Intelligent Invoice Processing (AP)

Invoices arrive from three separate channels, get OCR'd and structured by AI, checked for duplicates before anything is booked, routed for human approval in Slack, and archived — with a full audit trail of every rejection, duplicate, and review.

The problem

Manual invoice entry is slow and error-prone. Invoices arrive from email, an upload portal, and a shared drive folder with no single intake point. Duplicate invoices get paid twice without a systematic check, and finance has no consistent approval trail for what got paid and why.

The system

Invoices can enter through a webhook upload, a monitored Gmail inbox, or a watched Drive folder — each normalized into the same internal format so the rest of the pipeline doesn't care where the file came from. The file is validated, OCR'd, and the extracted text is passed to an AI extraction step that returns structured fields (vendor, amount, date, line items) against a strict schema. Before anything is saved, the system checks for a duplicate invoice; if one exists, finance is notified and the record stops there instead of double-booking. Valid, non-duplicate invoices get an AI-generated summary and go to Slack for a one-click approval decision — approved invoices are marked and archived, rejected ones notify the submitter, and anything ambiguous routes to human review rather than being forced through.

Architecture

flowchart TD
    A1[Webhook Upload] --> N[Normalize Source]
    A2[Gmail Trigger - Invoice Inbox] --> N
    A3[Drive Trigger - Invoice Folder] --> N
    N --> B[Validate and Extract File]
    B --> C{Is Valid File?}
    C -->|No| L1[Log - Validation Error]
    C -->|Yes| D[Google Vision OCR]
    D --> E[Extract OCR Text]
    E --> F[AI Extract Invoice Data
Structured schema]
    F --> G[Normalize Extracted Data]
    G --> H{Has Required Fields?}
    H -->|Yes| I[Check Duplicate Invoice - Postgres]
    I --> J{Is Duplicate?}
    J -->|Yes| K[Notify Finance: Duplicate
Slack + Email + Log]
    J -->|No| M[Save Invoice]
    M --> P[AI Invoice Summary]
    P --> Q[Email + Slack Approval Request]
    Q --> R{Approved?}
    R -->|Yes| S[Mark Approved]
    S --> T[Notify Submitter: Approved]
    T --> U[Archive to Processed Invoices - Drive]
    R -->|No| V[Mark Rejected]
    V --> W[Notify Submitter: Rejected]
    R -->|Unclear| X[Notify Finance: Human Review]
    U --> Y[Log - Execution Success]
  

What makes it production-grade

• Three intake channels normalized to one pipeline — no change needed in how the business sends invoices
• OCR + structured AI extraction, not a single free-text prompt — line items and totals come back schema-validated
• Duplicate detection before booking — the single most common cause of AP overpayment, handled automatically
• Human-in-the-loop approval — AI drafts and summarizes, a person still authorizes payment
• Original document reconstruction and archival, not just the extracted data
• Full audit trail — every duplicate, validation error, and human review is logged

Expected business impact