# An autonomous AP agent handling 2,000+ invoices a month
> Aralab's finance team keyed more than 2,000 supplier invoices a month into the ERP by hand, each in a different supplier layout. Twistag built an accounts-payable agent: Claude reads every invoice, and separate code recalculates each line, tax and total against the purchase order. One engineer shipped it in six weeks, and three finance roles moved to forecasting and supplier work.

Client: Aralab · Industry: Manufacturing · Practices: AI & Agents · Timeline: 6 weeks
## Key results
- **2,000+** supplier invoices processed a month
- **3** finance FTEs redirected to strategic work
- **6 weeks** to live, built by one engineer
## The challenge

**Two thousand invoices, no two alike.**

Twistag built an autonomous accounts-payable agent for Aralab, a manufacturing business, that processes more than 2,000 supplier invoices a month and redirected three finance FTEs from data entry to strategic work — live in six weeks, built by one engineer. Every supplier sends a differently formatted PDF, and template-based OCR breaks on that variety daily. Twistag built a hybrid system: Anthropic Claude reads and interprets each document, while deterministic code independently recalculates every line total, unit price and tax figure against the matched purchase order. The language model proposes; the validation engine verifies.

Aralab's operations generate over 2,000 supplier invoices a month, each arriving as a PDF, and that is where the consistency ends. Every supplier uses its own layout, line-item conventions, tax formatting and reference-number scheme. Processing them meant opening each invoice, reading it as a human, and working out where each piece of data belonged in the ERP. Aralab's three-person finance team had done this for years, because the volume left no other option. The real cost was a finance team buried in transcription, with no time for cash-flow forecasting, supplier negotiation or financial planning. The business was paying for judgement and getting data transcription instead.

## The approach

**The model reads. The code checks.**

1. **Read any layout.** Claude turns each supplier PDF into structured data: supplier and NIF tax number, invoice number and date, line items, taxes and totals. A supplier changing its format needs no new template. 
2. **Verify every figure in code.** Deterministic checks recalculate each unit price, line total, tax amount and grand total against the matched purchase order. Exact matches resolve at zero model cost; Claude handles partial deliveries and split orders, with a confidence score and its reasoning. 
3. **Review only what is uncertain.** A dashboard sorts every invoice into fully automated, awaiting line-item review, or flagged for full review. LangFuse traces each model call, so every decision can be audited and the cost per invoice is visible in real time. 

## The impact

**Same team, different work.**

The system shipped in six weeks with one engineer. Aralab's three-person finance team now processes more than 2,000 invoices a month through the review dashboard. No human is required for the patternable 80%; the ambiguous 20% gets a fast review surface. Three FTEs that were buried in data entry moved to cash-flow forecasting, supplier negotiation and financial planning — same headcount, different output. LangFuse makes each decision auditable after the fact, which is what let a finance team delegate at all, and cost per invoice is visible in real time, so the workflow cannot become quietly more expensive than the people it replaced.

> The LLM proposes; the validation engine verifies. Neither alone is sufficient.

## Technologies used

- Anthropic Claude
- GCP
- Firebase
- LangFuse

> Twistag built an autonomous accounts-payable agent for Aralab that processes more than 2,000 supplier invoices a month.

---

Source: https://twistag.com/case-studies/ai-invoice-automation-aralab-manufacturing