Taking over revealed how often the same data was copied
In late July 2026, Matt, who had recently joined, was preparing to take over order processing. After reviewing it, his reaction was: “It is hard to follow. There are so many rules to judge, and many steps seem repetitive.” A later review confirmed that the same items were manually copied from quotations to order emails, BC orders and tracking sheets. Vida and Matt of Team 1 began improving this everyday process.
Office Hours turned a broad idea into a practical first step
The original idea covered quotation through ordering, with multiple system comparisons and checks. More complete requirements brought more permissions and decisions. After CoE discussions and Office Hours, phase one narrowed to “read the order email and generate a BC order preview”: produce a draft that can be checked against the source, then expand gradually.
With a narrower scope, Matt explained each field on the order-entry screen: where the data came from, how to fill it and what exceptions applied. These rules became a field-mapping table and exposed the hardest field: the item number.
The answer to item numbers was in past orders
A single item could match several similarly described item numbers, making even people check old orders before choosing. Correctly reading the item name did not tell AI which number to select. The team decided to pair historical order emails with actual BC orders to map names to numbers, but obtaining the historical data became another obstacle.
During project discussions, AI team colleagues suggested retrieving history through browser interaction with an already accessible order screen. This enabled the mapping table. AI supplies a number when the match is clear; missing or multiple matches return to a person for judgment. Workshop support addressed a concrete problem: identify the missing material, then find a feasible way to obtain it.
An empty column sent the team back to the instructions
During the pilot, the vendor column was entirely blank, and the team considered removing it. Review revealed that the instructions told AI to look in the quotation, while the vendor appeared only in the email body. Correcting the source kept the column useful. Missing information and unreadable information were also recorded separately so reviewers could tell whether to request data or check the source.
The workflow now connects email intake, field extraction, item-number mapping, SA review and writing to BC after confirmation. Pending items and exceptions are recorded. People provide corrections and AI helps enter them. The change improves not only data entry, but also the basis for handover and review.
Before and after: how the work changed
Compare the same work steps to see the roles of tools and people.
| Work step | Before | After |
|---|---|---|
| Preparing order data | SA staff read order emails and quotations field by field and enter data into BC manually. | AI extracts a draft for SA review, then helps write the verified data to BC. |
| Order-entry rules and handover | Rules and exceptions depend largely on individual experience and verbal handover. | Sources, entry rules and exceptions are documented in a field-mapping table and extraction instructions for shared review. |
| Selecting item numbers | Similar item numbers require checking old orders or asking experienced colleagues. | Historical-order mapping supplies clear matches; missing or multiple candidates still require human judgment. |
| Missing data | Blank fields make it hard to tell whether information is absent from the document or was missed by AI. | Absent and unreadable information are recorded separately, alongside low-confidence and multiple-candidate exceptions. |
| Email intake | Manual CC or forwarding is required; forgotten emails may never enter the workflow. | The platform receives emails directly, reducing manual forwarding. Duplicate arrivals are added and marked. |
Applying this experience to your own work
Workflow connections, field specifications and review design are complete, but no formal time or accuracy statistics are available. Other teams transcribing emails and attachments can start with a task whose output is easy to check, document field sources and exceptions, then decide what AI handles and what people confirm.