AI Invoice Processing with Claude Code and n8n

This workflow watches a dedicated Gmail inbox for invoice email attachments. When a PDF or image invoice arrives, n8n triggers the pipeline and passes the file to Claude Code through the n8n MCP server. Claude extracts structured invoice data including vendor name, invoice number, line items, totals, tax amounts, and due dates using its vision capabilities. The extracted data is written to a QuickBooks invoice draft for review. A Slack message is sent to the finance team with a summary and a QuickBooks link for approval. Once approved via a Slack emoji reaction, the invoice is finalized in QuickBooks and logged to a Google Sheet audit trail. Rejected invoices are flagged in a separate sheet with the reviewer's notes.

BUSINESS PROBLEM

Finance teams at small-to-midsize businesses spend an estimated 12 hours per week manually entering invoice data from email attachments into accounting systems. The AP process involves opening each email, downloading the attachment, reading the data, typing it into QuickBooks or Xero, filing the PDF, and logging the entry. Human data entry error rates average 1-3 percent on invoice processing, leading to payment delays, duplicate payments, and reconciliation headaches. [ STAT ] Mid-market companies receive 250-500 invoices per month, with 68 percent still processed manually — Source, Ardent Partners, 2024. By offloading data extraction and entry to Claude via n8n, finance teams eliminate the most error-prone and tedious phase of accounts payable.

WHO BENEFITS

FOR small business owners SITUATION: handles all billing themselves with no dedicated AP staff PAYOFF: cuts invoice processing from 90 minutes to 5 minutes per day. FOR accounting firm associates SITUATION: manages invoice entry for 15+ client accounts across different accounting tools PAYOFF: eliminates context-switching and re-keying errors. FOR finance operations managers SITUATION: oversees a team of 3-5 AP clerks processing 400+ monthly invoices PAYOFF: reallocates team capacity from data entry to exception handling and vendor negotiation.

HOW IT WORKS

1. Gmail trigger node polls for unread emails with invoice-type attachments (PDF, PNG, JPEG).

2. Download attachment and pass binary file data to an HTTP Request node that calls Claude Code via n8n MCP server.

3. Claude analyzes the invoice image or PDF text and returns structured JSON with vendor, date, line items, subtotal, tax, total, invoice number.

4. QuickBooks node creates a draft invoice using the structured data, mapping Claude's output to QuickBooks line item fields.

5. Slack node sends an approval request message to the #invoices channel with invoice summary, QuickBooks link, and instructions to react with checkmark or X.

6. Slack webhook listener captures the emoji reaction. If approved, QuickBooks node updates invoice status to active. If rejected, Google Sheets node logs the rejection with timestamp and reviewer identity.

7. Google Sheets node appends a row to the master invoice log with all extracted fields, approval status, QuickBooks link, and processing time.

8. Gmail node moves the processed email to an Invoices/Processed folder and applies a processed label.

TOOL INTEGRATION

Gmail trigger requires OAuth 2.0 credentials configured in n8n. Use n8n's built-in Google OAuth credential type. GOTCHA: The Gmail trigger node polls on a schedule — minimum interval is 1 minute in production n8n instances. For Claude Code access, install the n8n-mcp npm package and configure the MCP server in Claude Code's settings file. The Claude HTTP tool node in n8n calls the MCP server endpoint. Claude returns JSON; use an n8n Set node to parse and normalize field names before feeding into QuickBooks. QuickBooks Online requires OAuth credentials with the Accounting API scope. GOTCHA: QuickBooks line item tax codes must match existing codes in your QuickBooks instance — Claude may extract a tax rate but not the exact QuickBooks tax code ID. Add an n8n Function node to map tax percentages to QuickBooks tax code IDs before creating the invoice. Slack webhook for approval reactions requires a Slack app with the reactions:read and chat:write scopes. GOTCHA: Slack reaction events fire only if the app is subscribed to the reaction_added event in the Slack API dashboard.

ROI METRICS

1. Invoice data entry time: 6-8 minutes per invoice manually → 45 seconds automated. [ STAT ] 2. Data entry error rate: reduces from estimated 2.3 percent → below 0.1 percent with Claude extraction validation (Ability.ai, 2026). 3. Approval cycle time: average 2.3 days → same-day processing for standard invoices. 4. Annual labor savings: one AP clerk at 12 hours/week at 28 USD/hour equals 17,472 USD yearly saved per clerk. 5. Late payment penalties eliminated: 5-8 percent of invoices incur penalties when processed beyond terms — near-zero with same-day automation.

CAVEATS

1. (critical risk) Claude Code MCP cannot set credentials — QuickBooks OAuth tokens, Gmail credentials, and Slack app tokens must be configured in n8n UI manually. Without this, the workflow fails at step 4. 2. (significant risk) Invoice layouts vary widely — Claude may mis-extract fields from non-standard or handwritten invoices. Add a human review step for totals above 5,000 USD. 3. (moderate risk) n8n MCP server must be restarted after Claude Code is fully restarted for the connection to reload. 4. (minor) QuickBooks rate limits: 650 requests per minute per app. Bulk invoice processing may require a throttle node between steps 3 and 4.