Skip to content
BeryxTech

Lighthouse Projects · Industry

Supplier invoice automation with AI

From intake to posting in the ERP, with people only on the exceptions.

Problem

High monthly invoice volume processed manually. Long average processing time. Posting errors that generate rework and supplier friction.

Solution

Centralized intake (email, portal, EDI), OCR + LLM extraction, rule engine validation (3-way match), automatic ERP integration. Exception workflow for humans only where they actually add value.

KPIs

  • Average invoice processing time
  • % of invoices in straight-through processing
  • Reduction in posting errors
  • Hours/month freed in accounts payable

How it is measured

Every KPI with a baseline and a target.

The numbers for this project come from your data. We measure the baseline during framing, agree the target with you before building and, in production, measure the same KPIs the same way.

  1. 01BaselineMeasured during framing, with the data of the process as it runs today.
  2. 02TargetAgreed with the case owner before building, starting from the baseline.
  3. 03In productionThe same KPIs, measured the same way, compared with the baseline.
Average invoice processing time
From intake to posting in the ERP, using the dates recorded in the ERP and the inbox.
% of invoices in straight-through processing
Invoices posted with no manual touch, divided by the total received in the period.
Reduction in posting errors
Credit notes, reversals and corrections caused by posting errors, counted in the ERP.
Hours/month freed in accounts payable
Accounts payable hours spent on invoices, measured by sampling before and after.

Before starting

Prerequisites and risks.

Prerequisites

  • Read access to the ERP (suppliers, purchase orders and goods receipts) and to the mailbox or portal where invoices arrive.
  • Purchase orders and goods receipts recorded in the ERP: without them there is no 3-way match.
  • Someone from accounts payable to validate the rules and handle the exceptions during the pilot.

Risks and how we handle them

  • Poor scans or unusual layouts

    Extraction fails more often on weak scans and rare layouts. Every extracted field carries a confidence score, and below the threshold the invoice goes to human review.

  • Wrong postings in the ERP

    The project starts in proposal mode: the system prepares the posting and a person approves it. Automatic posting is only switched on when the errors measured in the pilot justify it.

  • Supplier data

    Invoices carry tax IDs, IBANs and addresses. Access follows the ERP roles, and the data scope is agreed with your IT team before the sources are connected.

Want to see this in your operation?

90 minutes with you and your team. We leave with 3 concrete opportunities and estimated payback.