A distributor whose customers still send orders and price changes by email should, before replacing its order system, use AI to pull the item, quantity, price and delivery date out of those emails and their attachments for one quarter and count how often each disagrees with the current system.
MRC Global, a distributor of pipe, valves and fittings, began implementing a cloud-based ERP system for its U.S. segment in August 2025, replacing a mainframe system in use for over 30 years. Its quarterly report, filed November 5, 2025, says the challenges that arose, which it called temporary, adversely affected our ability to process orders, fulfill customer shipments, integrate with other necessary software systems and issue timely invoices, and that a 15% fall in U.S. sales for the quarter against the same quarter of 2024 was primarily due to those challenges. DNOW completed its acquisition of MRC Global on November 6, 2025. In its annual report filed February 26, 2026, DNOW wrote that the combined company continues to address configuration, data-migration and stabilization matters arising from that implementation. Neither filing mentions email.
The count comes from a log, one row per order line. For every email and attachment in the quarter, a language model reads the customer, the customer's part number, quantity and unit, the unit price, the ship-to address and the delivery date, and a lookup pulls the keyed order line, contract price and cross-reference from the current system and marks each field as a match or a mismatch. The customer service manager then sorts every mismatch by cause: the email was wrong, the system was wrong, or the system holds no rule for what the customer sent, as with a part number that has no cross-reference or a quantity in cases for an item sold by the each. That third group should go to the replacement project as test cases for its data migration.
How well the model reads depends on the layout. A preprint from the University of Las Palmas de Gran Canaria, posted April 1, 2026, tested two general-purpose models on 2,400 synthetic Spanish electricity invoices in six layouts, and names document template structure as the primary determinant of extraction difficulty. Its best configuration gave the models worked examples from three layouts and scored 97.61% and 96.11% on the other three in F1, a measure that combines fields read wrongly with fields missed. The documents were invoices rather than orders, and the authors write that validation on real-world invoices would strengthen its practical applicability. So the customer service manager should give the model a sample order in each large customer's format and check part of every week's rows by hand.
In the first two weeks, IT exports the order inbox while the pricing manager exports contract prices and the customer and item cross-references. From week three the extraction runs weekly and the customer service manager reviews the sample. By week 13 the COO has each customer's mismatch count by cause, and the ERP project lead has the list of missing rules.
The quarter's work ends at that count and that list. The model writes nothing into the current system, and customer service keeps entering orders the way it does today until the replacement goes live.
In the Census Bureau's E-Stats, a sale counts as e-commerce when the order is placed or the price and terms are negotiated over the Internet, mobile device (m-commerce), Extranet, EDI network, e-mail, or other comparable online system, and the wholesale survey behind those figures asks merchant wholesalers to report three numbers: total sales, e-commerce sales and e-commerce sales made through EDI networks.
