InvoiceFlow AI：面向财务运营的 LLM 智能审核与催收工作流系统

InvoiceFlow AI is an LLM-driven workflow system designed for financial operations, supporting two core scenarios: accounts payable (AP) invoice auditing and accounts receivable (AR) collection follow-up. Key capabilities include structured extraction, policy retrieval, anomaly detection, human audit checkpoints, and full audit trail.

Source: GitHub project by GargiGupta-io (original title: invoiceflow-ai), released on 2026-05-27. Link: https://github.com/GargiGupta-io/invoiceflow-ai

Enterprise financial operations involve repetitive but high-risk decision scenarios. AP teams need to audit supplier invoices for policy compliance and duplicate payment risks; AR teams need to draft follow-up emails for overdue accounts while maintaining professional relationships.

Traditional manual methods are inefficient and prone to missing risks. Generic chatbots lack structured output, policy basis, and audit trails, failing to meet financial compliance requirements. InvoiceFlow AI addresses these pain points as a specialized workflow system.

System Architecture: 6-layer agent workflow:

1. Document input (PDF/text/email)
2. Extractor Agent (structured extraction)
3. Workflow router (split into AP/AR flows)
4. Policy context retrieval (RAG)
5. AP decision flow / AR drafting flow
6. Tool call tracking + human audit checkpoint → structured results with evidence

AP Flow: Checks include field integrity, PO matching, duplicate invoice detection, payment term verification, approval threshold evaluation, invalid invoice identification, and amount calculation validation. Outputs one of four actions: approve, review, reject, missing_info, with anomalies and evidence.

AR Flow: Analyzes overdue days, counts historical reminders, identifies unproven payment claims, detects missing info, and evaluates escalation triggers. Outputs escalation level, email draft (with TTS-friendly version), and evidence.

Structured Extraction & Schema Validation: Supports heuristic (rule-based) and LLM (schema-constrained JSON with validation repair) modes. LLM gateway includes PII desensitization, delay metadata recording, and token usage tracking.

Policy Retrieval & Evidence: Built-in financial knowledge base with vector-based policy retrieval; all recommendations include cited policy fragments for traceability.

Human-in-Loop Audit: High-risk/low-confidence cases are routed to manual audit. Auditors can view workflow paths, key fields, anomalies, and evidence via the /ui interface.

Audit Trail & Evaluation: Complete audit records include extraction mode, prompt version, LLM calls, retrieval repairs, time costs, final recommendations, and agent tool chains. Built-in evaluation dataset and runner support CI/CD integration to prevent reliability regression.

Applicable scenarios:

• Medium enterprises: Need automated invoice auditing without high ERP costs.
• Financial shared service centers: Require standardized processes and unified audit records.
• Compliance-heavy industries (finance, healthcare, manufacturing): Need complete decision evidence chains.
• Remote financial teams: Need asynchronous collaboration and workflow status visualization.

Quick experience steps:

1. Use provided example cases:
   • ap_002_missing_po (missing PO invoice)
   • ap_004_duplicate_invoice (duplicate detection)
   • ar_001_first_followup (first collection follow-up)
   • ar_003_payment_claim_no_proof (unproven payment claim)
2. After starting the system, access the UI at http://127.0.0.1:8000/ui to run examples or upload local files.

InvoiceFlow AI demonstrates how to embed LLM capabilities into specific enterprise workflows instead of just chat interfaces. Through structured extraction, policy retrieval, tool tracking, human audit, and full audit trails, it provides an intelligent and trustworthy solution for financial operations. It serves as a valuable architectural reference for developers exploring LLM enterprise applications.