Enterprise Accounts Payable
Automation Agent

Monitors Gmail and Outlook for invoices. Accepts PDF uploads via REST endpoint. Parses EDI 810 feeds. Classifies documents and stores raw files with metadata.

GPT-4o vision extracts all invoice fields: vendor, invoice number, dates, line items, quantities, unit prices, totals, PO number, and payment terms. pytesseract OCR fallback for scanned documents.

Checks required fields, amount tolerances, due dates, vendor master, duplicate invoices, and currency matching. Flags exceptions with specific error codes for downstream routing.

Compares invoice against purchase order and goods receipt within a configurable tolerance (default 2%). Returns full match, partial match, or mismatch. Routes mismatches to the exception handler.

Applies a configurable approval matrix. Below-threshold invoices auto-approve. Above-threshold routes to the designated approver via email and SMS with a one-click approve or reject link. Escalates after configurable timeout.

Creates exception records. Drafts vendor communications using GPT-4o. Queues for human review with full context. Tracks resolution status in Supabase.

Prioritizes by due date. Groups payments for batch processing. Generates payment run reports. Triggers ERP payment entries on schedule.

Writes approved invoices and payments to QuickBooks, Xero, SAP, or NetSuite. Handles field mapping per ERP schema. Retries on transient failures. Logs transaction IDs back to Supabase.

Writes a complete audit trail entry for every state transition. Sends email and SMS notifications to the AP manager. Generates daily summary reports automatically.

Classifies pages in multi-page architectural PDFs using Claude vision. Routes schedule pages to a dedicated extraction agent that pulls window and door data — type, size, quantity, location, finish, elevation reference — and exports a formatted Excel takeoff file for non-technical users via a browser upload interface.

Ingests sales orders from email bodies and attachments in English and French. GPT-4o extracts SKUs, quantities, delivery dates, customer references, and line items into a structured SalesOrder model. Extractions below a 0.70 confidence threshold are automatically routed to a human review queue. Validated orders are mapped to the target ERP schema and pushed via the Order ERP Sync Agent with retry logic and a full audit log of every attempt, source email, extracted fields, and outcome.

Reads raw .docx files using python-docx. Extracts full heading hierarchy, section content, table structure, and document metadata including title, author, and creation date. Outputs a structured dict ready for AI normalization. Handles files with missing or inconsistent heading styles gracefully.

Sends extracted document content to GPT-4o with a strict JSON schema prompt. Maps content to a master 10-section structure: Title, Document Control, Purpose, Scope, Definitions, Responsibilities, Procedure, Records, References, and Revision History. Returns confidence scores per section. Documents below 0.70 overall confidence are automatically flagged and routed to a human review queue.

Rebuilds a clean, formatted Word document from normalized JSON content. Python controls all Word styling via python-docx: Title style, Heading 1, Heading 2, Normal, Table Grid, headers, footers, margins, and page numbering. OpenAI never controls formatting. Outputs clean .docx files to the output folder with a confidence footer note on every document.

Processes entire input folders in batch. Logs each file as successful, failed, or partially processed with per-file confidence scores and processing time. Includes retry logic for OpenAI API failures (up to 3 attempts). Generates a JSON summary report after each batch run. Wired to FastAPI via /docx endpoints and accessible through a non-technical browser UI at ap.datawebify.com/docx-ui.