IntelliFlow AP — AI-Native Accounts Payable Orchestration

Inspiration

Accounts payable is one of the most expensive, error-prone back-office functions in enterprise. The average company spends $12.44 processing a single invoice manually — taking 3–5 business days per cycle. AP clerks spend 60%+ of their time on exceptions: mismatched POs, duplicate invoices, missing goods receipts. Fraud losses in AP exceed $300B globally per year.

We asked: what if an intelligent system could handle 75%+ of invoices end-to-end in under 60 seconds, with humans only touching the genuinely hard cases?

What it does

IntelliFlow AP is a production-grade accounts payable automation platform that:

1. Ingests invoices from email, PDF scan, or web portal
2. Extracts 15+ structured fields using Groq LLaMA-3.3-70B with function calling — vendor, amount, line items, GL codes, payment terms
3. Performs 3-way matching against purchase orders and goods receipts with configurable tolerance
4. Auto-approves invoices that pass matching below the $500 threshold (no human touch)
5. Routes exceptions — 9 types including fraud flags, duplicate detection, PO mismatches — to the right AP team member via UiPath Action Center
6. Escalates large approvals — invoices >$10K go to manager, >$50K to director, >$500K to CFO — via Action Center human tasks
7. Captures early-pay discounts using AI payment timing optimization, generating 36% annualized ROI
8. Posts to ERP automatically once approved
9. Tracks all SLAs with real-time dashboard — 4-hour SLA for fraud, 72-hour SLA for standard exceptions

How we built it

Backend: FastAPI + SQLAlchemy + SQLite with an 11-state invoice status machine

AI Layer (Groq):

• llama-3.3-70b-versatile — invoice data extraction via structured function calling
• llama-3.2-90b-vision-preview — scanned PDF / image invoice OCR
• llama-3.1-8b-instant — payment timing optimization and exception advisory

UiPath Platform:

• Maestro BPMN — 8 service tasks, 3 exclusive gateways, 3 user tasks orchestrate the complete AP lifecycle
• Action Center — exception resolution and approval tasks with custom forms (IntelliFlow-AP-Exception, IntelliFlow-AP-Approval catalogs)
• Agent Builder — wraps Groq AI extraction and anomaly detection as callable agents
• Orchestrator — webhooks deliver human decisions back to the FastAPI backend in real time
• Robot / UiPath Assistant — email monitoring, ERP posting, vendor notifications

Integration: HMAC-SHA256 signed webhooks, OAuth2 machine-to-machine auth, ngrok tunnel for local dev

Challenges we ran into

• Groq function-calling schema differences from Anthropic SDK — had to rewrite all tool schemas to OpenAI-compatible format (parameters instead of input_schema, tool_choice="required" instead of {"type":"any"})
• UiPath service tasks require public HTTPS — solved with ngrok tunnel for local development
• 3-way match tolerance — implemented configurable percentage-based tolerance at both invoice-level and line-item level to handle real-world vendor rounding
• Webhook signature replay attacks — HMAC-SHA256 with timing-safe comparison, configurable secret per environment

Accomplishments that we're proud of

• 75%+ touchless rate on well-formed invoices in testing
• <60 seconds end-to-end processing for standard invoices
• $0 LLM cost thanks to Groq free tier (500 req/day on 70B, 14,400 on 8B)
• 9 exception types with distinct SLAs and routing rules
• Full audit trail on every invoice — every state transition, every AI decision, every human action is logged

What we learned

• UiPath Maestro BPMN is genuinely powerful for process-centric automation — the visual designer maps 1:1 to the business process, making it easy for both developers and business analysts to understand
• Groq's free tier is production-usable for hackathon-scale demos (and beyond) — the 70B model with function calling is excellent for structured document extraction
• Human-in-the-loop design requires as much thought as the automation itself — the exception routing logic and SLA tracking are as complex as the AI extraction

What's next for IntelliFlow AP

• Multi-tenant SaaS with tenant-isolated databases and configurable thresholds per client
• ERP connectors for SAP, Oracle NetSuite, Microsoft Dynamics via REST APIs
• Email ingestion robot — UiPath Robot monitors AP inbox, attaches PDFs automatically
• Vendor self-service portal — vendors check invoice status, upload supporting docs
• Predictive analytics — ML model trained on historical AP data to predict exception likelihood before matching

5. UiPath Components Declaration

Agent type: Low-code agents (Agent Builder) + custom Python FastAPI service layer External AI: Groq Cloud — llama-3.3-70b-versatile, llama-3.1-8b-instant, llama-3.2-90b-vision-preview

Built With

• bpmn
• fastapi
• groq
• llama-3.1-8b
• llama-3.2-90b-vision
• llama-3.3-70b
• maestro
• python
• sdk
• sqlalchemy
• sqlite
• uipath