why this Beats AWS Lambda Scripts
Many supply chain teams have tried to hack this together using AWS Lambda scripts and Textract. The problem? Fragility.
If a vendor changes their invoice layout, the Textract template breaks. If the logic gets complex (e.g., partial shipments), the Lambda script becomes unmanageable "spaghetti code."
Migrating to a Google Cloud Agentic approach moves the logic from code (hard to change) to natural language prompts (easy to change).
Old Way: Rewrite Python code to handle "Freight" charges.
New Way: Update the Agent's prompt: "If you see a line item for freight, code it to GL Account 5000 instead of rejecting the mismatch."
The Evonence Approach: Fix the Cash Flow in 8 Weeks
We call this the "Invoice-to-Cash" Accelerator.
Weeks 1-2: Map your ingestion points (Email/FTP) and ERP integration.
Weeks 3-4: Configure the Eventarc triggers and Agent reasoning prompts.
Weeks 5-6: "Shadow Mode" – The agent reviews invoices but only suggests actions to humans.
Weeks 7-8: Go Live with auto-approval thresholds.
The Goal: Let your AP team focus on vendor relationships, not data entry.
An agent that sleeps until work arrives, fixes the errors, and goes back to bed. Contact Evonence to automate your AP workflows.
