Ertas for Invoice & Receipt Processing

The Challenge

Accounts payable teams process invoices from hundreds of vendors, each with a different layout, terminology, and level of detail. Extracting structured data (vendor name, invoice number, line items, quantities, unit prices, tax amounts, payment terms, and bank details) from these diverse formats is tedious manual work that is prone to errors. A single transposition error in an amount field can cascade into payment discrepancies, vendor disputes, and accounting reconciliation issues.

Template-based extraction tools work for high-volume vendors with consistent formats but fail on the long tail of vendors that send invoices in unique layouts. Machine learning-based extraction tools improve coverage but still struggle with handwritten invoices, scanned documents with OCR artifacts, multi-page invoices with complex line item tables, and invoices in multiple languages. The accuracy gap between what extraction tools deliver and what AP teams need means that every invoice still requires human verification, defeating the purpose of automation.

The Solution

Ertas lets AP teams train extraction models on their own invoice corpus, so the model handles the exact vendor formats, layout variations, and field naming conventions they encounter. In Ertas Studio, teams upload annotated invoices as JSONL, each entry pairing the OCR text from an invoice with the corresponding structured data fields, and train a model that maps unstructured invoice text to clean, structured output matching their ERP system's field schema.

The trained model handles vendor-specific quirks that generic tools miss: the vendor that lists 'Net Amount' instead of 'Subtotal,' the one that embeds tax in line item prices instead of listing it separately, the European vendor that uses a comma as the decimal separator. Because the model learns from the organization's actual invoices, it reflects the real mix of formats and edge cases. Training runs on Ertas's cloud GPUs, so the annotated invoices are uploaded to Ertas; teams can swap real bank account numbers for realistic placeholder values before upload. The finished model is exported from Hub and runs on the team's own servers through Ollama, llama.cpp or vLLM, processing invoices as they arrive and outputting structured data ready for ERP import. Outputs that fail the team's validation checks, such as line items that do not add up to the invoice total, go to human review, and the corrections become training examples for the next run.

Key Features

Example Workflow

A construction company's AP department processes 3,000 invoices monthly from 400+ vendors: material suppliers, subcontractors, equipment rental companies, and professional services firms. The team annotates 8,000 invoices with structured field mappings, replaces real bank details with placeholder values, and uploads the dataset to Ertas Studio. They fine-tune a model targeting their ERP's 22-field invoice schema, including project codes, cost categories, and retention amounts specific to construction billing, running a few base models in parallel and comparing field-level accuracy on a held-out set of invoices. The best model is exported and served on the company's own server, connected to their invoice intake workflow. Incoming invoices are OCR-processed and sent to the model, which outputs structured data matching the ERP import format. Invoices that fail the team's validation rules go to an AP specialist, and the corrected outputs are added to the dataset for the next training run. Each month the team tracks how many invoices pass without edits to decide when to retrain.