Oldcastle APG, a North American building materials manufacturer, says its new AI order-entry system automated 95% of the order lines it processed during its first two weeks in production. Customer service staff still check orders before they enter the company’s business software.

According to the firm, the company’s previous process relied on more than 200 representatives to enter roughly 13 million order lines a year into Infor M3, its enterprise resource planning system. Representatives spent upward of four hours a day retyping details from emailed PDFs, faxes and scans.

The work involved more than transcription. Representatives had to identify customers and match their product descriptions to Oldcastle’s records. Mistakes in quantities, product codes and prices could then carry into fulfillment and invoicing. The new workflow separates reading the document from checking the order against company records.

Repeat orders bypass the model

Incoming documents are collected through Amazon’s email and storage services and Microsoft Power Automate. Anthropic’s Claude Sonnet 4.6, running through Amazon Bedrock, extracts customer details, products, quantities and prices. The system matches that information with Oldcastle’s customer and product records in an Amazon Aurora database.

A representative then compares the proposed order with the original document and checks the suggested matches. Once confirmed, the order enters Infor M3 with an audit trail. Oldcastle also saves corrections made by staff, allowing repeat orders from the same customer to skip the AI processing step altogether.

In its first two weeks, the workflow processed more than 10,000 lines across 1,100 orders. The case study estimates that the system redirected roughly 170 hours of manual work, although it does not explain how that figure was calculated. The case study puts AI processing costs below $0.04 per page but does not specify whether that includes infrastructure, maintenance or employee review. The eight-week build involved two developers and an architect.

Oldcastle targets receiving operations next

Oldcastle’s system also handles purchase orders written in French, including those from customers in Quebec, without additional configuration. Oldcastle receives purchase orders containing up to 100 line items, in formats including PDFs, images, HTML and text. French-language orders follow the same extraction and matching process as English-language orders.

Rather than developing a separate extraction system for each workflow, Oldcastle plans to reuse the existing Amazon Bedrock engine. AWS and Oldcastle say this approach can reduce the additional development work to weeks, allowing the company to extend document automation into other parts of its operations.

Oldcastle says it plans to extend the system to purchase orders and delivery notes used in receiving. It is also routing orders from major retail partners to be automatically prepared for further processing. Its early results indicate less manual data entry.

Personalized Feed
Personalized Feed