Invoice Intelligence
From inbox to ERP in seconds — AI that reads, classifies and routes every invoice automatically.
An LLM-powered pipeline that intercepts employee invoice emails, classifies each one using few-shot examples from past cases, updates Eloficash in real time, and surfaces edge cases through a human-in-the-loop dashboard.
Hundreds of invoice emails.
No automated processing.
PwC teams received daily emails from employees and suppliers containing invoices to process. Every email had to be read, understood, categorised and forwarded to the right department — a wholly manual process, slow and error-prone.
Without automation, some invoices went unpaid for too long, others were settled without validation, and disputes accumulated with no structured follow-up.
LLM + few-shot learning.
From inbox to ERP in real time.
Invoice Intelligence intercepts incoming emails, extracts invoice content (message body and attachments), and uses an LLM enriched with examples from past cases to decide the action to take — with no human intervention in standard cases.
Ambiguous cases are escalated to a supervision dashboard where operators validate or correct the decision. These corrections automatically feed the model's example base, improving it continuously.
End-to-end pipeline
The main flow processes each email in under 5 seconds. Ambiguous cases branch to the human dashboard, whose corrections continuously feed the few-shot example base.
When the model's confidence falls below the configured threshold, the invoice is escalated to the dashboard. The operator sees the full context, chooses the correct decision, and their correction is automatically added to the few-shot example base. The model improves with every correction — no costly retraining required.
Few-shot prompting
anchored in PwC context
Rather than fine-tuning an expensive model, Invoice Intelligence dynamically builds a prompt enriched with similar examples retrieved from a base of past cases. The model sees real precedents before deciding.
The invoice is recognised, the amount matches a purchase order in Eloficash, and the sender is a registered supplier. Payment triggered automatically.
Amount diverges from the contract, duplicate billing detected, or unrecognised supplier. The invoice is frozen and a dispute ticket is opened automatically.
Missing information, hierarchical approval required, or insufficient model confidence. Escalated to the human dashboard for a final decision.
The product in pictures
LLM-first,
ERP-native
No costly fine-tuning. The power comes from few-shot examples and direct ERP integration. The model improves through human corrections, with no heavy ML infrastructure.