Gmail Invoice Cost Tracking Workflow (OpenAI + Airtable + Slack)

Automatically track invoice costs from Gmail PDFs—then log confidently in Airtable or route questions to Slack

This Gmail Invoice Cost Tracking Workflow monitors your inbox for new invoice emails, extracts cost details from attached PDFs using OpenAI, matches invoices to your active jobs, and records results in Airtable. If confidence is low, it sends a review request to Slack.

What this workflow does

• Checks Gmail every minute for new unread messages that include attachments and match invoice-like subject keywords.
• Downloads the first PDF attachment and extracts its text into a single combined document.
• Uses OpenAI (gpt-5-mini) to parse extracted invoice text along with your list of active jobs, extracting:
  • Vendor
  • Amount
  • Category
  • Invoice date
  • Job match confidence score
• Applies a confidence threshold of 0.75 to decide whether to auto-log or request confirmation.
• Creates a new Airtable record (in your “Job Costs” table) with Job, Amount, Vendor, Category, and Invoice Date when the match is confident.
• Posts a Slack message when confidence is below 0.75, prompting a human to confirm the job match.

Use cases

• Cost tracking for agencies or service teams receiving invoices by email.
• Reducing manual work by automatically logging invoice costs into a “Job Costs” Airtable table.
• Ensuring accuracy with a Slack review step when invoices don’t clearly map to a job.

Technical details

• Gmail trigger: runs every minute for new unread messages with matching subject/attachment rules.
• PDF text extraction: downloads the first PDF attachment and combines extracted text.
• OpenAI: parses invoice fields and computes job match confidence (model: gpt-5-mini).
• Decision logic: compares confidence score against 0.75.
• Airtable: writes confirmed invoice costs into the “Job Costs” table.
• Slack: sends review messages for uncertain matches.
• Workflow nodes used: if, slack, airtable, sticky note, gmail trigger, n8n-nodes-langchainagent.